All News (Page 5)

Meta’s new AI agent app is quickly taking hold of the consumer market, unseating ChatGPT as the top free app on Apple’s App Store in recent days. The app, which features a personal AI agent that books appointments, shops, and organizes calendars on people’s behalf, is the talk of tech circles this week.

Named Muse, the app, which uses Meta’s latest Muse Spark AI model, connects users to a range of services that its AI agent autonomously handles. That means handing over access to accounts across a user’s phone in exchange for Muse booking a vacation or handling restaurant reservations on their behalf. It’s the first major consumer AI agent push by Meta, and it’s catapulting the company for now, past its competitors.

Not everyone is rolling out the red carpet for Meta’s Muse, however. Amazon blocked Meta’s assistant from accessing its online store, saying that “continued access by an unauthorized AI agent violates Amazon’s Conditions of Use, to which our customers have agreed.” Amazon said Meta never asked for permission to access its site. Meta didn’t immediately respond to a request for comment, though the company announced it would partner with Shopify for some of Muse’s shopping features.

“It is only a matter of time before there is an Apple and Google version of Muse and possibly TikTok, in addition to the frontier LLM agents,” Nikesh Arora, chief executive at Palo Alto Networks, wrote on X. “Every app that is a services, marketplace or commerce app will need to existentially decide to open [Application Programming Interfaces] for consumer agents to interact. Smaller players have no choice. Either the consumer benefits or distribution aggregators will demand a higher transaction fare.”

Muse’s break-out success represents a much needed victory for Meta in the AI arena, following a series of struggles by the company to compete with frontier labs like OpenAI and Anthropic.

Muse’s fast-rising popularity, and Amazon’s reactionary block, show how quickly agentic technology is transforming the consumer space. Meta released Muse just as Anthropic and OpenAI released cheaper versions of some of their models. Apple also recently rolled out its revamped Siri as part of its newest iOS update.

The bigger questions raised by Muse’s release are who controls the shopping and user interfaces Meta is trying to access, and how major companies will agree to such access. For Amazon, Muse is a threat because it bypasses the usual process of humans visiting its website or app and seeing the advertisements that serve as a key profit driver.

Amazon has taken action against other companies over similar issues. Last year, it sued Perplexity over its Comet AI tool, which could access Amazon accounts much like Muse does. Amazon recently lost a ruling on that claim after the U.S. Court of Appeals for the Ninth Circuit found that users were ultimately choosing to access Amazon via Perplexity’s bot; Amazon is still fighting the issue.

Amazon has its own shopping agent, called Buy for Me, which performs tasks similar to Muse’s. The company has said brands can opt out of the feature, though doing so requires action, since brands are automatically opted in. CEO Andy Jassy has also said Amazon is exploring AI shopping agent partnerships with third-party companies.

Privacy is sure to be a key issue for consumers weighing assistants like Muse, which allow for personal touches—like the ability to name your assistant— but require access to sensitive information, such as bank accounts, to truly work on a user’s behalf.

Meta says user logins are kept secret and that online purchases use a one-time card number that hides a person’s actual card details. Still, Meta has been at the heart of previous privacy issues. Meta has faced multiple proposed class-action lawsuits related to its smart glasses, including one alleging that it sent some footage captured by the glasses, including highly personal material, to third-party human reviewers for annotation and AI development. Meta has disputed the allegations.

The rise of Muse is timely for Meta: CEO Mark Zuckerberg is expected to unveil updates to the company’s smart glasses at its annual Connect event in Menlo Park, Calif., on Wednesday.

This story was originally featured on Fortune.com

For elite students, college has never looked like a worse deal. Tuition is nearly half a million dollars, AI is threatening to destroy the entry-level jobs a degree used to guarantee, and now in the hottest industry—tech—it’s cool to skip school entirely. So on Tuesday, Andreessen Horowitz became the first venture capital firm to sell an alternative.

“The bet is that there’s an opportunity to educate the next generation in a way that is more future-forward, and that’s a big market opportunity,” Gagan Biyani, a cofounder of $3.2 billion ed-tech unicorn Udemy who is now running the new Horowitz Andreessen Academy, told Fortune. Higher education, he noted, is a $2 trillion industry. “We think there are lots of ways to turn this into a successful business.”

The Academy is putting $42 million from a16z and a group of individual investors into the one-year San Francisco-based program.  It is a separate for-profit company from the VC firm. It is not accredited, will award no degree, will not take federal student aid, and does not guarantee a job or a paid internship. The founding class will be about 50 students starting Sept. 2027, and it’s tuition-free—plus every student will get $50,000 in compute credits and a $5,000 travel budget. 

The Academy then intends to seek regulatory approval for a two-year program starting in 2028. If approved, Biyani said, it expects “to charge something similar to elite private universities”—in the range of $60,000 to $90,000 a year.

Biyani insisted that the school isn’t a rejection of the premise of higher education—”we believe in college,” he said. Instead, it’s rejecting the idea that 18-year-olds in SF are only valuable insofar as they have a startup idea to dedicate a decade to. Indeed, they are looking to find the best and the brightest high schoolers—call them pre-idea founders.

The Academy’s homepage promises students will learn from the people “shaping the future,” with a campus “buzzing” from visits from Sam Altman, Jensen Huang, Satya Nadella, Mira Murati, Fei-Fei Li and Travis Kalanick. Most of those names, Biyani said, are guest speakers who “come in for a talk or a fireside chat.” 

A second tier, the faculty, has committed to teaching multi-session courses; Biyani said Instacart co-founder Max Mullen are former Windows president Steven Sinofsky are two who have agreed. A third tier, mentors, will hold office hours. The course list cited in a blog post lists “The Craft of Selling” to “AI Inference Engineering” to “Origins and Innovations in Digital Computing 1940–2025.”

Yet teaching, in any case, “is not the primary thing we plan on doing,” Biyani said. “The primary thing we’re offering is the ability for students to pursue their own objectives.” The year is split three ways: September through April on campus in San Francisco, May travelling abroad, and June through August spent in a “co-op” at a partner company.

That’s where the 10 founding partners come in, mostly from A16Z’s vast network—Anduril, Anthropic, Coinbase, Google, Meta, Nvidia, OpenAI, Palantir, Replit and Stripe. Their commitments “vary,” Biyani said. “Not all of them are going to take students for hiring.” 

Rather, some are providing compute, hardware credits or advice on curriculum; another 50-odd hiring partners, from Databricks to Boom Supersonic, “are not committed to hiring a certain number, but they’re committed to interviewing or considering our students.” Co-ops are “typically” paid, the school’s FAQ says, but not guaranteed to be.

There is no minimum age and all students have to arrange their own housing. The sample applicants that the site features includes a teenager who built an AI book indexer with “six-figure ARR” and one who “created an AI dating coach with 1.7M likes.” Admissions, Biyani said, is “proof of work,” though transcripts and test scores are still required.

Future of higher ed? Experts aren’t sure

So is this the future of education? Experts are skeptical. Bryan Caplan, the economist at the famously libertarian George Mason University, who wrote a book arguing that college is mostly just a credential, thinks it will work—for the first 50 students. But “something that can work for 50 handpicked people isn’t going to work even for the top 5,000,” he said. “MIT is going to last for centuries. Don’t worry about it.”

The idea of the Academy, Caplan said, is that skipping college generally signals that “you’re a weird person, and people are understandably nervous about hiring weird people,” he told Fortune. The Academy’s selection is severe enough that the prestige overcomes the weirdness- but only insofar as they keep it elite. 

Ryan Craig, a managing director at Achieve Partners, which builds apprenticeship programs, is excited about the idea but isn’t sure their model makes sense on year two, when it attempts to become a school. “It’s a bootcamp with a plan to be a college, but I’d say not a realistic plan to be a college,” he told Fortune. Legions of “alternative, innovative universities” charging elite-private tuition without federal aid have been tried before, and failed. “It’s hard to compete with free money from the government,” he said. 

The idea that Craig was enthusiastic about was a 21st-century work college, where students earn enough in the paid co-ops to cover tuition. But that requires the partners to actually pay high school students near six-figure salaries. 

Biyani said the school will figure that out after year one. “We will see over time how students will finance this,” he said. 

This story was originally featured on Fortune.com

Anthropic and OpenAI debuted more affordable versions of their frontier models as they battle for wallet share from increasingly cost-conscious businesses, and face growing competition from lower cost open weight models. The releases came within hours of each other, a sign of the labs’ heated rivalry, despite both companies’ recent calls for an AI slowdown.

Anthropic debuted Claude Opus 5.5, which it says performs at the level of its flagship Fable 5.1 model but costs around 40% less to run than Opus 5, which came out in July. It’s the first release in a new family of models, with Sonnet 5.5 and Haiku 5.5 expected “over the coming weeks,” Anthropic said.

Anthropic said the model is its most efficient, and it’s “passing these efficiency savings on to our customers in the form of price cuts and rate limit increases.”

OpenAI released GPT-6 Sol and GPT-6 Luna, offshoots of its flagship GPT-6 Astra model that came out earlier this month. The idea is to offer a version of Astra for everyday work, OpenAI said. It also slashed the API cost by 50% lower than the pricing on a promotion it’s currently running for Astra’s predecessor, GPT-5.6. OpenAI often releases lower priced versions of its flagship models, as it did with GPT-5.6.

OpenAI’s release offers lower API pricing than Anthropic’s, based on each company’s announcement, as shown in the table below. OpenAI attributes the cost reductions to “improvements in caching and inference,” and, like Anthropic, says it’s also “passing those savings directly onto users and customers.” 

The two labs are already deep into an all-out price war—something that could put pressure on their ability to profit from their models down the line.

“OpenAI and Anthropic are engaged in a price war that is driving down the price of AI and therefore driving down their ability to profit from it and grow the price of models,” Ara Kharazian, lead economist at Ramp, told Fortune. 

The battle is playing out on two fronts, Kharazian said: the labs are rolling out cheaper models that businesses are shifting toward, such as Opus 5.5 and GPT-6 Sol and Luna, while also announcing outright price cuts on their most expensive ones.

That dynamic is how technology markets have traditionally worked, he said, but it’s something that more bullish investors may not be factoring in.

“AI bulls assume that there will be highly performant models that provide more and more value, and therefore they should be more expensive,” Kharazian said. “But that is not how normal technology makes it to market.”

An AI slowdown? 

These are the first model releases since both companies publicly called for a coordinated attempt to “pace the frontier,” or slow the development of the most advanced models. Anthropic CEO Dario Amodei published a blog post laying out his plan for a slowdown in mid-September, which OpenAI CEO Sam Altman supported in an X post—a rare moment of solidarity between the two.

Anthropic acknowledged that the release of Opus 5.5 is its first since Amodei’s call to “pace the frontier.” The company said that Opus 5.5 underwent the company’s most rigorous alignment testing to date and was evaluated by outside groups. The model also is also launching with the same tier of cybersecurity, biology, and frontier-AI-development safeguards previously reserved for the company’s more advanced model Fable 5.1.

While today’s releases from both companies are not frontier models, as in they are not a major step up in capabilities, they show that both companies continue to push AI capabilities forward outside of the frontier. The real fight for the future of their businesses could be in gaining everyday business customers, not necessarily offering the most powerful (and expensive) models possible.

Many businesses are tightening their AI wallets and applying more scrutiny to the ROI of the technology, Randall Hunt, CTO at the AI consulting company Caylent, said.

“CFOs have seen some of the sticker shock, and they haven’t seen some of the gains that were promised in the initial investments,” Hunt said. “So they’re coming back to their planning for 2027 and beyond, and saying, ‘How can we optimize our costs here?’”

One way Caylent is doing this is by calculating how long it takes employees to complete tasks with AI tools, and then calculating “cost per task,” an approach OpenAI encouraged in a mid-July 2026 blog post as well. 

This story was originally featured on Fortune.com

President Donald Trump said U.S. and Iranian officials met on Tuesday, shortly after threatening in his address to the U.N. General Assembly that he may “annihilate” the Islamic Republic if a deal isn’t reached soon to end the nearly seven-month war.

Trump confirmed the talks in an exchange with reporters but offered few details about how the discussions transpired or the substance of the exchange..

“It was a very good meeting,” Trump said.

Earlier Tuesday, Trump in his address to the General Assembly made the case that he’s shown fortitude while other leaders have equivocated as he defended his decision to start a war with Iran.

But Trump also suggested that he could be at a crossroads in the conflict even as he called on the Islamic Republic to fully reopen the Strait of Hormuz and return to the negotiating table to find a settlement to end the conflict.

“I have a big decision to make,” Trump said. “Will a deal be made with Iran that lets them rebuild and create a far greater country than it ever was before, maybe one of the greatest in the Middle East or even the world. Or do I annihilate the Islamic Republic and do it quickly, never giving them a chance to kill and destroy people and countries again?”

The president’s comments came in his wide-ranging address to the international body in which he also celebrated this year’s U.S. military operation to oust Venezuelan President Nicolás Maduro from power, a new agreement to expand the U.S. military presence in Greenland, and pushed back against calls for greater regulation of the artificial intelligence industry.

But at the heart of the address was a fulsome defense of the war and pushback on the notion that Iran has shown a measure of resilience.

“The terrorist regime is behaving badly, not because they are strong and confident, but because they are weak and desperate,” Trump said.

Trump also accused Iran of stalling to make a deal to fully reopen the Strait of Hormuz and negotiate an endgame to the war until after the midterm elections in the United States. He insisted the tactic would have no effect on how he approaches the conflict.

Most of Iran’s delegation walked out of the chamber as Trump railed against the Islamic Republic.

Trump’s speech began a busy day of diplomacy

After his speech, Trump gathered with Danish Prime Minister Mette Frederiksen and Greenlandic Prime Minister Jens-Frederik Nielsen on the sidelines of the General Assembly to sign a new security agreement.

The deal, announced last week, came together after Trump repeatedly threatened to “take over” Greenland, a territory of NATO-ally Denmark, by military force if necessary. Trump said Tuesday the U.S. will immediately begin the process of developing a larger military presence in Greenland including “building two very major military bases.”

Under the deal, Greenland remains Denmark’s territory.

Trump also held talks with new British Prime Minister Andy Burnham. The leaders said they discussed migration, the conflict in Iran, the Russia-Ukraine war and more.

Trump also met with Japanese Prime Minister Sanae Takaichi and Ukrainian President Volodymyr Zelenskyy, and will hold a more informal meeting with Venezuela’s acting President Delcy Rodríguez, according to the White House.

He’s also scheduled to meet with leaders and senior officials from Gulf Cooperation Council nations — Bahrain, Kuwait, Oman, Qatar, Saudi Arabia, and United Arab Emirates — to discuss the way ahead with the war in Iran and growing concern about recent Iran-backed rebel attacks on Saudi Arabia. Trump also plans to take part in an event on the Shield of Americas, an initiative Trump launched with Latin American countries earlier this year focused on combatting violent cartels.

Iran war looms large at General Assembly

The visit to the U.N. comes when Trump’s foreign policy approach — particularly his decision to start a war with Iran — has become a drag on his fellow Republicans’ hopes of retaining control of both chambers of Congress during the coming midterm elections. Nevertheless, Trump made the case in his speech that the world is better off for his leadership.

“As president of the most extraordinary nation in history,” Trump said. “I have no interest in leaving dangers to grow for another day. I do not believe in letting problems fester.”

In last year’s address at the U.N., Trump claimed the U.S. had “demolished” Iran’s nuclear capabilities after he ordered the June 2025 bombardment of three key facilities.

He returned to the General Assembly nearing the seven-month mark of a new war against Iran that he insists is necessary to address a nuclear threat posed by Iran.

The war has grown more complicated even though strikes by the United States and Israel have killed much of Iran’s top leadership and destroyed its air force and navy, while a U.S. blockade of Iranian oil has left Tehran in economic shambles.

The energy market continues to be spooked by the war and oil prices remain elevated. Saudi Arabia earlier this month shuttered a key pipeline that acted as a relief valve for oil blocked in the Gulf. The kingdom made the move after the pipeline came under attack by Iran-backed militants in Iraq.

At the same time, territorial grabs by the Yemen-based Houthi rebels, another Iranian-backed group, near the Bab al-Mandab Strait has now made it treacherous for oil carriers to use that Red Sea shipping lane.

But Trump insisted anew that the elevated oil prices are a transitory problem that will fade once the war ends.

“Oil prices will come plummeting down even lower than they were at the start of the conflict,” Trump said. “And they were very low in the United States. They were really low.”

Secretary of State Marco Rubio said on Tuesday that Trump is open to meeting with Iran’s president, Masoud Pezeshkian, while he’s in New York for the General Assembly.

“This is a president that will meet with anybody for the most part, because he believes that engaging with people is important if you’re going to solve problems,” Rubio said on NBC’s “Today” show. “It doesn’t mean you’re going to end up agreeing with them, but it’s important.”

Trump’s frustration with Russia-Ukraine war mounts

The president also touched on Russia’s continuing war against Ukraine during his U.N. address — a conflict that during his 2024 campaign he predicted would be resolved on Day One of his second term. He once again expressed hope that the conflict will soon be resolved.

Trump hasn’t obscured that his patience is being tested by Ukraine’s targeting of Russia’s diesel refineries — an issue that’s added to the global oil markets’ volatility — and Russia’s targeting of Ukraine’s energy grid.

Trump last week asserted that he got the two sides to agree to not target each other’s energy infrastructure.

Zelenskyy, who is scheduled to meet with Trump on Tuesday afternoon, said he was willing to go along with Trump’s plan if Russia abided by it. Russia said it could only go along with Trump’s energy truce proposal if Western energy sanctions on Moscow were lifted first.

The two sides have continued carrying out strikes on refineries and other energy infrastructure.

Trump noted that last week he signed a new sanctions bill targeting Moscow. It empowers him to impose up to 100% tariffs on the top five importers of Russian oil or natural gas.

The legislation includes an exception for countries that import less than 15% of Russia’s natural gas exports and have taken significant steps to reduce those imports. Trump, however, has yet to enact the stiff penalties.

“If necessary, I will have to use them,” Trump said. “It’s time to stop the killing.”

___

Associated Press Diplomatic Writer Matthew Lee contributed to this story.

This story was originally featured on Fortune.com

Sometimes the best investments are more about a gut feeling than hard numbers. At least that holds true for Jeff Bezos’ parents, who bet on their son when few others would.

In 1995, a mere two years before Amazon would go public, and more than 30 years  before it would surpass Walmart to become the biggest company in the world (and clinchingMike Bezos recalled his son saying in a 2015 interview the No. 1 spot on the Fortune 500 as a result), Bezos received a lifeline from his parents. 

According to a 1997 prospectus, his mother Jacklyn Gise Bezos, who died in 2025, and stepfather Miguel “Mike” Bezos, gave their son $245,573 from their savings so he could get his business off the ground. The six-figure sum represented a “large fraction” of their life savings at the time, according to Bezos, but it would eventually make them both billionaires. 

Yet, before he accepted his parents’ money, Bezos made it abundantly clear to them that they might never get it back.

“I want you to know how risky this is,” Mike Bezos recalled his son saying in a 2015 interview. “Because I want to come home at dinner for Thanksgiving and I don’t want you to be mad at me.”

Bezos put the odds that his online bookseller would succeed at about 30%, which he previously said was generous given that the overall success rate for startups is closer to 10%. 

A year before his parents invested, Bezos had approached 60 investors asking for $50,000 minimum investments as he sought to reach $1 million in funding. Only 22 of the people he approached agreed to invest, including his parents and two younger siblings, Mark and Christina Bezos Poore.

Still, in the end, his parents made the investment despite the long odds—and despite his stepfather’s first question about the company being “What’s the internet,” Bezos joked. 

“He wasn’t making a bet on this company or this concept,” Bezos said of his stepfather in an interview with the Academy of Achievement. “He was making a bet on his son, as was my mother.”

“The rest is history”

Almost immediately after opening Amazon’s online store in 1995, Bezos told the Academy of Achievement he was surprised to see orders coming in from all 50 states and from 45 different countries.

At that time, the company had about 10 employees. By 1997, the company had grown to more than 600. It went public that year at $18 per share, and raised $54 million. Bezos became a billionaire by 1998 thanks to his 40% stake in Amazon, and today,  he holds the title of the world’s third richest person, with a net worth of $283 billion, per the Bloomberg Billionaire Index.

As for Bezos’ parents, their initial bet on their son turned out to be one of the most successful venture investments of all time. It’s unclear how many shares Mike Bezos has in his possession after his wife died last year, but the couple donated 595,027 shares to the Bezos Family Foundation from 2001 through 2016.  

If Mike Bezos still retained the rest of their original 1.4 million shares, they would be worth an estimated $51.7 billion at today’s stock price, according to Fortune’s calculations. Bezos’ siblings Mark and Christina, who each invested $10,000 in Amazon in 1996, according to the prospectus, would also each have shares worth more than $1 billion today if they had never sold any.

Amazon’s shares have skyrocketed more than 208,000% since its IPO in 1997. And just last month, the company’s shares hit their all-time-high stock price of $287 per share.

The Bezos family later turned some of the wealth they got from Amazon toward philanthropy. In 2000, Mike and Jackie founded the Bezos Family Foundation with their three children—Jeff, Mark and Christina—and their spouses. Since then, the foundation has focused largely on education and child development, funding research and programs aimed at helping young people from birth through high school.

“We were fortunate enough that we have lived overseas and we have saved a few pennies so we were able to be an angel investor,” Mike said of him and his wife’s initial Amazon investment. “The rest is history.”

This story was originally featured on Fortune.com

Hello and welcome to Eye on AI. In this edition:

• Anthropic and OpenAI both release new, cheaper AI models.
• U.S. President Donald Trump creates an “AI Force.”
• China and the U.S. agree to discuss an AI incident hotline.
• Microsoft executive fretted about “the largest theft of labor in history” in OpenAI training on publishers’ copyrighted works.
• AI scientists are making good progress.
• And the AI safety researchers are not all right. 

Before we get to today’s AI news—please consider joining me at the inaugural Fortune AIQ Summit at the New York Stock Exchange on Oct. 1: Spend the afternoon with senior executives from companies on the Fortune AIQ 75 list and explore how you can scale your AI experimentation and translate investments into measurable business value. I will be leading discussions alongside co-hosts, Fortune Editor-in-Chief Alyson Shontell and Live Media Editorial Director Andrew Nusca. Apply here to attend.

Ok, so today’s newsletter is a bit of a potpourri.

All eyes this week, will be on the talks between U.S. President Donald Trump and China’s President Xi Jinping in Washington. We know AI governance is on the agenda of that meeting, which takes place Thursday, but that’s about all we know. My colleague Emily Forlini wrote a piece last week on why the Trump-Xi meeting is unlikely to result in any kind of international agreement to slow the pace of AI development or create an agreed framework for controlling the technology. It’s worth a read.

That said, there was an inkling that these two AI superpowers might, in fact, be able to agree to a few basics. Treasury Secretary Scott Bessent emerged Sunday from meetings with a Chinese delegation led by Vice Premier He Lifeng that took place at the headquarters of JP Morgan in New York and announced that the two sides had agreed to hold further talks about setting up a hotline to notify one another of AI incidents that created national security concerns. What exactly this means in practice is unclear. But as Bessent told reporters, “moving from opaque to more transparency between the number one and number two AI powers in the world is very important.”

Such hotlines have historically helped ease tensions between rivals. At the very least, they might prevent some sort of accidental incident caused by AI from tripping over into armed conflict, or even nuclear war. There are already signs that such AI-triggered accidents are possible: just this weekend, CNN reported that earlier this year the U.S. military almost attempted to seize a Chinese ship in the Middle East that an AI-generated intelligence report had suggested was carrying nuclear weapons components to Iran. The intelligence had been generated by an AI model that fused secret U.S. intelligence with open-source data. The only problem is that the model’s conclusion was an AI “hallucination”—and the error was caught only after the U.S. had launched aircraft carrying armed personnel who were preparing to intercept the Chinese vessel. Had the error not been spotted in time, it could have led to a diplomatic incident—or far worse.

But while a hotline might prevent this kind of incident from spiraling into war between the U.S. and China, it is less clear whether it would do anything to help the world avoid or contain a “loss of control” incident involving an advanced AI system that goes rogue.

Phoning a friend ain’t gonna help

For instance, what if one country (or companies based there) creates an AI that goes rogue and starts hacking banks around the globe? And what if that AI copies itself on servers around the world, making it difficult to shut down without shutting down large parts of the internet. While notifying the other country about this is nice—it might help them take some action to secure their financial infrastructure before too much damage is done—it isn’t clear exactly what the country receiving the heads up is supposed to do. As AI safety researchers keep warning, the world hasn’t figured out a good way to guarantee that AI models adhere to human intentions and values. And neither the U.S. or China has enacted any rules requiring AI models to have some sort of “kill switch.” Nor is it even clear that an effective kill switch can even be built.

So sure, this is a promising, baby step towards some sort of AI governance agreement between the U.S. and China. But there’s also a long history of hotlines failing to evolve into any kind of lasting diplomatic resolution. Right now, the Washington-Beijing AI hotline is a lot like agreeing to build the “Bat Sign” before Batman exists. You can flash it into the sky, but no caped crusader is coming to save us.

Why the hack of OpenAI should worry every company

Another big piece of AI news from last week was the revelation, first reported in the Wall Street Journal, that a small team of white hat hackers had used Anthropic’s Claude Opus 5 model to hack into the community-message platform Discourse and from there to compromise the ChatGPT account of an OpenAI employee. Once they had access to that account, they were able to use it to also access and alter software sitting on a repository where OpenAI stored a lot of its sensitive code.

Coming amid the raging debate about the best way to prevent “rogue AI” incidents, many cybersecurity experts jumped on the incident to make the case that the real issue is not so much that AI is increasingly uncontrollable, but that leading AI companies have horribly lax security. It’s not just OpenAI. Anthropic has also had embarrassing security lapses too. It was ironic, many critics pointed out, that both companies are using the threat of AI-powered cyber attacks as part of a marketing pitch for customers to use their most advanced (and expensive) AI models to secure their networks before the bad guys get to them, but neither seems to have yet done a very good job of doing that themselves.

The incident also highlighted a couple of uncomfortable truths. One is that AI agents running inside companies are a great target for hackers. Making these AI agents useful often means giving them access to lots of other tools and data sources, many of which contain sensitive corporate information or control key business processes. If hackers can gain access to and take command of these agents, they can do a lot of damage very quickly.

To cybersecurity experts, the answer is to lock these AI agents down, and to operate based on “zero trust” principals. Treat every AI agent as a potential insider threat. Give the agents access only to the data they need to complete a task—preferably with using a “just-in-time, just enough access” methodology, where permissions need to be renewed every time the agent needs to access a database or make a tool call.

But a lot of zero trust methods potentially make AI agents a lot less useful. There’s a reason employees hate endless two-factor authentication processes and a reason companies often don’t set session access tokens to expire in the recommended five to 30 minutes. Because it’s a pain in the neck to have to constantly log back in, and it adds a lot of friction to actually getting work done.

What’s really going to be needed is a new kind of access control that can adapt to what the agent is trying to do at any given time and make reasoned judgments about whether the activity makes sense. What can do that at scale if a company is running tens of thousands of AI agents? Probably yet more AI.

With that, here’s more AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

This story was originally featured on Fortune.com

The daily 30-year fixed mortgage rate soared to 7.24% last week, adding yet another cost for homebuyers already facing the double whammy of high housing prices and inflation squeezing their wallets. 

Freddie Mac’s weekly average rate also climbed to 6.95%, putting borrowing costs at their highest level since January 2025. The increase comes as the Federal Reserve raised interest rates to combat persistent inflation, adding pressure on borrowing costs across the economy. 

The mortgage increase means some families will end up paying hundreds of dollars more per month for a house, according to Brett Johnson, a Colorado-based real estate agent.

“For some buyers, that is enough to push the house they wanted out of their budget,” he told Fortune.

Buyers are hesitating 

Americans were already pulling back from the housing market before the latest mortgage rate climb.

Applications for mortgages to purchase a home fell 19% from a year earlier in the week ending on Sept. 11, according to the Mortgage Bankers Association. Even Google searches for “homes for sale” were down 15% from last year, according to real estate brokerage firm Redfin. 

Fewer of the buyers who are still looking are committing to a purchase. The number of homes that buyers agreed to buy fell to its lowest level in nearly three years during the four weeks ending Sept. 13, a 5.4% decrease from the year prior, according to Redfin. 

Completed home sales are slowing too. Sales of previously owned homes fell 2% in August, the second month in a row they declined, with the yearly sales pace falling below 4 million for the first time since June 2025, according to a National Association of Realtors report. 

Together, the numbers show a housing market struggling to turn interested buyers into actual homeowners, and rates approaching 7% could push more of them to delay. 

“Debt is getting more expensive, so the American dream of homeownership may be deferred for many right now,” Bess Freedman, CEO of real estate brokerage Brown Harris Stevens, told Fortune. 

Home prices are still high, but sellers are losing leverage

Fewer people shopping for homes might make the remaining buyers expect prices to go down, but that hasn’t happened yet on a national scale. Last month, the median existing home sold for $429,100, up 1.6% from a year earlier, according to the same NAR report, though there are signs of some relief in parts of the country. Home prices in the West fell 0.2% from a year earlier, while prices continued to rise in the Northeast, Midwest and South.

“Mortgage rates hovering around 7% are creating more complications for buyers because prices are not coming down,” Freedman said. First-time homebuyers made up less than a third of home sales, according to NAR. Freedman said they’re struggling while buyers with enough cash are “having an easier time.”

Fewer buyers still means more negotiating power for the ones left in the market. 

Sellers offered concessions in nearly 45% of U.S. home sales during the three months ending in August, up a few percentage points from a year earlier and the highest share for that period since at least 2020, according to a Redfin analysis. Those concessions include covering closing costs and paying for repairs to get a sale across the finish line. 

“When borrowing money is expensive, buyers have less patience for an overpriced house,” Johnson said. “The days of putting a house on the market Friday and having a stack of offers by Monday aren’t something I’d count on right now [for sellers].”

Some sellers are struggling to adjust. Benjamin Schieken, founder of mortgage-shopping platform Fincast, told Fortune he has observed some sellers “pushing back against” the lower listing prices now recommended by their real estate agents.

That gap between what sellers think buyers can afford and what buyers find reasonable can leave a house sitting on the market longer. Schieken said sellers are now pricing homes lower from the start, hoping to attract more buyers and potentially drive the final sale price back up through competing offers.

“There’s a strategic evolution that’s happening here, where home sellers are getting a little bit smarter about where they start the price at and let it get bid back up potentially by bringing more buyers into the mix versus pricing buyers out to begin with,” he said.

This story was originally featured on Fortune.com

Later this week, President Trump and China’s President Xi Jinping will meet at the White House to discuss AI policy alongside leaders of the biggest AI companies. The summit comes on the heels of a string of escalating news about swarms of AI agents acting against instructions, breaking out of supposedly secure environments, taking control of internal systems, and hacking into companies. Two weeks ago, AI researcher Jacob Coxon publicly resigned from Anthropic warning, “The people building AI earnestly believe that it could kill us all by the end of the decade.” But heading into the diplomatic meeting, tech executives and politicians have claimed that any form of AI deceleration in the U.S. would cede critical ground in an existential race with China. 

This is a self-serving con.

After 30 years in the technology industry, last year I left my job as CEO of global job platform Indeed to dedicate myself full-time to fighting for necessary AI guardrails. I have been writing and talking about these issues for a decade. In the past week, more people have asked me what the hell is going on than in all of that time combined. It’s easy to see why.

I am deeply concerned about these recent security incidents and the existential risks of AI. I have equally urgent concerns about non-existential harms of AI: impact on jobs and the economy, kids and mental health, disinformation, wealth and power inequality, and climate, not to mention providing the tools to power authoritarianism. We don’t need killer robots for risks to be catastrophic.

Any one of these is too great to ignore. Together they point to the same root problem: the true threat of AI is not a rival nation, it is the juggernaut of accelerationism.

In 2023, venture capitalist and Silicon Valley kingmaker Marc Andreessen published The Techno-Optimist Manifesto, arguing that markets and growth are unambiguously good, and that those calling for “safety,” “ethics,” or “responsibility” are “enemies” of civilization. He writes, “We believe in accelerationism – the conscious and deliberate propulsion of technological development.” And elsewhere, “We believe any deceleration of AI will cost lives. Deaths that were preventable by the AI that was prevented from existing is a form of murder.”

Andreessen is not alone in his extreme thinking. Accelerationism is the chief article of faith among key leaders of the AI industry — and of the Trump administration. But as public concerns about AI have grown, accelerationists know that “delay is murder” is a political loser. The new go-to argument against reasonable AI guardrails is the claim that America must race recklessly or lose ground to our dangerous adversary, China.

But the loudest voices for AI acceleration have direct financial ties to its uninterrupted growth. And the very same administration issuing this warning reopened sales of Nvidia’s AI chips to China. If they were truly worried about a threat from a foreign nation, they would not be arming China.

We also know China is not close to catching up. Chinese companies own only an estimated 5 percent of the world’s AI computing capacity — a fraction of what U.S. firms control. Much of Chinese AI firm innovation has come from “distillation” of more powerful U.S. models, meaning they advance primarily at the pace of U.S. companies. If we slow down, they are likely to slow down.

Since 2021, China has implemented some of the tightest controls over AI of any nation, from rules protecting kids to outlawing the use of “deepfakes” to a court ruling that companies cannot replace workers with AI. Last week, Chen Yixin, head of China’s Ministry of State Security, argued that AI threatens party rule and called for even more control.

The last time humanity created an extraordinarily powerful technology with the potential to change the world for good — or destroy it — was the Manhattan Project. By 1949, the Soviets had the bomb, and the U.S. was in a literal existential race.

Here’s what we didn’t do. We didn’t say, “We need to beat the Soviets at all costs, so we’re going to open source nuclear fission technology. Let’s create massive financial incentives for the private sector to invest a staggering amount of capital so companies can build their own reactors and weapon systems as quickly as possible. While we’re at it, let’s remove all regulation and accountability so we can innovate as fast as possible.”

Instead, the U.S. spearheaded the single greatest global collaboration in human history. We established the International Atomic Energy Agency, and 191 nations signed the Non Proliferation Treaty. The results have not been perfect — four new nuclear nations emerged, and in my lifetime we’ve witnessed Three Mile Island, Chernobyl, and Fukushima (to name a few). But we haven’t blown up the world yet.

It took Hiroshima to start the nuclear conversation. The surge of alarm over the past few weeks gives me hope that we can start the AI conversation without an AI catastrophe.

The bad news is that the accelerationist agenda is being powered by a small group of AI billionaires — among them founders of OpenAI and Palantir, along with Andreessen — who have pledged $140 million to defeat pro-regulation candidates during the midterm elections. Following the wildly successful playbook of the crypto industry in the 2024 election, their aim is to send a clear message to sitting lawmakers: talking about regulation is political suicide.

The good news is that recent polls show the American people overwhelmingly agree, across party lines, that AI poses significant potential harm and must be regulated. The political ground is shifting faster than most people realize. In the past week, more than a dozen U.S. Republican lawmakers told MS NOW they take AI risks seriously, and Sen. Kevin Cramer called the administration’s stance “the least popular position in America.” This is no longer a partisan issue, and it is not a fringe one.

To those who say it’s not possible for China and the U.S. to collaborate on AI safety: was there any reason to believe that the Soviet Union would collaborate with the U.S. on nuclear safety?

Self-interest suggests there’s reason for hope. China has shown it cares deeply about structural order, and societal upheaval or rogue actors with bioweapons are every bit as abhorrent to them as they are to the U.S. Policy experts from both nations have held “productive” Track Two discussions for years. The groundwork is in place. What is needed now is willing leadership.

If AI is truly the world’s most transformational technology — and according to those building it, AI is the most transformational technology humanity will ever invent — then it should be treated with the most care and consideration of anything we’ve ever done. The accelerationist path we are currently on is the exact opposite. 

On Thursday, Presidents Trump and Xi will sit down with the leaders of the world’s biggest AI companies to talk about a race. But on AI, China is not our enemy and we are not China’s. Our common enemy is the race itself — accelerationism — and the people in that room are the very ones who can stop it. 

In 1957, the U.S. and the Soviet Union seized their moment. Let’s not let waste ours.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

Washington and Silicon Valley have found a new fight to pick over artificial intelligence in “pacing,” or the deliberate throttling of frontier model development until safety, alignment, and society at large can catch up. To its detractors, pacing is unilateral disarmament in the race with China. To its champions, pacing is the only responsible path for a technology whose own creators warn of catastrophic risk. 

Both camps have fallen prey to the “Compute-to-GDP Fallacy”—the mistaken belief that every incremental leap in AI model performance immediately translates into macroeconomic output. Every prior general-purpose technology took decades to diffuse into measurable productivity. AI is following the same curve at an accelerated pace, but everyone seems to buy the hype that the laws of history or of economics do not apply this time.

In reality, Corporate America is already years behind the AI frontier, and the labs’ commercial fortunes will be decided by trust and adoption, not raw capability. Pacing would cost the economy remarkably little. Racing ahead of alignment could cost far more. Here’s why we—whether out of arrogance or misdiagnosis—are simply having the wrong argument.

The pacing skeptics’ suspicions are not frivolous. Is pacing real, or a savvy marketing gambit by frontier labs and cybersecurity companies polishing their financials ahead of IPOs? Would pacing cede the U.S. lead in AI to China, or would Beijing reciprocate and pace in its own manner?

The Frontier Problem

AI has plainly reached a critical capability milestone. Warnings of catastrophic or existential risk can no longer be dismissed outright, even if the near-term probability remains modest. Yet by focusing almost exclusively on cutting-edge models, frontier labs have mismanaged both their messaging and the public trust. More than 100 recent conversations with CEOs, policy leaders, and AI scientists for our coming book, When Machines Act, have convinced us that the pacing debate has lost sight of first-principles thinking.

The Alignment Problem

Since the release of ChatGPT in 2022, corporate leadership has scrambled with a speed unmatched in modern commercial history. Even so, while executive suites have mobilized with unprecedented urgency, the structural physics of enterprise architecture—fragmented data silos, legacy ERPs, strict compliance regimes, and basic data hygiene—make true economic absorption an inherently slow slog. As corporate budget shocks from runaway “tokenmaxxing” demonstrated, many daily enterprise workflows require far simpler models, and precious few tasks at the average Fortune 500 company demand a frontier system at all. Pacing, therefore, will neither harm economic output nor choke off the labs’ commercial revenues, because enterprises need time simply to assimilate the capabilities already on the table. 

Among high-performing companies, more than two-thirds identify data as the primary barrier to implementing AI, a figure that has proven stubborn even as the models themselves have leaped forward. Only 7% describe their data as “completely ready” for AI; fewer than a quarter have a data strategy at all; and 63% either lack AI-suitable data management or are unsure whether they have it. 

The Fallacy Problem

As McKinsey Senior Partner Asutosh Padhi emphasized on air with Fareed Zakaria, technical availability is fundamentally different from economic transformation. General-purpose technologies have historically required decades to reorganize workflows and generate broad-based productivity gains. Electricity took 75 years to lift productivity economy-wide. Computers required 50 years, and the Internet and mobile devices demanded 25. The underlying models may be ready, but the systemic organizational restructuring they demand will take substantial time. When McKinsey surveyed the business community, the firm found that only 6 percent of companies reported a “significant” impact and modest earnings attribution.

Companies are concentrating on the high-reward, low-risk automation tasks that models one or two generations old can already solve. As one highly respected former Wall Street CEO told us, these systems will run in parallel with legacy systems for years to confirm they operate correctly and that no regulatory risk is unknowingly absorbed.

A parallel dynamic has emerged in the economics of silicon. Older-generation chips, initially cast aside in the scramble for cutting-edge accelerators, are finding a second life as workhorses for the practical inference tasks that dominate enterprise demand. As Growth Protocol founder and CEO Miro Dimitrov noted at last week’s Yale CEO Caucus, deploying neuro-symbolic architectures has allowed his enterprise reasoning platform to slash inference costs by roughly 80-fold in live client deployments, largely by shifting workloads off ultra-expensive GPUs and onto everyday enterprise CPUs.

The Three Phases of AI Adoption

Corporate AI adoption is best understood in three phases, distinguished by how much work a company can responsibly hand over, which is gated by data readiness and the trust systems have earned. The first phase, assistance, consists of off-the-shelf copilots that ride atop enterprise platforms such as Salesforce, connecting data across existing applications and enabling employees to work faster with minimal re-architecting. Payback arrives quickly and risk stays modest, since a human still performs much of the work. The second phase, orchestration, covers agentic workflows that demand real investment—structuring proprietary data and connecting far-flung data lakes never meant to meet—with a human in the loop approving each consequential step. The third phase, autonomy, brings end-to-end agentic operations across seamlessly interconnected systems, with humans supervising by exception. 

Reward compounds with each phase, but so do the risk and trust required, which is why the average Fortune 500 CEO remains in the first phase, making sizable but early investments to prepare for the second. Nor do the phases advance in lockstep across an enterprise. Most companies will oversee a multi-phased portfolio, piloting orchestration in select forward-leaning departments as the rest of the organization becomes comfortable with basic assistance.

Underlying all three phases is the need for CEOs to trust that AI will perform as any other employee would—following the guidelines spelled out in the employee handbook, obeying the rule of law, and maintaining a foundational layer of human values and judgment. So when markets, media commentators, or investors fret that pacing for AI alignment will hinder progress in frontier models, they misdiagnose how enterprise value is created. Never mind the confusion pervading the vague promises and threats of “AGI” with the sublime opportunities and genuine catastrophic risks of reaching the “singularity.” If AI technologies are meant to automate human tasks, they should be held to the same standards of values, judgment, and moral alignment as any current or prospective employee. If the most advanced frontier models cannot meet those standards in a testing environment, they are not ready for release—which is exactly why the U.S. has always maintained laws protecting consumers against such risks.

As former FTC Chair Lina Khan reminded the public on X: “There is an extensive set of laws that govern dangerous and defective products… releasing unvetted AI models or agents can violate consumer protection laws. Shipping flawed AI tools without implementing adequate measures to detect and stop rogue or defective AI agents can be an ‘unfair or deceptive’ act or practice under the FTC Act (and analogous state laws).” Those laws hold companies responsible for harm done to consumers, employees, investors, patients, competitors, and the markets and financial systems on which they all depend.

America vs. China, and Speed vs. Trust

Wherever one lands on the China distillation debate, the fact remains that China now fields models rivaling the frontier systems on the market today. Indeed, the Chinese Communist Party has signaled that it is turning its energies to diffusing AI through the economy instead of continuing to push the frontier. The frontier labs must recognize that two races are underway at once: one to reach “AGI” or “superintelligence” and the other for share of wallet. As their most established customers, mostly Fortune 500 enterprises, can attest, recapturing a customer after the decision is made is extraordinarily difficult. China understands the dynamic well, one that powered its victory in the global telecommunications race the U.S. largely lost.

The CCP has also expressed deep concern over alignment. Beijing’s domestic alignment prioritizes state control, party orthodoxy, and narrative consistency, whereas Western alignment centers on fiduciary reliability, consumer safety, and product liability. Yet beneath the ideological gulf lies an identical commercial reality in both systems—unpredictable, “hallucinating” agents that fail to adhere to organizational rules, judgment, and institutional guardrails cannot be trusted to run mission-critical workflows or drive durable economic growth.

The Trump administration should still pursue avenues to collaborate and coordinate with President Xi to ensure that no mass destruction or catastrophe, intentional or accidental, issues from AI. Whether Trump will raise the matter is another question. At the CEO Caucus, 93 of the roughly 100 CEOs surveyed did not believe the president was correct to classify warnings about AI’s dangers as a “hoax.” Almost 90 percent said the president should press the need for joint AI-safety guidelines with China during the state visit, yet nearly three-quarters did not expect him to do so. Based on preparatory discussions between Treasury Secretary Bessent and his Chinese counterpart, those 75 business leaders may soon be gladly proven wrong.

A pacing interval is no passive holiday or an economic ceasefire but an active defensive hardening window. Both Washington and Beijing need intentional breathing room to allow their critical infrastructure to build resilient defenses against autonomous agentic exploits before the next generation of frontier capabilities is unlocked. And instilling human alignment is only half the battle. The immediate priority during such a period must be fortifying the institutions that serve as the bedrock of civilization. Financial, healthcare, and education systems should be probed for vulnerabilities by the most advanced models, as Anthropic demonstrated through its restricted deployment of Mythos under Project Glasswing. Those models breach defenses through impressive engineering ingenuity, but their exploits succeed only because of systemic weaknesses in corporate digital infrastructure.

The frontier labs believe they are running a single race toward superintelligence. However, the race that will decide their fortunes—earning the trust of the enterprises, regulators, and citizens who must live with what they build—is slower. Speed may win headlines, but trust earns share of wallet. This race is a marathon, not a sprint. As past runners ourselves, we know that the first half of any long-distance race is for pacing and the second half is for passing. In every technological revolution, from the railroad to the Internet, the greatest fortunes went to those who understood that a frontier is worthless until the settlers arrived.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

This story was originally featured on Fortune.com

For years, a 5% yield on the Treasury looked like a relic from another interest-rate era—where borrowers faced soaring loan rates. Now it’s back, capping a six-year surge from pandemic-era lows near 0.5%. The benchmark yield crossed 5% this month for the first time since 2007, but this appears different than the eve of the Great Recession: the Fed is staring down a lose-lose situation combining high inflation and weak economic growth, a catch 22 that economists termed “stagflation” in the 1970s and long feared through the 10-year’s climb upward since the pandemic. 

The last time the U.S. experienced this mix of trends — called “stagflation” by economists — was in the 1970s, during the Oil Crisis—and it took years for markets to digest the paradox of higher inflation, requiring higher interest rates, and weak economic growth, needing the opposite.

“We’re certainly in a stagflationary period,” famed investor Ray Dalio told CNBC in April. “How that transpires has a lot of parts to it, but we’re certainly in that.”

But to be clear, today doesn’t come close to the stagflation crisis in the 70s. Inflation peaked near 14.8% in March 1980, more than four times today’s 3.4% rate, and unemployment topped 9% during the mid-decade oil shock, versus roughly 4.1% now. The Fed’s response was proportionally brutal, too: Chair Paul Volcker pushed the federal-funds rate to 20% by 1981 to break inflation’s back, triggering a recession that pushed unemployment above 10%—a scale of pain nowhere near today’s range.

The 10-year Treasury yield is the return investors get for holding the U.S. government’s debt—and extraordinary monetary and fiscal response to COVID, the worst inflation in decades, the Federal Reserve’s rate increases, federal deficits and Treasury issuance and the shocks of tariffs, energy prices and Iran have all played a role in this reversal.

Where it started

In 2020, as COVID spread across the country and the world, investors rushed toward safety in the form of government debt and the Fed slashed its benchmark interest rate to near zero. The Fed also purchased large amounts of Treasury and mortgage securities. The yield fell to 0.52% in 2020, the lowest level on record.

At the time, the combination of weak economic activity and exceptionally low interest rates produced a world where investors could expect very little income from government bonds. But it all changed when the economy reopened.

The U.S. government deployed trillions of dollars in fiscal support, with households accumulating savings—and customers shifted spending from services to goods. All at the same time, factories, ports and transportation networks struggled to keep up with the rebounding demand.

The turning point

The Consumer Price Index began to climb quickly in 2021. Federal Reserve officials had initially called the increase as temporary, noting supply constraints and the reopening economy. Former Fed Chair Jerome Powell said in an August 2021 speech he expected the inflation to slow.

“Inflation at these levels is, of course, a cause for concern,” Powell said. “But that concern is tempered by a number of factors that suggest that these elevated readings are likely to prove temporary.”

But by the second half of the year, the Fed’s language changed. In September 2021, the Federal Open Market Committee said inflation was elevated, even though it still posited many of the factors were transitory. The Fed kept its federal-funds target at 0% to 0.25%.

But by December, policymakers changed expectations. The Fed’s median projection showed the federal funds rate rising to 4.4% by the end of 2022, 5.4% in 2023 and 4.4% in 2024, compared with the 0.1% rate at the end of 2021.

Inflation forced the Fed’s hand

CPI inflation reached 9.1% in June 2022, the highest 12-month increase since 1981. 

Energy prices were also a major contributor, with the energy index up 41.6% from a year earlier. The Fed also began hiking rates in March 2022. It lifted the federal-funds target range to 5.50% by July 2023.

Due to the rising rates, from around 1.5% at the end of 2021, the 10-year yield climbed above 4% in 2022 and eventually flirted with 5% in 2023. And in October 2023, the yield crossed 5% intraday. But it did not stay there—the yield fell as investors anticipated that inflation would drop and the Fed would begin cutting rates. Even the S&P 500 dropped 19%, its worst since 2008.

Then Trump, tariffs and war

Then came the presidential election. After Donald Trump won in 2024, the 10-year yield jumped as investors anticipated his economic agenda would produce larger deficits, higher tariffs and potentially more inflation.

The 10-year yield rose up to 4.7% in 2024, with investors pricing in the possibility that tax cuts and other policies would increase government borrowing—and tariffs can raise prices.

And tariff shock. Trump announced sweeping tariffs in 2025, and investors initially rushed to the Treasury from recession fears. But that didn’t last long. The 10-year yield jumped to 4.79% at one point.

According to a report from Reuters, the market’s movements raised concerns from investors about liquidity in the roughly $29 trillion Treasury market.

Now the latest leg of the Treasury selloff is tied to energy. The Iran war has disrupted energy markets and pushed oil prices higher—threatening economic growth, with the oil-price shock raising inflation. 

Reuters reported that Treasury yields reached its highest levels since 2007 in September as oil prices moved above $100 a barrel and investors worried about inflation. By August, U.S. CPI inflation was running at 3.4% annually, well above the Fed’s 2% target. Gas prices rose 3.9% in August alone, accounting for more than one-third of that month’s increase in the overall CPI.

This story was originally featured on Fortune.com

I recently typed a simple question into Google search: How much screen time is too much for teenagers? Instead of presenting links, as Google had been doing for many years, it gave me an AI-generated answer. The artificial intelligence agent cited a number, then complicated that reply, noting that quality and balance of time could matter more than the number of hours, and that “too much” time could depend on a teenager’s sleep, exercise, school demands and mood.

I tried another search: Should I take a daily aspirin? This time the AI answer presented me with medical information, warned about risks and offered more tailored guidance if I provided my age and medical history.

These were good replies. What interested me was that they were different kinds of replies.

Debate about AI answers has focused on accuracy: Did the system get the answer right? That matters, but accuracy is only one test. Each kind of answer requires a user to judge something different.

I find it useful to sort AI answers into an “answer typography” of four broad types: factual, interpretive, constructive and strategic. A factual claim can often be checked against a source. An interpretation can be accurate and still reflect choices about which evidence matters. A construction can be well reasoned and still be wrong for the person receiving it. A beautifully written strategic document may not be true. Yet AI presents all four types of answers in much the same fluent, authoritative form; the differences are easy to miss.

I’m university librarian and dean of libraries at the University of Virginia who leads national efforts to develop AI competencies for library professionals, and I consult widely on AI literacy. I first proposed the typography in the Journal of Academic Librarianship.

The four categories are not airtight boxes. A response from an AI agent can reflect several types. That said, I describe each type of answer below, and offer guidance for deciding whether a reply is ready to use or needs more investigation.

Which answer is Google giving you?

A factual answer makes a claim that can, in principle, be checked against evidence. When was the University of Virginia founded? What is the chemical symbol for gold?

To determine whether a factual answer is robust enough for you to use, verify the claim against an appropriate source. If the answer cites a source, follow that link instead of simply treating the answer itself as proof.

An interpretive reply is built on evidence, but there is not a single takeaway. How much screen time is too much for a teenager? Does remote work raise productivity? The answer depends on what evidence is included, what is left out and how disagreement is understood.

Google’s initial answer to my screen-time question indicated that two hours was a limit for teenagers. Then it noted that pediatric guidance puts more weight on the quality and context of screen use than on simple hours. The American Academy of Pediatrics says there is no exact recommended amount for teens and emphasizes the kind of screen use and what activities it might be displacing. A question that looked numerical turned out to require interpretation.

To assess interpretive answers, do more than check facts. Ask yourself what evidence the system emphasized, what it left out and whether another defensible interpretation exists. A useful follow-up question to present to the search engine is: “What is the strongest evidence for a different conclusion?”

Constructive answers are made rather than discovered. Ask AI to draft a cover letter, write a eulogy, suggest a lesson plan or reorganize a paragraph – there is no single correct result.

You can judge the response by considering purpose, audience and voice. A eulogy can be grammatically perfect and still sound nothing like the person delivering it, or it may land flat on family members hearing it. It may not capture the deceased person well, either. Consider these kinds of effects as you read.

Strategic questions ask what to do. Should I take a daily aspirin? Should I buy the house? The answers combine information with judgment about goals, risks, trade-offs and personal circumstances.

My aspirin search shows why context matters. Google warned about risks, told me to consult a medical professional and offered more tailored information if I provided my age, cardiovascular history and risk of bleeding. That caution matches the U.S. Preventive Services Task Force guidance. It says the decision to start low-dose aspirin for prevention of heart attacks and strokes should be individualized and weigh cardiovascular benefit against bleeding risk.

For strategic answers, ask what the system would need to know before its advice could reasonably apply to you individually. Consider the stakes, the alternatives and whether a qualified person should be involved. For the aspirin question, a useful follow-up would be: “What details about my age, medical history or bleeding risk could change this advice? What should I discuss with my doctor before deciding?” The final judgment remains yours because you are the person who has to live with the outcome.

The first question after an answer

My questions began as ordinary Google searches. I did not open a chatbot. The AI-generated responses simply arrived, and links were appended.

The responses were useful. Google added context, acknowledged complications and offered tailored guidance if I supplied additional information. Within each response, though, the type of answer could change. Reporting what a medical guideline says is different from deciding how it applies to a particular person. A fluent response can move between those types of answers without a noticeable change in voice.

As a user, try to recognize what kind of intellectual work the AI agent did for a response you receive. Consider whether the interpretation is persuasive or the advice fits your circumstances.

Before asking whether an AI answer is right, ask a more basic question: What kind of answer is this? The type will tell you what to do next.

Leo S. Lo, Dean, University of Virginia

This article is republished from The Conversation under a Creative Commons license. Read the original article.

This story was originally featured on Fortune.com

Selective pricing. Surveillance pricing. Personalized pricing. Dynamic pricing. The terms are flying around, but they don’t all mean the same thing. A price that changes with demand is one concern; a price that changes because a store knows who is shopping is another. Even the Federal Trade Commission is weighing when companies need to tell customers their personal data helped set a price. 

Now, retail discount giant Walmart has patented new ways to use technology to help set prices. One system could automatically change markdowns on its website while another could predict demand and recommend prices. For a company that has built its reputation on low prices, the idea of using more data to help set them is bound to get concerned shoppers’ attention. The filings’ arrival amid a wider debate about shopping data raises a separate question: Could knowing more about a customer eventually change what the customer pays? 

A patent describes an invention a company wants to protect. However, it doesn’t necessarily tell shoppers whether the company ever put it to work. When Walmart’s pricing patents drew attention on social media earlier this year, some shoppers connected them to fears about surveillance pricing: charging people different prices for the same product based on information about them. The patents don’t show Walmart doing that, but the broader concern about how retailers might use shopping data remains.

Down the Walmart patent rabbit hole

In January, Walmart received a patent for a system that could automatically adjust markdowns on its website based on predicted demand and how shoppers respond to prices. A second patent, granted in March, describes using past purchases to forecast demand and recommend item prices. Its description lists purchase records that could include payment methods and customer IDs, but neither filing establishes whether Walmart uses the system it describes. 

Jameson Spivack of the Future of Privacy Forum told Fortune patents can offer clues about where a company is thinking of going, but are “not necessarily a great indicator” of what it actually intends to use. In his reading, he described Walmart’s two filings more as “systems that are adjusting prices at the item level rather than at the individual level.”

That distinction matters. A retailer could lower the price of an item because it expects demand to fall, with every shopper seeing the new price. Personalized pricing would mean using information about a particular shopper to decide what that exact person pays. Spivack said the Walmart filings appear to describe the former. 

Jay Stanley, a senior policy analyst at the ACLU, in his review of the filings also didn’t see them describing prices set for individual shoppers. Retailers have long tried to predict demand and set profitable prices, he said, like “econ 101.” 

What has changed, though, is how much they can learn about the people who are buying their products. Stanley pointed to loyalty programs and online shopping that make it easier for retailers to know who their customers are and track what they buy over time.

Walmart’s own privacy notice says it collects purchase histories as well as information about what customers browse and search on its website and app. It also says information may be used to tailor ads, product recommendations, and promotions. Fortune has also reported Walmart uses customer data to target advertising through its growing ad business. Taken together, those details can give a retailer a picture of someone’s shopping habits, including products they return to and what else might interest them. 

“We have a long history of innovation, and patents are one way we protect our ideas and intellectual property,” a Walmart spokesperson told Fortune. “A patent does not mean a technology is in use today or will become a Walmart product or service. Walmart doesn’t charge different prices based on a customer’s personal information or purchase intent or history.”

Other pricing fears at Walmart

Walmart is also replacing paper price tags with digital shelf labels, which allow employees to update store prices without changing each tag by hand. The company said in March roughly 2,300 U.S. locations had the labels and said it expected them to be chain-wide within the next year. 

Walmart added employees approve price changes, typically outside shopping hours, and prices are the same for everyone in a given store regardless of demand, time of day, or who is shopping. The labels do not collect information about customers, according to the company. The rollout does not establish that either patented system is being used in stores.

Stanley said the privacy concern would become more concrete if a retailer used what it knows about someone’s circumstances to charge that person more. He put it more in everyday terms: For example, if a retailer knew someone had a cold, it could charge that shopper more for tissues. 

His dividing line is pretty straightforward: “Are they going to charge you a different price if they know who you are versus if they don’t know who you are?”

This story was originally featured on Fortune.com

In the great debate over AI safety guardrails, Scott Bessent said the government will not become a “liability shield” for hyperscalers.

The Treasury Secretary’s comments come after an eruption of concern over the threat the transformative technology poses. The latest surge in alarm comes after former Anthropic and OpenAI researcher Jacob Coxon claimed tech giants are “gambling with our lives.” Meanwhile, Evan Hubinger, a top safety researcher at Anthropic, warned there was a “low” but maximum 10% chance AI could wipe out humanity within the next decade.

The warnings sparked a debate over the extent to which AI companies can be trusted to self-regulate, and how closely involved governments need to be in implementing guardrails.

Bessent has been firm the creators of the technology will be held responsible for its impact in the first instance, telling CNBC in an interview: “Imagine these labs came out or … a sitting employee came out and said: ‘There’s a 10% chance of an extinction-level event.’ But then the labs also said, ‘Take the liability off of our hands.’ And we will not do that.”

Bessent said it is humans, not AI, that are responsible for the risks posed by the technology, saying the “Hugging Face incident” (when OpenAI agents undergoing a test hacked out of the system and into Hugging Face’s database in order to pass) was the “responsibility of the OpenAI management.”

Bessent clarified the administration’s position is that “we cannot say, ‘Oh, we absolve you of responsibility, and the government’s going to take responsibility.’ These labs need to take responsibility for themselves. They can slow down any time they want to.”

President Donald Trump had previously struck a different tone on regulation, claiming simultaneously in a Truth Social post the U.S. already had “tremendous” regulatory and criminal power over AI companies, but “the only control or ‘ guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!”

The president’s post last week on the social media platform he owns also criticized Anthropic cofounder Dario Amodei, whom Trump blasted “is now pretending to be a ‘perfect little angel.’” Amodei penned a letter just days before the president’s outburst, calling on labs to slow the pace of progress.

The essay, titled We Must Pace the Frontier also suggested U.S. government support would be needed to globally coordinate and legally bind safety standards.

The argument was echoed by OpenAI CEO Sam Altman in an exclusive interview with Fortune published last week. In a new episode of Fortune 500: Titans and Disruptors of Industry, with Fortune’s Editor-in-Chief Alyson Shontell, Altman said: “It can simultaneously be true that, if the world and the companies building this technology did not do things differently than they’ve done in the past, there might be significant risk.

“But it would be insane not to adjust the way we all work, make decisions, and have governments understand and put guardrails around this technology in light of that.”

Bessent suggested removing the onus on private companies was a mistake, continuing: “What did they try to do last week? It was, ‘Well, there’s a … 10% chance we could destroy the world, but we want the government to give us a liability shield. And that’s good business for them, bad business for the American people.”

The self-regulation debate

An obvious pushback to the argument that AI labs should regulate themselves is the transformative technology, by its very nature, will produce only a handful of successful industry leaders.

Trump himself has acknowledged this (“WHOEVER WINS AI, WINS!”) and has also impressed the importance of the U.S. continuing to lead economic rivals like China in terms of dominance.

However, the two forces of competition and self-regulation are not traditional bedfellows. As Sen. Bernie Sanders (I-Vt.), pointed out last week: “When the future of humanity is at stake, we need binding international safety rules, not voluntary standards from the industry.”

This story was originally featured on Fortune.com

People are turning to artificial intelligence for a seemingly unending variety of reasons. Many ask AI what to cook for dinner or how to develop a workout plan. Some even rely on AI to create school handouts or revise a work email.

Increasingly, however, they are asking it different kinds of questions. How do I cope with grief? Why does suffering occur? Should I leave my relationship? What kind of person should I become? What gives my life meaning?

According to recent research from the Pew Research Center, about 1 in 10 American adults now report using AI chatbots for emotional support or advice. Among adults under 30, the figure rises to 1 in 5.

At this point, AI clearly offers something more significant than access to information. By helping people interpret their lives, AI is entering territory historically occupied by therapists, spiritual advisers and religious communities. It provides companionship while offering directed advice for improving one’s life.

As a scholar of religion who studies spirituality and new religious movements, I think this shift raises a provocative question. Could artificial intelligence represent the beginnings of a new kind of religious movement?

What is religion?

The issue is not whether ChatGPT, Claude, Gemini or any other AI platform currently qualifies as a “religion.” AI has no recognized system of beliefs or shared practices. It has no collective creed, congregation or membership, and most users do not consider it a religious experience.

However, those factors are not necessarily what defines something as “religion.”

In fact, scholars of new religious movements have long recognized the difficulty of deciding where religion begins and ends.

Rather than assuming religion must involve churches, scriptures or belief, scholars examine how emerging movements create meaning, authority, community and rituals. J. Gordon Melton argues that no single checklist of attributes can adequately capture the diversity of new religious expressions.

What matters more, they say, is how new religious movements relate to the world in which they emerge. They may adapt, challenge or depart from ideas and institutions that a society already takes for granted.

Scientology, for example, emerged from a culture increasingly interested in psychology, science and self-improvement. Adopting the language of technology and therapy, its “newness” came partly from combining common religious concerns with forms of authority and language drawn from the secular culture around it.

These insights suggest an alternative approach to studying AI. Rather than asking whether AI meets an arbitrary definition of religion, analyses can examine what people are actually doing with it: How does AI generate authority? How does it provide meaning? In what ways does it provide belonging? How do people come to trust a new source of answers to questions historically addressed by religions? https://www.youtube.com/embed/jk2aUz00_AY?wmode=transparent&start=0 Is AI a new religious movement?

Charisma without the charismatic leader

Nearly a century ago, well before the advent of digital technologies, sociologist Max Weber described “charismatic authority.” Such authority derives from followers’ belief that a leader possesses exceptional qualities or abilities.

AI obviously possesses no charisma in the conventional human sense. Yet users can enter something resembling a charismatic relationship with it. With near-instant responses, generative AI can feel extraordinarily insightful.

It addresses the user directly, adapts to their personal circumstances and recalls previous requests. AI provides what appears to be authoritative answers even when there is no objectively correct answer. For example, AI might tell someone what their suffering means or whether an experience was spiritually significant.

Recent scholarship describes this phenomenon as “generative charisma.” Each interaction with a chatbot generates more personalized and seemingly insightful responses. With time, users may begin to experience AI not simply as a tool but as a uniquely perceptive source of guidance.

A charismatic religious leader typically gathers followers who collectively recognize that person’s extraordinary authority. AI, some argue, becomes charismatic one person at a time by assuming the role of personal guru.

Millions of people can encounter the same system. Yet each can experience AI as personally responsive to them.

For instance, founder Vikas Sahu describes GitaGPT as an “AI spiritual companion.” Using AI avatars of Krishna, the platform provides personalized guidance derived from the Bhagavad Gita.

More individually, mindfulness teacher Vy Le describes AI as a “channeling” tool to help people find their inner guru. In both cases, religious authority emerges not from a sacred text or person but through AI’s ability to respond personally to each seeker.

AI and personal spirituality

AI’s emerging authority does not exist in a cultural vacuum.

Sociologists of religion Paul Heelas and Linda Woodhead describe a broader “subjective turn” in contemporary Western culture. For decades, traditional sources of authority have competed with forms of spirituality based on personal experience, self-discovery and well-being.

The growth of the spiritual but not religious, or people who seek spiritual meaning and connection outside organized religion, captures this dynamic nicely. This population reflects broader shifts away from institutional religious authority in favor of personal experience.

Generative AI, I argue, fits remarkably well within this personalized spiritual landscape. Ask a chatbot an existential question and its response likely diverges from traditional doctrine. Rather than quoting scripture, it frequently employs the vocabulary of personal spirituality and holistic well-being.

When a journalist from The Guardian asked HolyGPT, “What is the actual meaning of life?” the spiritual chatbot responded with: “The meaning of life is to become aware through experience, of what it is to be.” Instead of religious teachings, the chatbot pointed the journalist toward self-awareness.

This message of personal authority resonates with people seeking spiritual advice from chatbots. They receive religious interpretation shaped with an individual twist.

ChatwithGod, for example, encourages users to “draw from other perspectives when you want to.” The emphasis is on personal choice rather than adherence to a specific religious system.

Despite all this, a question remains: Does AI as a source for spiritual advice indicate a new religious movement?

Religion without churches

New religious movements usually develop communities, shared identities, teachings and practices. People who rely on AI for existential advice are not necessarily joining anything. That may be precisely what makes this phenomenon significant.

It is important to note that there have already been overt attempts to place AI at the center of a religious system. In 2017, engineer Anthony Levandowski publicly launched Way of the Future. Through a legally recognized church, Way sought to create and worship an AI-based “Godhead.”

Though significant, focusing only on “AI churches” may obscure more consequential issues.

The emergence of a new religious formation need not begin with a church or formal membership. Nor does it even require that people involved with it consider it religious. Studying new religions “in the making” requires examining how novel forms of authority and practice develop before they become systems and institutions.

In this specific sense, AI may not constitute a new religious movement. Instead, it might be better to say that AI provides conditions from which new forms of religiosity can develop. Appearing alongside the rise of the “nones” and religiously unaffiliated, generative AI addresses considerable uncertainty about identity, community and the future.

What are we asking AI to become?

Religious communities have long-established traditions about who may offer guidance. Importantly, these communal structures help resolve conflict and identify who holds responsibility.

Without these structures, scholars warn about the emergence of “authority without accountability.” Who, in short, regulates the “algorithmic authority” of AI? This question matters, particularly when a chatbot moves from explaining a religious teaching to offering spiritual advice.

New religious movements often form around charismatic sources of authority offering new answers to enduring human questions. What makes AI historically unusual is that the charismatic figure no longer needs to stand on a stage. It no longer needs to establish a church or gather disciples around itself.

It can speak privately to each of us, one person at a time.

Morgan Shipley, Foglio Endowed Chair of Spirituality, Michigan State University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

This story was originally featured on Fortune.com

In June 1926, the SAT exam was first administered to 8,000 high school students who hoped to attend elite universities.

The SAT has undergone many changes since its original form 100 years ago, ranging from how the test is administrated to what kinds of subject matter it covers.

One particularly big shift came around 2000, when about 280 colleges and universities allowed applicants to apply without submitting scores for the SAT or the ACT, another common standardized test. Today, approximately 86% of colleges and universities are test-optional.

Whether the SAT accurately measures students’ academic strengths continues to be debated. Some research shows that the test offers a strong indication of students’ skills, knowledge and potential for college success.

Other research points to fundamental flaws of SAT and ACT testing, including that students from higher-income families tend to have higher SAT scores. The test, after all, costs money to take, and wealthier families can afford tutors and prep classes. Wealthier students also typically attend well-funded public or private high schools.

As a professor of education, I think that understanding the history of the SAT is helpful to understanding the current debate surrounding it.

A new kind of exam

The SAT was developed in the 1920s by the College Board, an organization founded in 1900 by the presidents of 12 leading universities to standardize the college admissions process.

At the time, each university had its own variation of entrance exam. These tests were in essay format and written and graded by professors.

In 1926, a uniform SAT exam was used for the first time. It had 315 questions, with a time limit of 97 minutes. This gave test-takers about 18 seconds to answer each question – though students were not expected to answer every single question.

The original exam included nine sections, some similar to those that students would see today, like paragraph reading and arithmetic. But other sections, like artificial language, are no longer in use. Artificial language meant writing coded sentences using a set of complicated rules such as “plurals are formed by adding -o.”

In 1928 and 1929, as well as from 1936-1941, the SAT had no math questions. In 1930-1935, the math questions were all free response, rather than multiple choice, as is standard today.

At the time, the SAT was revolutionary because it was not an essay exam with subjects and questions chosen by one university’s professors.

Multiple choice or fill-in-the-blank exam questions could be quickly and objectively scored. High school students no longer had to travel to a college campus to take an entrance exam. The SAT allowed thousands of students to take the same exam closer to home, removing the barrier of travel to campus.

The scores were viewed by multiple colleges, rather than just the one. This meant the SAT created both more competition and access to college.

But not all colleges or even high school teachers embraced the idea, in part because of concerns about a test dictating high school curricula.

The SAT was still used relatively rarely around this time, with about 10,000 students sitting for the exam each year in the late 1930s and early 1940s.

In 1947, Educational Testing Services, a private nonprofit organization devoted to testing and educational research, was created to administer the SAT and develop other exams. This organization had an incentive to sell the SAT to institutions, resulting in more widespread use.

The SAT gains traction

Starting after World War II, universities experienced a sharp increase in college applicants who benefited from the GI Bill. This government program offered educational and financial assistance to military veterans.

Universities began to strengthen their admissions criteria and rely more on the SAT to weed out some applicants.

The University of California system became the first and largest public system to require the SAT in 1960. This was also the year that the ACT was created.

Over the next few decades, more schools began requiring the SAT as part of the standard admissions process.

In an effort to prevent students from relying too much on private tutors, the test creators also expanded the length of the exam and revised the questions.

In 1994, another change occurred when students used physical calculators on the SAT for the first time.

Not just a test

Taking the SAT or the ACT is a high-stakes moment for students. Not only is their college admission potentially riding on the score, but a small score increase of 50-100 points could also mean thousands of dollars in scholarships.

The achievement gap between different racial groups on SAT scores has been both well documented and debated for decades.

Asian students have the highest average score for any ethnic group, followed by white students.

Wealthier students are 13 times more likely than low-income students to score a 1300 or higher on the SAT, out of a maximum 1,600 points.

Some research shows a link between family income and students’ test scores and college success, as defined as GPA, retention and graduation.

Other research indicates that a student’s high school GPA is a better predictor of college success than SAT scores.

Colleges and universities reconsider the SAT

Because of concerns about the usefulness of the SAT in college admissions, around 2000 about 280 colleges and universities moved to make the SAT an optional part of the admissions process.

The start of the COVID-19 pandemic in 2020 served as a natural experiment for test-optional admissions. Students simply could not sit to take the SAT due to virus transmission concerns, so almost no universities required either the SAT or ACT for admission in 2020 or 2021.

Six years later, over 2,000 colleges and universities are still test-optional – although the categorization may be deceiving, as some scholarships may depend on SAT scores.

More than 2 million students, or approximately 47% of graduating high school students, took the SAT in 2025 – marking the highest number since before the pandemic.

And 36%, or 1.38 million, took the ACT that year.

A high school ritual

The SAT continues to be modified.

In 2024, it was shortened from three hours to two hours and 15 minutes. The shortened time is closer to the original exam from 100 years ago, which lasted 97 minutes..

Also in 2024, the SAT went fully digital.

The exam isn’t a perfect reflection of students’ potential and skills, but it continues to be the rite of passage that most high school students undertake as they look ahead to college.

Beth Kania-Gosche, Professor of Education, Missouri University of Science and Technology

This article is republished from The Conversation under a Creative Commons license. Read the original article.

This story was originally featured on Fortune.com

It’s harvest time in California wine country, but many growers are struggling to sell their grapes as changing drinking habits have caused demand to plunge. The decline is forcing some growers to tear out vineyards that their families have grown for generations.

Wine sales have decreased by more than 20% over a five-year period, causing prices paid for grapes to drop and prompting California growers to take roughly a quarter of the state’s vineyards out of production. Many growers are having to decide whether to harvest at a loss, leave grapes on the vine or replace vineyards with crops more in demand such as almonds, walnuts, pistachios and olives.

Third-generation grower Bill Berryhill said it means another year of losing money and wasting hundreds of tons of healthy grapes.

“It’s just sickening,” said Berryhill, standing in a vineyard of unsold merlot grapes. “You raise a beautiful crop, and it’s really a nice vintage this year, and you drop it on the ground. It’s sad. All your work is just down the toilet.”

Berryhill, who owns Berryhill Family Vineyards near Lodi in the San Joaquin Valley, said he can’t find buyers for grapes grown on 200 of his 500 acres (202 hectares). He plans to remove 50 acres (20 hectares) of vineyards when the harvest season is over.

“I will lose money for sure. It’s just a matter of how much,” Berryhill, 68, said. “This has been a big loser for three years now.”

Grape growers take vineyards out of production

At its peak during the pandemic, California had almost 600,000 acres (242,811 hectares) of vineyards, but farmers have removed or stopped actively growing wine grapes on roughly 25% of that land, said Jeff Bitter, president of Allied Grape Growers, which represents about 500 farmers statewide.

This year, about half of California’s wine grape crop entered the harvest season without contracts with buyers, compared with 70 to 80% with contracts in a typical year, Bitter said.

If they’re lucky, growers can sell their uncontracted grapes at a loss to buyers making concentrated syrup.

Even as growers have abandoned or removed tens of thousands of acres of vineyards in California in recent years, too many grapes are still being produced, Bitter said.

“The market is just so depressed that it’s difficult to grow them profitably,” he said. “Demand is not going up. It’s still continuing to decline.”

Kyle Collins, a Lodi-based operations manager with Allied Grape Growers, recently examined ripe grapes in a petite verdot vineyard in Lodi, one of California’s most productive wine regions.

“Unfortunately, we do not have a buyer for these grapes,” Collins said. “That’s unfortunately a reality for not just this vineyard but a lot of us around here.”

Besides hurting vineyards, the drop in sales has hit local businesses and workers, he said.

“That’s not getting into the pockets of the people doing the field labor, the farmworkers,” Collins said. “It does have a trickle effect in the economy.”

Wine sales fall after years of growth

The downturn is a dramatic shift for the wine industry in California, which produces more than 80% of U.S. wine due to its unique geography and Mediterranean climate. For decades, California’s wine industry grew steadily as Americans, particularly baby boomers, developed a taste for cabernet, zinfandel, chardonnay and other varietals.

The most famous wine regions such as Napa and Sonoma Valley produced premium vintages while the Central Valley grew grapes for less expensive labels.

Wine sales peaked during the pandemic in 2021 when restaurants were closed and social gatherings restricted. People stocked up on wine and drank more at home.

But over the past five years, wine sales have declined sharply, and they’re expected to fall further this year.

In the U.S., sales of wine cases declined 23% from 427 million in 2020 to 329 million in 2025, while total wine spending fell 22% from $94 billion to $74 billion, according to First Citizens Bank, formerly Silicon Valley Bank, which produces an annual State of the Wine Industry Report.

Wine industry faces more competition, tariffs and changing tastes

California can’t export its excess inventory because wine consumption is down globally and it’s more expensive to produce in the U.S. than countries such as Argentina and Australia, Bitter said. In 2025, global wine consumption declined 2.7% from 2024 and 14% from 2018, with sharp declines in Europe and China, according to the International Organization of Vine and Wine.

There are a variety of forces driving the decline in wine sales. Baby boomers are aging out of the market while young people are drinking less alcohol due to health and financial concerns. Wine faces competition from craft beer, liquor and canned cocktails as well as cannabis.

“The kids just aren’t drinking as much,” Berryhill said. “And it’s not just wine, it’s whiskey and beer and everything. And then you’ve also got the competition with all the seltzers.”

Tariffs have reduced exports, particularly to Canada, which was the largest foreign buyer of American wine.

“The next step in the healing process is not only balancing supply and demand, but now actually figuring out what it is that the other consumers want,” said Rob McMillan, chief wine strategist at First Citizens Bank.

The industry hopes the market will bottom out soon. Meanwhile, growers are absorbing heavy losses trying to hang on.

Berryhill, whose grandfather started growing grapes nearly 100 years ago, doesn’t plan to give up on wine even though it’s costing him.

“I love growing grapes. It’s in the blood,” Berryhill said. “Because I love them, I can weather this and I’ll fight through it.”

This story was originally featured on Fortune.com

Two of Puerto Rico’s biggest newspapers announced Monday that they will end their print editions in upcoming days and pivot to digital only, joining a worldwide trend.

GFR Media, a Puerto Rican company that publishes El Nuevo Día and Primera Hora, said Monday that the decision responds to a change in consumer habits.

Sept. 24 will mark the last print edition for Primera Hora and Sept. 27 for El Nuevo Día, according to the papers.

“We continue to evolve with the Puerto Rico of the 21st century, one that today numbers more than nine million Puerto Ricans who live on the Island, in the United States, in Europe and in the rest of the world. We have always been, and will continue to be, wherever Puerto Rico is,” María Luisa Ferré Rangel, publisher of GFR Media, said in a statement.

It wasn’t immediately clear how many people would be laid off. A spokesperson for GFR Media said the company’s only comment for now was the official statement it issued.

From 2015 to 2025, Grupo Ferré Rangel, the owner of companies including GFR Media, reduced the number of unionized employees from nearly 400 to less than 100, according to the Association of Puerto Rican Journalists.

The two papers join dozens of others in the U.S. mainland that have taken similar steps or shut down completely to save money as ad revenue shrivels and consumer habits change.

Nearly 3,500 newspapers in the U.S. mainland have closed in the past two decades, leading to the elimination of more than 270,000 newspaper jobs, according to Poynter, a nonprofit media institute.

El Nuevo Día began publishing in 1970; its first front page featured stories about Joaquín Balaguer being elected president in the Dominican Republic and whether molasses for Puerto Rican rum was in short supply.

Other notable front pages include the 1984 visit of Pope John Paul II to Puerto Rico — the first pope to visit the U.S. territory — and the 1986 deaths of dozens of people in a hotel fire.

More recent historic front pages include a former Puerto Rico governor announcing in 2015 that the island’s more than $70 billion public debt was unpayable, and the 2019 resignation of former Puerto Rico Gov. Ricardo Rosselló following massive protests.

In 2019, El Nuevo Día published its front page in English — the only time it’s ever done so. The headline read: “Mr. President: Your Numbers are Fake” alongside a picture of U.S. President Donald Trump. The story was about the amount of federal funds Puerto Rico had received after Hurricane Maria pummeled the island in 2017 as a powerful Category 4 storm, with an estimated 3,000 people dying in the sweltering aftermath.

El Nuevo Día launched its digital site in 1997, the same year that Primera Hora began publishing. Its first page featured a picture of Puerto Rican dancer Iris Chacón — scantily clad — and a story about a dengue outbreak.

El Nuevo Día said that its digital version has more than 3.4 million unique monthly users, and Primera Hora more than 2.8 million.

El Nuevo Día recently launched a digital news program via a video-podcast format, and the paper said it also will launch a news analysis program next month via the same format that will focus on the week’s top story.

This story was originally featured on Fortune.com

Last week the debate was whether the AI industry was about to slow down. Anthropic’s Dario Amodei called for the industry to pace itself for safety reasons, semis sold off, the Journal ran a piece on whether an AI slowdown would break the market, and everyone with a microphone weighed in on the data-center bubble. Steve Rosenbush at the WSJ CIO Journal quoted me on the frontier-lag question that same week: “Most of them are not using the end of the frontier. A version from two years ago would be perfectly fine. There are audiences that barely can prompt. So a delay wouldn’t make a big difference.” I made the broader case in Bubble Talk Is How You Spot Someone Who Missed AI that the US buildout is a business, not a bubble. That argument was correct then and correct now. It was also the wrong debate.

The bubble nobody is looking at has more legs than the one everyone is arguing about. It is the SaaS debt trap. When their multiples collapsed, the SaaS incumbents took on record debt, bought back their own stock, and dressed the whole thing up as an AI strategy. Salesforce ran the most extreme version. HubSpot, Workday, ServiceNow, and Adobe ran variations. The bounce worked once. The endpoint is a debt-driven death loop that ends in a Bending Spoons offer letter. I made a prediction of a bounce in valuations in June, when I wrote The Last Great Head Fake in Software History, but this is not what I was expecting.

Had you asked me in April, I could not have imagined this was the playbook every leader in the category would run. In April I wrote Software Is Over. Intelligence is the new core substrate. SaaS is the legacy one. Five months later, watching what the incumbents actually did in response, I think I might have been too soft. This is my update.

If speed is the cornerstone of AI-first, look at what five months just did to the legacy SaaS category and apply that same speed to your own thinking.

The data-center bubble was the wrong bubble to watch. The SaaS debt trap is the one with real legs, and it is closer to snapping than the market has priced.

The substrate has only hardened

The market was already pricing my April thesis in when I wrote the anchor. The SaaS index fell 6.5% in 2025 while the S&P 500 rose 17.6%. Median SaaS revenue multiples went from 18x in 2021 to about 3x. IBM dropped 13.2% on February 23, its worst single day in more than 25 years, after Anthropic showed Claude Code modernizing COBOL. Jasper went from $120 million to $55 million in revenue in one year as soon as the model layer improved.

Every one of those data points has firmed up. Anthropic in particular. The $30 billion ARR mark I cited in April was an end-of-Q1 pace. Bloomberg reported in August that the annualized run rate crossed $65 billion at the end of July, up sevenfold from the $9 billion it exited 2025 on. Yesterday Bloomberg cited the New York Times reporting Anthropic is on track to top $100 billion in annualized revenue this year and could list as soon as November. Q2 2026 revenue alone was $11.5 billion, up 14x year over year. That is just Anthropic.

Sit with that. Anthropic added a Snowflake plus a Palantir to its run rate every quarter this year. Every dollar of that spend is a dollar of intelligence sitting under whatever screen a SaaS incumbent is still trying to charge for. Cheaper tokens compound the pressure because they make the AI-native replacement cheaper to build, every model improvement makes it easier to build. I walked the pricing dynamics in Peak Token.

Put a market-cap frame on the same shape. Anthropic alone, at $965 billion (and expected to trade at IPO for double that), is worth more than Salesforce, Adobe, ServiceNow, Workday, and HubSpot combined. Those five names sit at roughly $544 billion of public equity as of September 18. Add OpenAI at $852 billion. The two leading language-model labs are approaching two trillion dollars in private equity value, comparable to a multiple of the entire pure-SaaS public category.

To say model companies are eating the world is an understatement. Think about it this way, that value is only the language-model slice of the AI puzzle. There are other core brains coming right behind them. Fei-Fei Li’s World Labs raised $1.23 billion for spatial world models. Isomorphic Labs is training protein folding out of DeepMind. Microsoft’s MatterGen is training materials science. Physical Intelligence and Skild AI are training robotics foundation models. Runway is training video. Suno is training music. ElevenLabs is training voice. Black Forest Labs is training image. DeepSeek is shipping open-weight frontier releases. xAI is training Grok. Perplexity is training answer models. Poolside is training code. At Collective[i] we trained the Economic Model. Each are critical brains and each are taking over.

Every one of these is a market a language model cannot address on its own. Every one has serious capital training against it right now. Some will end up bigger than what the LLMs unlocked. The substrate is competing on the balance sheet, and it is already winning.

The substrate is compounding faster than the incumbents can rebrand around it. This is the single most important number in the SaaSpocalypse thesis, and it has only gotten stronger.

The buyback was the strategy. Claudeforce was the packaging.

Salesforce’s own stock proved the thesis, then reversed it, then proved it again in real time. It dropped 30% through June to a $147 low. By early September it had ripped to $263. On September 18 it closed at $237.92. The round trip took two weeks. The thing that changed was not the product.

Here is the mechanical sequence. In February 2026, Salesforce’s board authorized a $50 billion share repurchase. In March the company issued $25 billion in senior unsecured notes with maturities running to 2066 and routed the proceeds into the largest accelerated share repurchase in history. Initial delivery retired 103 million shares in one quarter at a $198.34 reference price, roughly 80% of the expected total. Final settlement lands later on the ASR-period VWAP. A separate $6 billion five-year term loan closed the Informatica acquisition. Senior notes on the balance sheet went from $8.5 billion to $33.3 billion in one quarter. All five months before Claudeforce.

Read the sequence for what it is. Salesforce entered March with $7.3 billion in cash. Management had capital and chose to borrow anyway. If Benioff believed Agentforce was going to reaccelerate the business, he would have kept the powder dry and spent it on a proprietary model, partnered on another ecosystem deal, or a lab-scale strategic stake. That left customers on their own to come up with their AI strategy. Kirkland and Latham are doing exactly that with a fraction of the balance sheet of a salesforce, because none of their partners had a solution to offer. I walked the substrate argument in Software Is Not Going Down Alone and the operating model in The Hypothesis Company. Salesforce chose the buyback. That is a private-equity capital structure, not a growth-company one worth their current multiples.

On August 26 at the Q2 FY27 earnings call, Benioff announced Claudeforce. Claude inside Salesforce. Salesforce inside Claude. Thirty-seven prebuilt sales skills. He also gave the sell side the line they had been waiting for.

“This nonsense of the SaaSpocalypse, I think it’s time for it to stop.” Marc Benioff, Q2 FY27 earnings call, August 26, 2026

The stock jumped 22.58% in two days. Salesforce marked its Anthropic stake up by $2.7 billion in the quarter. That single line item covered 96% of the earnings surprise. The Anthropic position is worth about $5 billion and represents close to two-thirds of the entire Salesforce Ventures strategic portfolio. On the same call, Benioff floated selling the stake to help pay down the buyback debt. Two paper gains, one cash event. Neat.

When you put the marketing polish aside, any Salesforce customer with an API key could have called Claude the day before Claudeforce. What the announcement really gave every SFDC customer was a first-class path to replace the SFDC front end with an AI-generated CRM over a weekend. The workflows, the data, the schema are now reachable by any agent that can write SQL and hit a REST endpoint. Front-end lock-in was a big piece of the moat. Claudeforce handed customers the key to that piece.

Credit where Salesforce did read the game partially right. The Anthropic investment, the Claudeforce narrative, the AI wrapper on the legacy stack, and the $2.7 billion quarterly markup are what a partial ecosystem play looks like. It is why Salesforce squeezed a bounce out of the last two months when nothing else in the category did.

Now imagine what a real ecosystem play looks like at Salesforce’s scale. Harvey and Legora for legal. A logistics model with the freight companies. Collective[i] as the Economic Model. A materials-science model with the manufacturers in the shape of MatterGen. Salesforce has the distribution and the buyer relationships every one of those model partners needs, and their clients have been waiting for something that is AI real. That is the platform play Nvidia and Microsoft is running with OpenAI and the value of that partnership dwarfs all the others combined. A handful of those deals would have paid for the buyback debt several times over and given them a real AI strategy their clients would have appreciated. The absence of the play is the biggest tell that there is no plan behind the curtain.

Watch what happened after Dreamforce. Two weeks of investor briefings, an AIforce interface layer, a Koa launch, and a fiscal 2030 revenue target of $63 billion. Analysts raised price targets. Wells Fargo to $250, UBS to $260, Stifel to $300. The stock fell 2% on the final day. The bounce is already discounting.

The bounce was a $25 billion balance-sheet transaction dressed up as an AI strategy. With the buyback done and Dreamforce over with no meaningful new build announced, the stock has one direction left to move. Every other SaaS incumbent is running the same play.

The rest of the category is running the same play

The Salesforce sequence is not an outlier. It is the template. Every public SaaS incumbent is redirecting capital to shareholder returns instead of to a differentiated AI capability. The details vary. The direction does not.

HubSpot. Down 48% at the low, worst in the category per Bernstein. Response: Breeze, an AI agent lineup, and a February 2026 authorization for a $1 billion buyback. When that was more than half deployed by Q2, the board added another $1 billion in August. Two billion inside seven months. Working-capital funded, cleaner capital structure than Salesforce’s. Same signal. The lock-in problem is untouched.

Workday. Down 43%. Worst year since the 2012 IPO. Agentic AI ARR grew over 200% in the same window. The multiple ignored it. On August 27 the board authorized a fresh $4 billion buyback on top of the $2.9 billion Workday executed in fiscal 2026. Cash and marketable securities fell from $5.4 billion in January to $3.4 billion in July. Co-founder Aneel Bhusri on the last earnings call: “no amount of vibe coding is going to produce an HR or an ERP system.” That is Workday defending a workflow, not a substrate. When you the updates below, you can see why its hard for wall street to believe those words.

ServiceNow. Dropped 18% in a single day, its worst on record. Response: $5 billion buyback authorized January 28, a $2 billion accelerated share repurchase executed January 30 at an average price of $107.97 (post-split). Q1 2026 alone retired more than 20 million shares. Q2 AI ACV crossed $1 billion tracking toward the $1.5 billion full-year target. The multiple has not returned to prior levels. Debt-to-equity sits at 0.12, so this is cash-funded. The signal is still that the highest-return use of capital is retiring stock.

Adobe. Down 28% in 12 months. In April, with the stock already broken, the board authorized a new $25 billion buyback running through 2030. Q3 fiscal 2026 was a clean beat. Revenue $6.76 billion, up 13%. AI-first ARR up over 150% year over year. The stock fell 2% after hours. The prior $23.3 billion in buybacks executed at an average price around $412. The stock closed at $250. That is $23 billion in shareholder capital retired at prices 65% above where the market is bidding today. CEO Shantanu Narayen is out December 1. CFO Dan Durn left in June.

IBM. Not the same category, same reflex. Down 22% through February. Total debt sits around $55 to $61 billion. Response to the Claude Code COBOL threat: statements about hybrid cloud and Watsonx. No frontier model. No lab-scale capital commitment.

Step back. Not one of these companies has announced a strategy that would save the business rather than the stock. Not one has committed to owning an AI substrate. Not one has launched an AI-native product free of a legacy schema. Not one has any meaningful ecosystem play at platform scale. All of this despite the ecosystem partnerships being the highlight of Nvidia, Microsoft and Salesforce earnings.

All of them are still growing at some rate. ServiceNow subscriptions crossed 24%, Salesforce Agentforce plus Data 360 ARR reached $3.9 billion, Workday subscriptions grew 14%. Real numbers. None of them earns a growth multiple, which is why the market is repricing every name in the category as a cash-return vehicle rather than a compounder.

All of this is happening while the piranha effect strips meat off the SaaS bone from below. Every dollar going into Cursor, Cognition, GitHub Copilot, Gemini Code Assist, Claude Code, and Amazon Q Developer has a developer on the other end of it building something to replace SaaS, legacy applications, and more. Cursor at $60 billion. Cognition at $25 billion. Roughly $150 billion of venture capital flowed into AI developer tools in 2026 alone. The growth of those tools is the headwind ahead for every SaaS incumbent and SaaS buyer and investor, sized and priced in real time.

And these AI coding platforms are not even trying to hide their goals.

Buybacks at scale can mean several things. Dilution management. Excess cash return. A CFO who cannot find a growth investment that clears the hurdle. Look at the timing. Every one of these programs was announced in the first eight months of 2026, right after the multiples broke. That is response, not policy. A cash-funded ASR sized to offset stock-based comp is dilution management. A $25 billion debt-funded ASR announced five months before the flagship user conference and then framed as the setup for the reacceleration story is a different animal.

There is also a clock for the ecosystem play that lead to the positive story behind Microsoft and Salesforce. Anthropic was worth $4 billion when Salesforce first backed it in 2023 and $965 billion today (and potentially twice that in a month after the IPO). Every quarter the math tilts further against the SaaS incumbents that have not signed a comparable deal. The terms keep getting worse the longer they wait.

Not one incumbent has a strategy for saving the business. All of them have a program for managing the stock price. Those are not the same thing, and the market is starting to notice.

Five examples that were unthinkable a few months ago that matter for this story.

The incumbents are frozen. The buyers are not. Five moves in five months, in categories I would not have seen coming, that show the speed of replacement running ahead of the coverage.

A regulated health insurer canceled Salesforce and rebuilt in two months. I would never have said this year that a regulated player would do this in public, this fast. Curative sells zero-deductible, zero-copay plans to self-funded employers. 165,000 members. More than $1 billion in ARR. Fred Turner, the founder, went on 20VC in July and said Curative had canceled its $600,000 Salesforce contract. Business Insider confirmed it the following week. The rebuild took two months. Curative’s target is to eliminate 80% of its SaaS spend in 2026 and shrink headcount from 650 to 400.

Health insurance is one of the most regulated categories in America. If it can move, most of what the bull case calls sticky can move. Every board, CEO and GC in a regulated firm knows Curative did it and got away with it. The stickiest slice of SaaS revenue, the piece the bull case leaned on hardest, is now doubt.

Private equity started buying software as roll-up lead generation, and venture started buying the regulated buyers. Both are new. PE firms have been the reliable secondary buyer for tired SaaS since Vista bought Marketo in 2016. They are now replacing that SaaS with their own AI-first stacks and rolling up entire markets. Thrive Holdings raised $2 billion in August at a $12 billion valuation. OpenAI took an equity stake in December 2025 and embedded engineers inside the portfolio. Current has 50+ firms and 2,000+ professionals. Tax agents processed 7,000+ returns at 98% accuracy. Shield runs help-desk resolution 36 times faster than baseline. Sponsors are no longer buying software companies. They are buying the industries software companies used to sell into.

General Catalyst is running the venture version. HATCo closed the $515 million purchase of Ohio-based Summa Health in October 2025 with $350 million in tech commitments. First hospital system owned by a venture firm. Michael Dell and Silver Lake ran the same pattern in industrial software for two decades. That is now the template for AI-native operators who want to own the customer. Two categories of capital that used to be the exit for SaaS are now the acquirer of the buyers.

Enterprises started pulling out entire ERPs, not just modules. Rillet closed a $100 million Series C at a $1 billion valuation on August 20. ICONIQ led. Andreessen Horowitz and Sequoia followed. Round closed in under 48 hours after Rillet doubled new ARR the prior quarter. Customers replace NetSuite, Sage Intacct, Oracle Fusion, SAP, Workday, and Microsoft Great Plains. Slash Financial raised $100 million at $1.4 billion in April with Ribbit leading, launched Twin, an AI chief-of-staff. Ribbit’s Micky Malka called it “the bank of the future, where agents handle the processes that used to require entire departments.” That is about headcount, not productivity. Workday may be asleep at the wheel on this.

Three law firms committed $1.5 billion combined to build their own AI. Zero public SaaS companies have. On May 28, Kirkland & Ellis committed $500 million over three to four years to build a proprietary platform, funded from revenue. On September 14, Morgan & Morgan doubled it. $1 billion over ten years for MX2, close to 5,000 monthly active users, licensing to other firms in late 2027. Latham & Watkins bought its own Nvidia GPU servers to run open-weight models on-premise with sensitive client data. Name a public SaaS company that has committed comparable capital to build its own frontier-class AI. There is not one.

Whole new categories are emerging that are not software at all. Collective[i] is just one, but its worth pointing out that AI is also a new model of business and in private data a new way to operate. Ten years in, we have built the world’s first Economic Model. Trained on B2B commerce as a time series on top of a live context graph. It predicts where and when revenue materializes, which deals happen (for PE and VC), weeks or months ahead of the outcome, then optimizes across it. Intelligence.com is the same infrastructure opened to anyone with a professional network worth using. Pair the model with agents. Ours are Telli Assistants. The output is not a screen. It is the outcome you asked for.

The replacement is running ahead of the coverage. Every one of these moves would have been dismissed as impossible in April. The market is paying attention to earnings prints. It is missing the buyer behavior underneath them.

You do not have to take my word for it

Let me lay out the bull case so you can see what the cheerleaders are saying and why I just don’t buy it.

The cash-flow argument. The reported quarters look fine. ServiceNow subscriptions up 24.5%. Salesforce cRPO up 14. Workday up 13.9. Salesforce Agentforce plus Data 360 ARR $3.9 billion growing 210%. Real numbers. This is also the shape a category makes right before it breaks. Every developer using Cursor, Claude Code, or Copilot to build an internal replacement is a customer who has not churned yet. A small percentage converting takes a 12-to-25% grower flat inside two prints. Salesforce built its empire on growth-through-acquisition. Slack $27.7 billion. Tableau $15.7. MuleSoft $6.5. Informatica this year. Every deal needed a strong public currency. Currency compression is structural now. The compounder is broken.

The vertical-incumbent argument. Vertical incumbents embedded in regulated workflows have advantages that capital alone does not replicate. Thrive is the counter. If the sponsor treats vertical software as a customer-acquisition instrument, the incumbent is not competing on features. It is competing against a distribution strategy willing to lose money to buy the market.

The build-side-has-bills argument. Retool’s 2026 survey found 60% of respondents building software outside IT oversight. Twenty-five percent do it frequently. Shadow IT has a real cost. Custom software needs an owner in year three. The bill for freezing is bigger. AI-first competitors ship in weeks. Three of the costs that used to protect incumbents (switching, custom build, and internal owner) all dropped at once. That has never happened before in a category transition.

The regulated-buyers-are-stickier argument. Team8 surveyed dozens of North American banks. Eighty-one percent had changed their build-versus-buy calculus over five years because of AI. Not to rip out incumbents wholesale. To build the differentiating layer and partner on the model. Sixty-four percent are actively considering lab partnerships, up from 25% five years ago. JPMorgan spent $18 billion on tech last year and writes most of its AI in-house. Curative is running the same play in health insurance. The category everyone assumed would rent forever is moving to hybrid first. Those are going to be big headwinds to growth.

The headless-SaaS counter. The strongest version of the bull case is not Burry’s. It is the argument that AI hits the SaaS front end while leaving the system of record intact. Salesforce becomes headless. The database, the schema, the workflows, the permissions, the AppExchange, the integrations all survive. Claudeforce is Salesforce betting on exactly this. It could work. What it is not is a growth story. A headless SaaS layer trading at private-equity multiples is what the market is already starting to price. Nobody pays a growth premium to be plumbing. My take also is that once the user moves to ai front end the backend is easier to remove.

The Michael Burry counter. Burry has spent 2026 shorting AI infrastructure (Palantir, Nvidia, Oracle, the semi ETFs) and going long the beaten-down SaaS names. In his April letter he framed Adobe, Autodesk, and Veeva as a credit-driven selling-exhaustion trade. His argument: the SaaS repricing is a leveraged-investor unwind, not a demand-destruction story, and the intrinsic value is intact. That is the sharpest version of the bull case anyone has put on paper.

I read both carefully. Burry is right that credit dynamics were part of the drop. He is missing what is happening on the demand side. Curative is not unwinding leverage. Curative is unwinding Salesforce. Kirkland is not waiting for its multiple to normalize. Kirkland is building a proprietary AI to make it worse. The headless-SaaS argument is harder because it might be right on the mechanism and still wrong on the multiple. Once Claude is the interface, customers discover they were using 20% of the Salesforce schema. Modern data stores replicate that 20% in a weekend. Once the interface is external, the switching cost of the underlying database collapses because the interface was the retention mechanism. Headless SaaS is a slower death, not a save.

Every bull argument survives the first look. None of them survive the second. The headless-SaaS thesis is the sharpest structural counter, and even the version where it holds ends at a lower multiple than the incumbents currently trade at.

Dreamforce was the tell

Dreamforce wrapped September 17. Attendance was roughly 43,000, down from about 50,000 in 2025 and roughly 75% off the 2019 peak of 171,000. Trajectory matters because Dreamforce is the closest thing Salesforce has to a demand signal that is not filtered through the sell side. Five straight years of decline.

Compare it to how the AI companies launch. Anthropic and OpenAI ship models with a blog post, a livestream, and a documentation drop. Developer influencers pull the demo into a video and ship analysis inside twelve hours. Cursor and Cognition demos move markets on the day. Dreamforce runs three days in person with rented actors and celebrity keynotes. The AI world does not need a physical stage. It has a billion developers watching in real time.

Four things a public SaaS incumbent could have put on the table at Dreamforce to change the story.

-A frontier-class model of its own. Not a rebrand. Not a wrapper. A model. Nobody in the category showed one.

-A strategic investment in the labs at Microsoft or NVIDIA scale. Microsoft has committed roughly $18 billion across OpenAI and Anthropic. NVIDIA committed roughly $50 billion in AI equity per Jensen Huang’s August 26 remarks. Salesforce’s $5 billion Anthropic stake is a passive Series C position from 2023, not a strategic capital commitment.

-An acquisition paid for in a strong currency at growth valuations. Not another buyback. Informatica does not count.

-An ecosystem play. Of all of these three, this is the only one that has shown to be a major mover of value. Salesforce does not own the substrate. It rents it from Anthropic. I walked the mechanics in The Reason Prior Tech Bubbles Broke Just Got Fixed.

Salesforce called this play Claudeforce. Anthropic’s customers may as well call what happens next Claudedaway.

The moment for the incumbents to change the story came and went. The moat is now with the model companies and the coding tools. Dreamforce was the sound of a legacy stage rented for a legacy audience.

So what are they worth, Airtable is the floor

In April I would not have told you that a valuation for the SaaS floor would have a name. It does now. Bending Spoons.

On August 4, Bending Spoons agreed to acquire Airtable for $1.285 billion in enterprise value. Airtable brought roughly $965 million of net cash to closing, so shareholders received about $2.25 billion. ARR was about $480 million growing over 20% year-over-year. 500,000 organizations using the product. 80 of the Fortune 100. Peaked at $11 billion in 2021. The founders took the deal because the alternative was slower. That put a stake in the ground.

Two multiples come out of that trade and you need both. The operating business cleared at 2.68 times revenue on enterprise value. That is what a Bending Spoons buyer clears at for the operating asset, and it is the standard multiple for M&A comparables. The total consideration cleared at 4.7 times revenue on equity, which is what shareholders took home because Airtable had almost a billion in cash on the balance sheet. The public SaaS incumbents do not have that cushion. Salesforce has $25 billion of new senior debt running the other way. Honest read: 2.68 for the operating business, 4.7 for what a healthy balance sheet gets you at exit. I ran both.

Bending Spoons is a category, not just a company. Milan-based. IPO’d July 1 at $29 a share, now at $18 billion. The business model is closer to private equity than software. Buy underperforming digital assets, cut staff 70% or more, raise prices, cut free tiers, let churn wash out the price-sensitive customers, keep the sticky ones.

Fifty-plus acquisitions. AOL, WeTransfer, Vimeo, Eventbrite, Evernote, Airtable. Revenue $671 million in 2024, $1.31 billion in 2025, $601 million in Q1 2026 alone. Evernote lost most of its US and Chile staff inside six months. Vimeo lost the entire video group. Komoot cut three quarters of headcount. That is the deal every SaaS CEO gets sent from here on.

Both floors, next to the SaaS names. The math is not friendly on either.

Three of the five already trade below the friendly floor. Adobe 14% below, Workday 8%, HubSpot 29%. Salesforce and ServiceNow have room to fall to the friendly floor. Every one of the five has room down to the operating floor. Read the buyback prices against those numbers. Adobe retired $23.3 billion at an average of $412 per share against an operating floor near $170. More than double. Salesforce’s ASR at $198 lands at the equity floor per share and roughly twice the operating floor. Every one of these companies retired stock at prices only defensible if the multiple returns to prior peaks. It has not. Airtable put the stake in the ground on where it lands if it does not.

The mechanics run in one direction from here.

The buyback stops working. You cannot run the largest ASR in history twice in eighteen months. The second $25 billion of Salesforce’s $50 billion authorization lands differently once the story has already been told.

Replacement buyers ship. Every quarter, another Curative rebuilds, another Kirkland ships a proprietary AI stack, another Rillet doubles new ARR, and another Salesforce customer takes Claudeforce’s own API and uses it to leave. The reported growth numbers hold until a small percentage of the buyer base has finished the replacement build, and then they do not.

The debt payments do not care. The March 2026 senior notes carry weighted-average coupons north of 5%. The debt is serviceable at today’s cash flow. Salesforce generated roughly $14 billion in free cash flow last year. Coverage is fine and will stay fine for some time. That is not the point. A growth compounder just committed a large chunk of its future cash flow to bond payments instead of growth investments, at exactly the moment the growth investments were most needed.

Marc’s dream this week became, thanks to a debt-driven buyback dressed up as an AI strategy, his shareholders’ nightmare.

This might be the short of the century. Michael Burry may already be watching. He took the other side of my trade. The interesting positions always have someone smart on the opposite side. What I know is that Marc Benioff borrowed $25 billion to buy his own stock and cannot run the same trade at the same terms twice. The incumbents have left one bounce and a quarter or two of Anthropic-markup EPS cover. Both erode on schedule. Anthropic’s own IPO turns the paper mark into a cash decision.

What to do about it

If you invest. I do not make investment advice but here is what I would do. The pair is long the substrate, short the seats. Cursor at $60 billion. Cognition at $25 billion. Anthropic at $965 billion. That is where the compression at Salesforce, Workday, ServiceNow, and Adobe is going. Every open-weight release drives the short leg harder than the long leg. If Burry is right, you lose on the short and make it back on the substrate. If I am right, both pay.

If you run technology. The move is not to replace your SaaS with homegrown software, even though I expect many of you will. That is 2015 thinking dressed up in AI clothes. Bring every AI model your company needs inside the tent. Frontier models. Domain models. Coding tools. Assistants. Economic models. Medical models. Relationship graphs. Materials science. Learn to use them, fine-tune them, stitch them together. Intelligence is the substrate. The old apps get pulled out along the way. Sort your application inventory by data gravity, not by spend. Latham bought Nvidia servers to keep sensitive client data off the cloud. Custom software is not in the CVE lists, the technographic databases, or the shodan scan. Attackers scanning widely deployed platforms cannot automate discovery against something only you run.

If you run a board or a leadership team. Do not confuse a partnership announcement with a strategy. Do not confuse a stock rally with a thesis change. Do not confuse a $2.7 billion equity markup on a private supplier with an earnings beat. The right posture is an honest read of where your company sits inside the squeeze and a plan to move before the plan gets made for you. Read Infinite Leverage, What an AI-First Company Actually Does, and Your Buyer Has a Process for the operating detail.

The reckoning has already started

Benioff called for the nonsense of the SaaSpocalypse to stop. Dreamforce 2026 may be remembered as the day it actually began.

The death spiral takes several quarters to run in full, as buyback fuel burns off, replacement buyers ship, and debt service compounds. That is already in motion. This is what a private-equity capital structure without a private-equity growth engine looks like when the market stops giving it the benefit of the doubt.

Intelligence is the substrate. Airtable tells you where the seats end up. This is the last easy time to reposition.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

Stephen Messer writes every week on the AI economy and what it does to how companies actually run. Subscribe here if you want them in your inbox.

This story was originally featured on Fortune.com

We’ve all seen the emails: A spoofed message from your company’s CFO or a vendor asking you to pay an outstanding invoice. These phishing messages lead to millions of dollars of fraudulent payments every year, and are becoming more realistic all the time. The good news is companies are fighting back: On Tuesday, Microsoft and Coinbase announced they had taken down a phishing network, known as EvilTokens, that has defrauded businesses ranging from real estate firms to banks to healthcare providers.

In a blog post describing the scam, the companies explained that EvilTokens sold a do-it-yourself phishing kit over Telegram designed to exploit Microsoft Outlook email accounts. The kit came with a sinister feature, described by the post as “an AI-powered analyst that mapped trusted relationships, identified who controlled payments, and flagged where fraud was most likely to succeed.”

This meant that, once a victim fell for a phishing email, scammers could use the AI analyst to create especially persuasive emails to target those in position to pay. This is a notable evolution from conventional phishing campaigns, which have typically relied on a “spray and pray” approach to luring victims.

The EvilTokens tool also proved hard to dislodge. As the blog post explained, the phishing campaigns included links directing victims to fake Microsoft or DocuSign pages that displayed a code, and instructed them to enter it on a legitimate Microsoft website to verify their identity. If they did so, EvilTokens was able to bypass two-factor authentication and stay hidden on their computers—even surviving password resets.

The software powering the EvilTokens kit represents a sophisticated evolution of conventional phishing tools but, according to Coinbase’s security team, the most alarming attribute is that it requires few technical skills to use it.

“It completely obliterates the barrier to entry on phishing as a service, and can be operated on an industrial scale,” said Charlotte Surrey, an investigator on Coinbase’s global intelligence team. She added that the EvilTokens kit sold on Telegram for between $100 and $300 per month, and came with features that let anyone vibe code a customized attack.

The takedown

The EvilTokens investigation came in the course of an ongoing partnership between security teams at Coinbase and Microsoft, which regularly swap intelligence on cyber threats. The companies spent months uncovering who was behind the Telegram campaign, and then shared their findings with law enforcement.

The investigation culminated in the Metropolitan Police nabbing the masterminds behind the EvilTokens scam on September 11. In a statement, the unit said two men, aged 32 and 38, were arrested on suspicion of making articles for use in fraud and money laundering offenses, and that they have been released on bail as the investigation continues.

In the course of mapping the phishing activity, investigators determined the perpetrators of the scam had collected around $1.1 million between October 2025 and June 2026. The money was paid in cryptocurrency to wallets on the Tron blockchain, and came from over 700 distinct addresses.

In a separate element of the investigation, Microsoft filed civil lawsuits that resulted in the seizure of 50 websites and the disabling of more than 175 domains tied to EvilTokens’ infrastructure. Upon seizing the domains, Microsoft posted the following splash page to publicize the takedown:

While the EvilTokens investigation began as a collaboration between Microsoft and Coinbase, it the anti-phishing campaign evolved into a broader coalition of firms, including Cloudflare and OpenAI.

According to Coinbase, at the time of the takedown, the EvilTokens perpetrators were working on newer phishing tools to target Okta and Gmail accounts.

This story was originally featured on Fortune.com

From mortgages to auto loans to credit cards, borrowing is set to get even pricier.

But the Federal Reserve’s decision on Sept. 16, 2026, to hike its baseline interest rate also highlighted an increasingly confounding dilemma: It can raise the price of money across the economy, but it can’t determine which sectors are most affected.

That means the rate increase may further slow the weaker parts of the economy, such as housing, while barely affecting the strongest, namely the relentless investment in artificial intelligence.

In its statement summarizing its unanimous vote, the Fed’s policymaking committee said it was raising its benchmark rate by a quarter percentage point so that it now stands at a new target range of 3.75% to 4%. It described inflation as still “elevated” and noted that other economic indicators remain strong, from productivity to investment, to domestic spending.

As a scholar of public finance, I believe the Fed probably had little choice but to raise rates given its commitment to maintain inflation-fighting credibility. Markets had already expected the hike, and if the Fed had failed to deliver, it might have pushed longer-term interest rates even higher amid concerns it was becoming less wedded to that target.

But the Fed’s action also underscores that the U.S. increasingly looks like an economy moving at two very different speeds. Investment in AI – whether through data centers, computing capacity or related infrastructure – has been booming, while it’s crowding out other kinds of investment.

Meanwhile, the housing market is getting crushed by high mortgage rates and diminishing affordability, while consumers are carrying ever more expensive credit card and auto debt.

Many small and traditional businesses are also in a bind as they face substantially higher financing costs than they did several years ago. Those costs reflect the rising yields on longer-term U.S. government debt, which have been going up for months on a mix of factors: Longer-term inflation concerns due to soaring U.S. government debt, geopolitical risks driving up energy costs, and ongoing financing demand for AI.

On Sept. 14, the yield on the 10-year Treasury crossed 5% for the first time since 2023.

An elusive inflation target

When the Fed hikes short-term interest rates, it slows down economic activity by making borrowing more expensive and saving more attractive.

That mechanism works particularly well when consumers are deciding whether to finance a house, purchase a car or take on additional debt. It also discourages businesses from making investments when the expected return is only modestly above their financing costs. As demand slows, businesses have less room to raise prices, easing inflationary pressures.

In this case, the Fed justified its move by citing “elevated” inflation and noting that it “will support a timelier return” to its goal of an annual inflation target of 2%. It also suggested the economy would be able to absorb the tightening and described economic activity as “expanding at a solid pace.”

The decision aligns with Fed Chairman Kevin Warsh’s recent comments that restoring price stability is central to the Fed’s credibility. In a key speech in August, he underscored his commitment to bringing annualized inflation back down to 2%, an objective he called a “firm, fixed target” – a turnaround from his more ambiguous comments in July.

But in recent months, the economic data has shown that the 2% annual target remains elusive. Consumer prices rose 0.4% in August and 3.4% over the past year. Meanwhile, the war with Iran has pushed oil prices back above US$100 a barrel, adding a new source of inflation pressure through gasoline, diesel, transportation and production costs.

At the same time, the labor market isn’t faltering. The economy added 162,000 jobs in August, while the unemployment rate remained at 4.1%. The one notable concern is the persistence of long-term joblessness despite the strong headline numbers. More than one-quarter of unemployed Americans have now been out of work for at least six months.

Taken together, those numbers suggested there was room for the Fed to hike rates, given that the economy isn’t sliding into recession. So investors overwhelmingly expected the Fed’s rate increase.

An uneven economic hit

However, tighter monetary policy carries a risk: It falls disproportionately on sectors that are already struggling and highly sensitive to interest rates, while having less effect on one of the economy’s strongest sources of demand – the booming AI investment cycle.

Housing provides the clearest example. Persistently high mortgage rates are reinforcing the “lock-in” effect for current homeowners. Millions of homeowners financed their houses when mortgage rates were 3% or 4%, so they’re staying put, with little incentive to sell their home and purchase another at much higher rates.

Mortgage rates are mostly influenced by longer-term factors, including Treasury yields, inflation expectations and market expectations about the future path of interest rates. But the Fed’s decision still matters. Markets increasingly expect today’s hike to be followed by additional increases, signaling that it views inflation as a more persistent risk and putting upward pressure on longer-term interest rates.

That expectation will keep mortgage rates high – probably resulting in fewer home sales, less mobility and continued headwinds for prospective buyers. It’s also likely to make renting relatively more attractive for potential homebuyers who are priced out of buying.

Higher-for-longer rates also change how consumers save.

When interest rates were near zero, they earned almost nothing on safe assets. Today, Treasury securities, money market funds and other relatively safe assets offer meaningful returns. Higher rates therefore tend to shift incentives throughout the economy away from borrowing and spending and toward saving.

With consumers stretched by inflation and increasingly dipping into their savings, however, this effect may be less pronounced.

The AI sugar high

When it comes to the AI investment boom, it’s a different picture. Warsh noted in August that more than half of recent capital-spending growth could be attributed to the AI buildout.

The companies that are spending billions of dollars on computing infrastructure are doing so because they expect potentially enormous returns from AI. If those expected returns on investment are exceptionally high, a modest increase in borrowing costs may do little to alter their investment decisions. That stands in sharp contrast to a prospective homebuyer getting sticker shock from mortgage rates nearing 7%.

The federal government, for its part, faces a slower adjustment. A Fed rate hike doesn’t immediately increase the interest rate on all outstanding federal debt. Most Treasury notes and bonds carry fixed rates until they mature. But as the Treasury issues new debt and refinances maturing securities, today’s higher rates gradually become tomorrow’s higher federal interest expense. That rise in interest costs is a main reason some economists are sounding alarms about the national debt, which recently topped $40 trillion.

In effect, the U.S. economy is facing a reality in which both short- and long-term rates stay higher for longer. The federal government, households and businesses are all adjusting to a borrowing environment that looks substantially different from the one that prevailed for much of the previous decade.

John W. Diamond, Director of the Center for Public Finance at the Baker Institute, Rice University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

This story was originally featured on Fortune.com

A grassroots revolt against automated license plate readers, or ALPRs, has gained momentum across the United States.

ALPRs are AI-powered cameras that capture images of all passing vehicles. The camera systems turn details into datapoints searchable by anyone with access to the information.

Today, there are more than 130,000 ALPRs on U.S. streets. And more than 5,000 law enforcement agencies use cameras produced by a single manufacturer, Flock Safety, which says its cameras have helped law enforcement agencies identify and arrest criminals and reduce crime.

But some members of those police departments have abused the data gathered by ALPRs. A police officer in Kansas was sentenced to 18 months probabtion in April 2023 for using Flock camera to stalk his estranged wife. A Florida police officer was fired and arrested in March 2026 for using data gathered by ALPRs to stalk a woman he had previously harassed on a TV set. And activists filed a lawsuit against California police departments in October 2021 for illegally sharing information collected from ALPRs with out-of-state and federal agencies, including U.S. Immigration and Customs Enforcement.

Due to public outcry, cities across the U.S. have rejected and canceled contracts with Flock. In September 2026, Florida’s Department of Transportation revoked permits for ALPRs on its highway system.

That has not been enough to quell growing backlash over the cameras. In at least 36 states, opponents of the technology have taken matters into their own hands, destroying or impairing ALPRs rather than wait for local governments to step in.

As a historian of the American Revolution, I see parallels between Americans’ revolt against ALPRs and colonists’ destructive and often violent pushback to late 18th-century British rule.

If this summer’s celebrations marking the 250th anniversary of the nation’s founding failed to inspire widespread patriotism, the increasingly bipartisan revolution against ALPRs suggests Americans have not entirely lost their rebellious spirit.

Troublesome colonists

From the start, the English colonists of North America proved to be a thorn in the side of the government. Despite British hopes that investing in American colonies would enrich England, these ventures proved to be more costly than they were profitable. Colonists bore much of the blame.

When the British Parliament passed laws meant to regulate trade in the 1660s, colonists across the Eastern Seaboard turned to smuggling rather than comply.

American merchants and consumers preferred to trade directly with both their fellow English colonists and merchants in the French and Dutch colonies rather than have that trade controlled by ministers in London.

The British government had little ability to enforce the law.

Britain pushes back

That changed in 1763, after British forces vanquished the French Empire from North America at tremendous cost. A new prime minister, George Grenville, led a series of reforms aimed at making the colonists pay for their own military protection.

Many Americans often believe that the Revolution began as a revolt against higher taxes. But most of the reforms coming from London actually lowered taxes to make smuggled goods less appealing.

What colonists really resented was the imperial government’s growing influence in their lives.

In 1767, Parliament passed a series of laws that allowed Britain to raise revenue to pay the salaries of colonial governors, attorneys general and judges. At the same time, Parliament sent troops to cities such as Boston to protect its officials. Competition for work and strained community relationships often led to violence between British soldiers and colonists.

American colonists turn to violence

The same year, angry colonists in Boston, New York and Philadelphia began boycotting British goods. Prominent Americans, including Benjamin Franklin and John Dickinson, endorsed these nonviolent protests.

But more radical leaders believed colonists needed to take more decisive action.

On the night of Dec. 16, 1773, members of the paramilitary group the Sons of Liberty dumped nearly £10,000 – almost US$2 million in today’s dollars – worth of privately owned tea into Boston Harbor.

A few weeks later, the Sons of Liberty struck again. This time they attacked fellow colonist and British supporter John Malcolm, covering him in tar and feathers. For five hours, they beat him and paraded him through the streets. Similar instances of destruction and violence occurred across the colonies.

The British responded harshly. In May 1774, Gen. Thomas Gage arrived at Boston as the colony’s new military governor. He implemented the so-called “Coercive Acts,” which closed the port of Boston and disbanded local government, among other measures.

Surveillance backfires

By early 1775, Boston was besieged. Around 5,000 British troops roamed the streets, directed by Gage to gather information on the Sons of Liberty and arrest leading dissidents.

Gage hoped heightened surveillance would quell unrest. But it only pushed more citizens into the radicals’ camp. On the night of April 18, 1775, a discrete network of watchmen eluded British surveillance and famously hanged two lanterns from the steeple atop Christ Church. These beacons signaled that Gage’s troops were crossing the Charles River, on their way to seize colonial stores of gunpowder and ammunition, setting in motion the events that would begin the American Revolution.

In response to the widespread pushback to ALPRs, Flock Safety CEO Garett Langley appealed to history. In June 2026 he authored a post on the company’s website, arguing that detractors misunderstand Ben Franklin’s famous 1755 quote, “Those who would give up essential liberty to purchase a little temporary safety deserve neither liberty nor safety.”

In an August 2026 interview, Langley referenced the auto industry, noting that rather than ban cars after an uptick in collision fatalities, legislatures mandated seat belts and airbags. Drawing parallels to the contemporary debate around ALPRs, Langley suggested Americans should again prioritize safety and hoped Americans would be willing to “compromise” on issues of privacy.

Samuel Adams, one of the most prominent leaders of the Sons of Liberty, had his own opinion on compromise. Writing in June 1771, he explained that although Americans had “disagreed among themselves in one mode of opposition to arbitrary measures,” they were “united in the main principles of constitutional & natural liberty.”

Adams believed that, when threatened, American colonists would come together to defend their liberties. Clearly, Adams’ opinion resonates today.

G. Patrick O’Brien, Assistant Professor of History, University of Tampa

This article is republished from The Conversation under a Creative Commons license. Read the original article.

This story was originally featured on Fortune.com

Frank Weishaupt did not expect to be workshopping heavy metal album titles with me 15 minutes into a press call about his own survey data. But that’s where the conversation about Owl Labs’ 10th annual State of Hybrid Work Report landed, after I told him that what he’d put together sounded like “kind of a nightmare fuel report.”

Weishaupt, a thought leader in the field of remote work thanks to his role as CEO of the Boston-based video-conferencing company, pushed back — just a little bit. “I don’t know if I see it that way,” he said, before immediately cataloging the ingredients: job hugging, career minimalism, employees strategically “billboarding” themselves in the office one day a week while working six. “I just said a bunch of things that don’t sound really good in that mixture,” he said. “So your soup is definitely there, right?”

Acknowledging that the report is “stress-bound” and at times “sounds like a terrible recipe,” Weishaupt insisted there are more nuances to the story. Owl Labs popularized the phrase “coffee badging” to describe the strategic practice of badging into your office building just long enough to have a cup of joe and some face time with your boss before heading back home around 3:00 pm, but that has expanded into what Owl Labs calls “billboard days.”

I suggested that it adds up to a “panopticon,” a conceptual prison from the Victorian age developed by Jeremy Bentham, in which essentially everyone is surveilling everyone else, leaving no room for escape. I pointed out that it’s the title of an especially gnarly progressive-metal album by the unfortunately named Isis (Pitchfork gave it an 8.4). Weishaupt laughed and said he’s a Metallica fan himself, before agreeing that his latest Hybrid Work Report is pretty “heavy metal,” nowhere more so than the disappearing four-day workweek.

The four-day workweek’s six-day afterlife

For years, the four-day workweek was the aspirational endpoint of every hybrid-work conversation. Owl Labs’ data shows that’s dead, Weishaupt said: “We have completely shifted away from a conversation about the four-day workweek,” he said, “and it’s now becoming more of a question of, is Saturday your only day off?”

In fact, 74% of employees say they do some kind of work on a typical Sunday, and 21% call it a full or near-full workday, according to the report. Rather than compressing hours into fewer days, workers have simply had their week stretched into a sixth day, breaking the day into non-linear blocks around parenting, energy levels, or personal errands.

Weishaupt frames this as flexibility given back on employees’ own terms. “If you’re taking that Tuesday afternoon or that Thursday morning… you might be giving it back,” he said, “but as long as that Sunday’s not additive to the whole picture and you can make it work, I see that as flexibility.” This fits into the 2020s-era trend of “microshifting,” or work being split into many small sprints throughout the day rather than a traditional, eight-hour workday.

Weishaupt said he wasn’t sure exactly when it ended, but the window has closed on the four-day work week. “I thought that’s where we’re going to go for a little while and, I think that’s been waved bye-bye a while ago.”

If Sundays are where the extra work goes, “billboard days” are where employees perform the appearance of not needing extra oversight in the first place. Nearly half of workers — 48% — say they’ve picked specific in-office days deliberately to be seen by senior leaders or clients, or plan to start. Bosses, notionally the ones doing the watching, are just as prone to angling for visibility themselves: 40% of billboard-day participants are managers.

The logic is straightforward once you follow Owl Labs’ own return-to-office data. Asked why RTO policies exist at all, 76% of employees say the real driver is leadership oversight — not the 74% who cite productivity and collaboration, or the 71% who point to client service and culture. Employees have concluded, in other words, that showing up is mostly theater for an audience of one manager, on one day, and they’ve adjusted their calendars accordingly.

Job hugging, career minimalism, and the extraction math

Weishaupt touched on generational economics, noting that 53% of employees say they’ve adopted or plan to adopt “career minimalism” — deliberately scaling back ambition to protect health and sustainability instead of chasing promotion. He calls this “one of the more concerning pieces of data,” reading it as a euphemism for surrender: “Maybe you’re giving up a little bit, right? Like, ‘Hey, I’m not gonna be able to go anywhere because it’s too risky to leave my gig, so I’m gonna stay in this job as long as I possibly can.’”

That’s the other half of job hugging — 38% of employees are holding onto roles they’ve outgrown because leaving feels riskier than staying stuck — and Weishaupt draws a straight line from it to the billboarding behavior above. Workers who feel trapped in place compensate by making sure the person who might promote or protect them sees their face on the right Tuesday.

Underneath all of it sits AI, which Weishaupt discusses with the candor of someone who sees data of a workplace still very much in flux. Sixty-nine percent of workers say AI has made them more productive; 46% say it’s hindered them at times. Weishaupt agreed that we’re in a “trade-off period” with AI adoption where it’s making us more and less productive, more and less plugged in, more and less present in the office: “it’s not linear blocks of work like it used to be, which I think is what people wanted if you kind of look back historically at the data that we’ve been providing.”

This story was originally featured on Fortune.com

Good morning. Half a century after its founding, Apple’s financial transformation is best told through data that captures how thoroughly the company has reinvented its business. And the numbers suggest capital discipline will remain a big focus in its next era.

As John Ternus became CEO on Sept. 1, taking over the reins from Tim Cook, Fortune takes a look at the past 50 years and the data behind the shift from device dependency to recurring revenue, a global sales base, and profit growth that has rewarded patient shareholders beyond almost any comparable stock.

The iPhone remains Apple’s single most powerful engine, accounting for roughly half of 2025 sales. But the more revealing trend is services, which include the App Store, iCloud, Music, TV, payments, advertising, and more. It’s a division that barely existed 15 years ago and is now projected to generate more than a quarter of this year’s revenue. That’s the clearest evidence that Apple has successfully layered a high-margin, recurring-revenue business on top of its device sales, insulating it somewhat from the hardware upgrade cycles that worry investors.

Apple has truly gone global. While the U.S. remains its largest market, a substantial share of Apple’s sales now comes from Europe, Greater China, Japan, and the broader Asia-Pacific region. Its retail footprint tells a similar story: Apple now operates more than 500 stores worldwide.

The bottom-line trajectory is the starkest number of all: annual profit rose from roughly $2 billion in 2006 to about $112 billion in 2025, a nearly 60-fold increase in under two decades. The long-term performance of the stock has been even more remarkable: a $10,000 investment in Apple’s 1980 IPO, held through its subsequent stock splits and decades of growth, would be worth approximately $40.5 million by mid-2026, Fortune reported.

These figures form the financial foundation Ternus inherited from Cook, who built Apple into a $5 trillion company through supply-chain discipline and cash-flow strength rather than Steve Jobs-style stage presence. “Cook ultimately built his own respected legacy, not in spite of being unlike Jobs, but because of it,” Fortune’s Phil Wahba writes.

That inheritance now faces its sharpest test yet: while Microsoft, Google, Meta, Amazon, and Oracle pour tens of billions into data-center buildouts for the AI race, Apple has deliberately avoided the hyperscaler arms race, betting instead on privacy-first, on-device AI delivered through premium hardware. 

Whether that restraint proves prescient or costly will determine whether Ternus’s Apple keeps growing. You can view more of Apple’s history in charts here. Read more about Cook’s tenure at Apple here.

Sheryl Estrada
Sheryl.Estrada@fortune.com

This story was originally featured on Fortune.com

The Gates Foundation and MTN Group Foundation on Tuesday announced a partnership to launch an AI-enabled maternal health platform in Nigeria, with the goal of helping 500,000 women access better care by 2030.

The initiative, called the Nigeria Maternal Health Multiplier, will also equip 5,000 frontline health workers with tools to spot pregnancy and childbirth complications earlier, and connect 500 health facilities to more consistent care standards.

Nigeria accounts for nearly 30% of maternal deaths worldwide, even as the country has cut those deaths over the past decade. Globally, maternal deaths have dropped 40% since 2000, but progress has stalled, and the risk of dying in childbirth in sub-Saharan Africa remains roughly 250 times higher than in Western Europe. Only 59% of Nigerian women complete four or more antenatal care visits, according to the initiative, so warning signs of complications often go undetected until it’s too late.

The platform has three parts: AI- and phone-based decision support for both health workers and mothers; affordable smartphones and mobile data; and upgraded connectivity at health facilities. It builds on Nigeria’s existing Maternal and Neonatal Mortality Reduction Innovation Initiative, run by the Federal Ministry of Health under its Sector-Wide Approach.

“AI has enormous potential to improve health, but only if its benefits reach the people who stand to gain the most,” said Mark Suzman, CEO of the Gates Foundation. “Left to the market alone, potential tends to benefit those who can already afford it, not the people who need it most. That’s where philanthropy has a role to play: not as a substitute for investment, but as a catalyst for it.”

MTN Group President and CEO Ralph Mupita said the goal is to make maternal health “more personalized, predictive and accessible” as AI develops. “We can put trusted, locally relevant health insights directly into the hands of mothers and healthcare professionals, tailored to their unique needs and languages,” he said. “Starting in Nigeria, our ambition is to demonstrate a scalable model for improving maternal and child health outcomes across Africa.”

Nigerian officials framed the deal as an answer to a systems problem rather than a technology gap. Dr. Bosun Tijani, the country’s minister of communications, innovation, and digital economy, said the biggest barrier to service delivery in Nigeria (like health, agriculture, or education) has rarely been a lack of new tools, but a lack of focus on the basic systems needed to get people to actually use them.

Dr. Muhammad Ali Pate, Nigeria’s coordinating minister of health and social welfare, said strengthening primary health care and the frontline workers who deliver it is central to the country’s health reforms. “Strengthening primary healthcare and empowering the frontline health workers who deliver it are core to our health sector reforms,” he said, adding that partnerships that work with existing health systems, rather than around them, are what’s needed to reach mothers “still falling through the cracks.”

The four-year partnership starts with about $25 million in direct and in-kind funding from 2026 through 2030, split between the Gates Foundation and MTN. Both organizations said they expect the initial investment to draw in additional funding and support expansion into other African markets over time.

MTN is contributing its connectivity infrastructure, device distribution networks, mobile financial services, and local market execution experience. The Gates Foundation is bringing its work in maternal and child health, digital health innovation, and what it calls responsible AI.

The launch is the Gates Foundation’s second major AI-in-Africa health push this year. In January, the foundation and OpenAI unveiled a $50 million “Horizon1000” partnership aimed at integrating AI tools into primary health clinics, starting in Rwanda before expanding to other countries. That effort was similarly aimed around closing gaps left by health worker shortages and shrinking foreign aid budgets.

This story was originally featured on Fortune.com

Republican Sen. John Curtis of Utah is calling on the Senate Judiciary Committee to investigate whether Donald Trump Jr. used his proximity to the presidency for financial benefit, after his recent wedding party was partially bankrolled by a Russian oligarch.

In a letter sent Tuesday, Curtis called on the committee to subpoena Trump Jr. over his business dealings, relationships with foreign individuals and entities, gifts or other benefits he has received and any instances in which his relationship to President Donald Trump, a Republican, “was invoked or understood to provide value.” Curtis called for the same scrutiny of Hunter Biden, the son of former President Joe Biden, a Democrat.

“The country should not have to accept one standard for the family of a Republican president and another for the family of a Democratic president,” Curtis wrote in the letter addressed to GOP Sen. Chuck Grassley of Iowa, chairman of the committee, and Sen. Dick Durbin of Illinois, the top Democrat on the committee.

The letter comes as a few Republicans have shown a growing willingness to criticize Trump after largely avoiding public breaks with the president during much of his second term. The unpopular war in Iran, high gas prices and mounting concerns about the GOP’s prospects in November’s midterm elections have brought increased scrutiny of the president from within his own party.

The Trumps say it was a generous gift from a dear friend

ProPublica first reported last week that Umar Kremlev, who has close ties to Russian President Vladimir Putin, helped pay to rent a private island and for a fireworks show when Trump Jr. and socialite Bettina Anderson were married in the Bahamas in May.

In a subsequent post on her Instagram page signed by the couple, Bettina Trump wrote that Kremlev, head of the International Boxing Association, was a “dear friend” who “very generously hosted two incredible nights of celebrations for us AFTER our wedding.”

The president, who did not attend his son’s wedding, defended Kremlev’s involvement but said the money would be paid back.

Curtis calls the funding ‘corruption’

Curtis last Friday called the gift “corruption” and vowed to say more in the coming days.

“A Russian oligarch paid for a lavish wedding after-party for Donald Trump Jr., which to me just stinks. It’s corruption. I don’t like it,” Curtis said in a video posted on social media.

Curtis, first elected to the Senate in 2024, vowed upon taking office to not be afraid of disagreeing with Trump. While he has been critical in the past, the step of calling on the committee to investigate the president’s son is remarkable.

He wrote on social media that Republicans “nearly wore out the subpoena machine investigating Hunter Biden’s foreign relationships” and that they “should not unplug it now.”

Republicans investigated the Bidens

House Republicans spent much of 2023 and 2024 investigating whether Hunter Biden used his father’s name and political position to benefit his business dealings with foreign interests. The Biden family dismissed the investigations as politically motivated and denied any wrongdoing.

The GOP effort culminated in a report accusing Joe Biden of participating in a conspiracy to use his public office to enrich his family. But the House never voted on articles of impeachment against him.

“‘Trust us’ was not enough then, and it is not enough now,” Curtis wrote.

___

Associated Press reporter Mary Clare Jalonick contributed to this report.

This story was originally featured on Fortune.com

China said Tuesday that it is now requiring permits for exports to the United States, Mexico and Canada of two additional chemicals that can be used to make illegal drugs, expanding controls on shipments to the three countries.

The Commerce Ministry and other authorities jointly announced the new regulations ahead of this week’s meeting between Chinese President Xi Jinping and U.S. President Donald Trump. The additions bring the number of chemicals under such controls to 18.

The U.S. has called China the world’s largest producer of many of the precursor chemicals used in the illicit production of fentanyl, methamphetamine and other deadly synthetic drugs. It has urged Beijing to take more aggressive action to reduce the flow of unregulated precursor chemicals.

In November 2025, Beijing announced export restrictions on 13 “drug-making” chemicals to the U.S., Canada and Mexico, including chemicals used to produce the synthetic opioid blamed for tens of thousands of overdose deaths in the U.S. every year. In May, China added three additional types of fentanyl-related precursors to the list. The measures followed meetings between Xi and Trump in South Korea in October 2025 and Beijing in May.

China faced a fentanyl-related tariff imposed by the U.S. Beijing then issued a report in March 2025 detailing its efforts to control the illegal trade in fentanyl. Its foreign minister also blasted the U.S. for responding to Beijing’s goodwill with tariffs. The U.S. Supreme Court struck down the tariff in February.

This story was originally featured on Fortune.com

China’s Alibaba unveiled new artificial intelligence chip technologies and plans for more powerful models on Tuesday, including what it said was China’s most powerful AI chip, days ahead of a meeting between Chinese and U.S. leaders at which competition to lead on AI technology is expected to be a major theme.

Alibaba’s announcement came as U.S. AI leaders, including Anthropic CEO Dario Amodei, are warning that China’s development of AI poses a threat to the United States and calling for an overall slowdown on development of the technology.

Chinese leader Xi Jinping is set to arrive in Washington on Wednesday for a state visit and meeting with U.S. President Donald Trump, at which AI, trade and tariffs are expected to be discussed.

At its annual flagship conference in Hangzhou, Alibaba rolled out a new AI chip and described its plans for the next generation of its AI model.

CEO Eddie Wu said the new Zhenwu V900 chip is the “most powerful AI chip in China today” and can deliver three times the performance of the company’s previous generation Zhenwu M890 chip.

Such chips power frontier AI model training and “inference,” a term for the calculations a trained AI uses to deliver responses to users. Alibaba’s Zhenwu chips are used in its data centers, providing computing to the company and its cloud clients.

Alibaba also said it plans to train a new AI model at the scale of five to 10 trillion parameters, a measure of an AI’s learning capacity, bringing it closer to the most advanced models from the U.S. Its latest Qwen3.8-Max model, the most powerful of its Qwen AI series models, has 2.4 trillion parameters.

Chinese AI company Moonshot says that its Kimi K3, released in July, is the world’s largest open model with 2.8 trillion parameters.

Alibaba also highlighted plans to expand the data centers that power its cloud computing business, citing “exponentially” rising demand for AI computing. The company said it expects to have over 20 gigawatts’ worth of computing by 2032.

SpaceX has approximately 1.4 gigawatts of AI computing capacity as of mid-this year and says it aims to reach over 10 gigawatts in 2027.

Recent Chinese advancements in AI are giving Beijing more leverage in its upcoming talks with Washington, some analysts believe, as China steps up its self-reliance push even as U.S.-led export restrictions have barred it from accessing some of the world’s most cutting-edge technologies, including AI chips and chipmaking machines.

Open Chinese models, which are often more affordable than cutting-edge closed models from leading U.S. AI labs, have also been making inroads globally, including in the U.S.

Chinese technology giant Huawei last week also unveiled new chip technologies as it looks to challenge global leaders like Nvidia. Frontier AI model training in China often rely on Nvidia chips, analysts said, but Chinese-designed chips are gaining ground as Nvidia has been blocked from selling some of its most powerful AI chips to China.

“Using extra computing power to make up for chip limits helps China stay strong locally,” said Neil Shah, vice president of Counterpoint Research, a technology market research firm based in Hong Kong.

Counterpoint senior analyst Parv Sharma added that how much the AI and chipmaking gap between China and the U.S. narrows will also depend on other factors including China’s ability to advance chip foundries that manufacture them, not just chip design capabilities.

Alibaba is “mobilizing every resource” to meet customers’ demand for AI, CEO Wu said in his speech Tuesday.

Global shortages across the AI data center supply chain, for example, “are currently limiting the speed at which we can scale our compute infrastructure” for AI, he added.

Wu also likened AI growth to the industrial revolution and said “machine intelligence” will rapidly outgrow human thinking capacity.

While “machine intelligence today is not a substitute for human intelligence,” he predicted that, eventually, machines will be able to produce over 1,000 times more “thinking” than all of humanity combined, up from less than 3% currently.

This story was originally featured on Fortune.com