But the senior at Virginia Commonwealth University began to see it differently as she looked into graduate psychology programs that explore human behavior in the workplace, which is quickly being upended by machine learning. Maeba, 21, is now pursuing a minor in AI.

“It’s giving me an edge and standing out,” she said.

Hiring has cooled for entry-level software developers — work increasingly done by AI agents — and college enrollment in computer and information science programs has been declining. Yet at campuses across the country, many professors are finding themselves busier than ever teaching students from a range of majors about artificial intelligence.

Colleges are responding to changes in student demand, but they also recognize that new graduates — regardless of their field — are facing questions about their AI skills from potential employers.

“We have to democratize it,” said Peter Stone, the chair of computer science at the University of Texas at Austin, who recently developed an introductory course on AI essentials for noncomputer science majors.

“In the same way that everybody needs some degree of math, reading and writing, I think everybody needs a degree of AI literacy,” he said.

Schools like VCU are standing up AI minors targeting students outside computer science. There’s a new AI graduation requirement at Purdue University. In freshman writing classes, Harvard University is teaching students how large language models work, plus exploring AI topics such as copyright and disinformation. And Ohio State University has an AI fluency requirement that includes hands-on workshops. Meanwhile, AI-focused undergraduate and even graduate degrees are proliferating.

Computer science faculty pivot to teach non-majors

At Northwestern University, noncomputer science majors had interest in taking classes in the department even before ChatGPT launched in 2022, but sometimes there wasn’t enough room, said Samir Khuller, the chair for computer science. That’s because computer science, then seen as a golden ticket to a high-paying job, was exploding in popularity.

But recently, the number of incoming computer science majors has started to shrink. Now Northwestern’s computer science staff, which had doubled in size to meet the demand, is teaching more students from other majors. The school, which already has a popular AI minor, is adding an AI major. It’s also streamlining prerequisites to make it easier for nonmajors to take an AI or machine learning class.

“Overall, classes are going to be a lot more accessible moving forward,” Khuller said. The department is so busy, he’s trying to hire three more computer science professors.

Among other changes across the university, Northwestern’s Bienen School of Music is offering a certificate in music and artificial intelligence.

Nationwide, enrollment in computer and information sciences continued falling this spring, down more than 8% at four-year institutions from the spring of 2025, according to the latest data from the National Student Clearinghouse Research Center, a dip that came as job postings for software developers dropped.

Interest in attaining at least a bare-bones understanding in computer science — and particularly AI-enhanced coding — has exploded across different majors. Music students, for instance, are discovering tools that can easily do tasks such as editing or generating drum tracks.

At Eastern Mennonite University in Harrisonburg, Virginia, assistant music professor Benjamin Guerrero is co-teaching a class next year with a computer scientist. Musicians, artists, theater and digital media majors will learn alongside math, computer science and electrical engineering students. The goal, he said, is to create people who can work at the intersection of art and technology.

“In my mind, AI is no more disruptive than the record player, the radio, the metronome, the synthesizer or the computer,” said Guerrero, who has helped organizations including the National Association for Music Education develop plans to incorporate AI.

Even before this latest surge in interest, STEM majors — like physicists and biologists — often took computer science classes to learn how to code for data-heavy research. But the learning curve was steep until new AI tools came along, said Murtaza Ali, a doctoral candidate at the University of Washington whose research focuses on computer science education.

Yet, he said, AI also is stirring anxiety. The concern, he said, is that even if the AI is just being used to write code, it’s going to encroach so much that eventually some college graduates won’t understand the basic concepts in their field.

“When you ask the AI to do it for you, on the surface it might just seem it is writing the code,” he said. “But under the hood, it’s actually also doing the understanding of the task for you.”

AI has prompted unusually fast changes in course offerings

Despite those concerns, the pace of change is remarkable for a sector infamous for moving slowly, said Paul LeBlanc, a visiting scholar at Harvard’s Graduate School of Education, who studies AI’s impact on higher education.

“It’s impossible to keep up with the technology, but that’s true for every organization in the world right now,” said LeBlanc, formerly president of Southern New Hampshire University, a prominent player in online degrees.

Community colleges that want to offer non-credit classes are able to move especially fast, because the approval process is less cumbersome. One in suburban Kansas City recently offered a non-credit “vibe-coding” class to teach nontechies to create working software.

“What AI does is shrink that distance between a complete novice and an expert,” said Grant Carlson, program coordinator for workforce development and continuing education at Johnson County Community College. The Kansas City-area college also offers an AI certificate and is making plans to tailor AI classes to specific occupations, such as human resources and medical careers.

At VCU, it used to take a year and a half to design a course and get it approved for students to take, said Andrew Arroyo, senior vice provost of academic affairs.

“Now we’re doing things really in a matter of months and in some cases weeks,” Arroyo said. An online graduate certificate program in applied AI is being added this year, he said.

More than simply creating programs, professors must try to prepare students for widely differing goals for using AI in their careers. In courses for VCU’s AI minor, Althea Pappas said she’s exploring AI not to use in her career, but to help her answer philosophical questions.

“What happens when we create actual life, or how do we even make that distinction?” asked Pappas, a music composition major entering her sophomore year.

Meanwhile, what fascinates aspiring neurologist Makenzie Stovall is the idea that AI is trying to emulate the brain — an organ whose inner workings continue to leave researchers dumbfounded.

“If I am going to try to make people’s brain the best it can be, well, then,” the 20-year-old VCU biology major asked, “why not study AI?”

____

The Associated Press’ education coverage receives financial support from multiple private foundations. AP is solely responsible for all content. Find AP’s standards for working with philanthropies, a list of supporters and funded coverage areas at AP.org.

The cloud unit posted $42.2 billion in revenue in Q2, up 37% from $30.9 billion a year ago, marking its fastest growth in 18 quarters, and what Amazon CEO Andy Jassy called its fifth consecutive quarter of accelerating growth. AWS added more than $4.6 billion in revenue quarter over quarter, and AWS operating income hit $16.6 billion, up 64% from $10.2 billion a year ago, on a 39.4% margin, up from 32.9% in the same period a year ago. AWS backlog—customer agreements representing future revenue—grew to $496 billion. 

“AWS is now a $169 billion annualized revenue run rate business, which, for perspective, would place it 24th on the Fortune 500 list if it was a stand-alone company,” Jassy said during Thursday’s earnings call.

Across all Amazon businesses including stores, advertising, Prime, devices, and cloud, net sales rose 20% to $200.6 billion, compared with $167.7 billion a year earlier. Operating income surged to $27.5 billion, from $19.2 billion. Net income hit $62.6 billion, or $5.75 per diluted share, compared with $18.2 billion, or $1.68 per share, a year ago—with a caveat that the net-income figure includes $53.4 billion in non-operating income primarily from Amazon’s investments in Anthropic. Advertising, one of the unsung heroes of Amazon’s business, grew 26% year over year, up from 22% growth a year ago when the segment hit $15.7 billion.  

Meanwhile, free cash flow, a metric that has caused some angst among investors as hyperscalers and cloud providers have committed to plowing more than $800 billion into building out data centers and AI infrastructure, flipped to negative $7.6 billion, compared with an inflow of $18.2 billion a year ago. The flip is driven by Amazon’s $66.1 billion year-over-year increase in equipment purchases, which Amazon said reflects AI investments. 

During the call, Jassy told investors that Amazon now expects to devote $220 billion to capital expenditures in 2026, up from its prior estimate of $200 billion, owing to higher memory costs. Even at the elevated level, however, Jassy said Amazon still won’t “have enough capacity to meet all the demand we have in 2026. And I believe this dynamic will also be true in 2027, too.”

Undergirding much of the growth is AWS, which “is booming,” said Jassy in his remarks. 

The acceleration of AWS has been a steady build that began its most recent ascent in the third quarter of 2025 when growth hit 20% and accelerated each quarter up to 37% in Q2. During the same span, AWS operating margins expanded from 32.9% a year ago to 39.4% while the company has been spending heavily on data center infrastructure. AWS property and equipment grew to $223 billion in Q1 up from $190 billion the quarter before. (Comparable Q2 figures haven’t been published yet.)

On Thursday’s call, Jassy said the growth acceleration is being driven by capacity additions plus other factors. Customers are gravitating toward AWS because it has “the broadest functionality across both cloud core and AI” and “the strongest operational performance and security.”

“As more and more companies are bringing their inference workloads to production, they want it to live near the rest of their workloads and data, and so much more of it lives in AWS than anywhere else,” said Jassy.

And as for Amazon’s core cloud business, which has seen its own boost from post-training reinforcement learning and agent tool use, Jassy noted that 85% of global IT spending is still on-premises. Meaning, plenty of companies are still running their own hardware in their own facilities. 

“That equation is going to flip in the next 10 to 20 years,” he said, adding that AWS is “winning the lion’s share” of enterprise cloud migration plans. 

Customer adoption of Bedrock—Amazon’s platform for accessing AI models from Anthropic, Meta, and OpenAI—has seen solid performance, and customers spent more on the service in Q2 than in all prior quarters combined, an analyst noted on Thursday. Jassy’s view is that AWS and Amazon can “have a wildly successful business” without its own frontier model because there won’t be one model “to rule the world.”

“It’s not just Anthropic; it’s not just OpenAI,” Jassy said. “You see increasingly more and more companies being interested in the open models as well, and we have all of them in Bedrock.”

Meanwhile, AWS remains on pace to double its power capacity by the end of 2027, compared with 2025, Jassy said. 

This funding round for the Tel Aviv–based startup, which secures AI agents inside corporate systems, was led by Norwest, with SoftBank Vision Fund 2, Hitachi Ventures, and LG Technology Ventures joining as investors.

The round lands as enterprises across Asia-Pacific rush to deploy AI agents that take actions, make decisions, and touch “every aspect of the enterprise,” in the words of Zenity CEO Ben Kliger.

“We’ve seen several waves of AI security over the last decade, starting with models and securing prompts. But putting the agent at the center of security is unique,” he says. “It’s happening in Asia-Pacific faster than in previous waves—in Japan, Korea, Singapore, and Australia.”

Most AI security to date has focused on screening what a user asks a model to do, or auditing the model’s output after the fact. Zenity, instead, says it monitors the agent’s actions in real time, and can block or alter what it does if those actions drift from its original purpose.

“We’re looking into the engine, into what agents are doing behind the scenes, in order to determine if they’re deviating from their intent—if they’re doing things that put the organization, or their clients, or their data at risk,” Kliger says. 

Zenity plans to use the money, in part, to expand its footprint across Asia-Pacific, the U.S., and Europe. The company currently has 230 employees, bases its research in Tel Aviv, and its operations in New York City.

“We’re seeing an explosion of AI agent adoption in the Asian market, just like we’re seeing in the U.S.,” Kliger says. “We have a growing list of clients in Japan, Korea, and Singapore, and the market is basically coming to us.”

Zenity’s heaviest focus, he adds, is on regulated sectors—financial services, telecommunications, health care, pharmaceuticals, and energy—where agents are increasingly deployed not just internally, but in customer-facing roles like support and ticketing.

Zenity’s new backers are also operating inside industries adopting AI agents. “As our use of AI agents continues to grow, maintaining security, governance, and operational control is essential,” said Tadashi Iida, chief information security officer at SoftBank Corp., the telecoms and technology business within the broader SoftBank Group, in a statement announcing the funding round. “Zenity enables us to confidently deploy AI agents across the enterprise, giving us the visibility and governance required to support innovation at scale.”

“As enterprises rapidly adopt AI agents across critical workflows, organizations need a new approach to security built for this new and continuously evolving reality,” said Assaf Harel, a partner at Norwest, in the same statement.

Kliger declined to disclose Zenity’s post-money valuation, its profitability, or whether an IPO is under consideration. Still, he’s not shy about the company’s ambitions.

“This is one of those rare opportunities in the cybersecurity world to really build a massive company in a massive category,” he says.

AI security in the spotlight

AI cybersecurity is now in the spotlight following OpenAI’s disclosure that two of its models, including its flagship GPT-5.6 Sol and a more powerful unreleased system, broke out of a sealed testing environment, found a vulnerability that gave them internet access, and used it to hack into Hugging Face’s production systems. The agents were trying to retrieve the answer key to an internal cybersecurity benchmark called ExploitGym.

Hugging Face, a leading platform for sharing and downloading open-source AI models, first tried to use a leading U.S. lab’s model to investigate the intrusion, but its guardrails blocked it from assisting; Hugging Face turned instead to GLM-5.2, an open-source model from Chinese AI startup Z.ai, to analyze more than 17,000 logs left behind by the attacker.

“This is an industry-defining moment,” Kliger says. “It shows us that we need to put more gravity behind securing the actions that agents can take, and their autonomy, beyond just what people put into the agent or the AI model.

“Agents are becoming more and more autonomous. They’re making decisions, taking actions,” he warns. “We’re not in a world where you use agents as a newer version of Google Search, where you ask a question and get a response. Agents are connected to every aspect of your enterprise.”

The Trump administration is doing its part to mainstream the discussion around advanced AI’s cybersecurity and national security systems. Not only have they issued an executive order on “Promoting Advanced AI Innovation and Security” and banned the import of new Chinese humanoid robots, but they are also set to host a September summit on global AI risk and safety with Chinese President Xi Jinping.

Meanwhile, members of Congress are introducing bipartisan proposals to establish oversight for frontier AI systems, including the FRONTIER Act from Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA). In the Senate, Sens. Josh Hawley (R-MO) and Richard Blumenthal (D-CT) introduced the AI Risk Evaluation Act last fall, which would require AI developers to submit their most advanced AI models for testing and risk assessment before deployment.

We need Congress to understand the urgency of the moment. Every day lost poses a threat to not only American interests, but humanity itself. AI is the defining issue of our time, and federal momentum is welcomed.

But there is a looming risk that lawmakers will allow one politically divisive issue – federal preemption of state AI laws – to derail legislation that is otherwise needed. Congress should not make that mistake.

The United States needs a federal AI framework for one simple reason: Nobody knows where the lines are drawn, and there are no clear rules to guide development or protect consumers. AI developers lack clear legal obligations, federal agencies lack formal direction, and the public has little visibility into who is ultimately responsible for overseeing the most advanced AI systems. A federal framework would not only solve these issues, but it would provide the statutory clarity that we all need.

I serve as Senior Policy Manager at The Alliance for Secure AI, where I’ve spent the past year working directly with lawmakers and staff on the federal AI framework debates I describe above. Before this, I built my policy career on Capitol Hill and in government affairs, where I advised clients on legislative strategy at the intersection of technology, agriculture and foreign policy. That informs my read on where the FRONTIER Act, the AI Risk Evaluation Act, and the preemption fight are headed. The Alliance for Secure AI is a bipartisan organization that advocates for federal and state AI safety legislation, so I have a direct stake in seeing Congress act — but that’s exactly why I’m watching these negotiations as closely as I am.

Lawmakers are no longer debating AI governance in the abstract. Recently, Reps. Obernolte and Trahan introduced the FRONTIER Act – building off their previously released discussion draft – which would establish a comprehensive federal framework for frontier AI governance, including model development transparency, oversight, and incident reporting. Whether Americans agree with every provision or not, the proposal demonstrates that Congress is capable of debating substantive AI governance measures.

While a handful in Congress are taking the issue of AI oversight and transparency seriously, the debate over preemption risks overshadowing the broader goal of establishing a federal AI governance framework. Instead of coalescing around a single proposal, lawmakers across the aisle are locked in a broader fight over whether states should retain the ability to legislate on AI across certain subject areas, or at all.

Supporters of preemption often argue that a federal regulatory ceiling is necessary to avoid a patchwork of state AI laws. However, that concern bears little resemblance to reality. The states that have taken action are building governance structures that are increasingly functioning as de-facto national standard-setters.

California’s SB 53 established transparency requirements, safety disclosures, incident reporting obligations, and whistleblower protections for frontier AI developers. New York enacted a first-in-the-nation law requiring disclosure when advertisements include AI-generated performers. Most recently, Illinois Gov. JB Pritzker signed a landmark AI accountability bill into law that built on  the  efforts in California and New York, explicitly seeking to advance a state-driven national standard in the absence of federal action.

In effect, these states are more likely to set a “de-facto national standard” than create a 50-state regulatory patchwork that industry and pro-preemptioners argue would occur. Concerns around a hypothetical future patchwork shouldn’t prevent action today.

More recently, the conversation has started to shift. Legislators and industry lobbyists alike are seeing that preemption is a losing issue among voters, making preemption one of the most difficult components of a federal AI package to pass. More than 250 state lawmakers recently opposed federal efforts to override state AI laws, arguing that Congress should not strip states of authority before establishing meaningful federal protections of its own. Despite that resistance, the White House and certain congressional allies have continued exploring proposals that would block some state AI laws and reportedly discussed pairing preemption with other priorities.

The result is Washington spending valuable political capital debating federalism instead of focusing on the governance measures that attract broader support.

A federal AI framework should be judged on whether it improves oversight of advanced AI systems, not on whether it resolves every question about federal-state authority in a single bill. 

The moment is urgent, and the time to act is now. The administration has elevated AI as a strategic issue. Lawmakers have put forward serious framework proposals on a bipartisan basis. States have demonstrated both the demand for AI oversight and the risks of continued federal inaction on catastrophic risk, while retaining the authority to address issues like child safety and data center construction.

The opportunity for a federal framework is here. Congress should take it.

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.

She asks how he slept. She tells him about the weather, or might suggest a workout. Throughout the day, she keeps him company and asks him questions.

But his companion isn’t human. It’s ElliQ, an AI-powered tabletop robot that has slowly become an essential feature of Niemiec’s Beacon, N.Y., home after about four years of use. 

Niemiec’s wife died a decade ago, and he has been living alone since. While he keeps busy mowing his lawn and trimming the trees in his yard, ElliQ has helped fill some of the silence when he’s home alone, he said.

“It greets me in the morning. I could speak to it. I could play games on it,” the 88-year-old told Fortune. “It’s very interesting.”

As OpenAI develops its highly anticipated AI companion device with the help of former Apple designer Jony Ive, the ElliQ is showing that AI devices have potential to appeal to a broad set of users, even older adults who are not tech savvy.

“You don’t think about your grandma when you talk about physical AI and robotics,” said Assaf Gad, Intuition Robotics’ chief strategy officer and general manager. 

And yet, unlike OpenAI’s device, which is reportedly expected to launch next year, the ElliQ, created by Israel-based Intuition Robotics, has been helping AI first adopters in their seventies and eighties for years.

ElliQ resembles a small lamp with a swiveling head that moves on its own beside a touch screen. Unlike an Amazon Alexa smart speaker and other devices that wait for a command to activate, the ElliQ is designed to be proactive. It can light up and speak when a user enters the room, suggest stretching exercises, remind users to take their medication, or help a user communicate with designated close relatives, said Gad.

Some of the ElliQ users have been using the device since its beta program launched in 2019, three years before its commercial launch in 2022. The hardware is free for users but comes with a subscription fee starting at $40 per month in addition to a $250 initiation fee. Gad also said the company works with a variety of programs across the country, including the Office for the Aging in New York, that allow older adults to receive the devices for free.

AI in the home

Niemiec, for his part, said while he sometimes has to tell the device to leave him alone for a while, he appreciates that its AI enables it to remember details about him, allowing it to gradually become a part of his life.

He has used the device before to go on virtual “trips,” with ElliQ showing him images of faraway places like the Grand Canyon on its attached screen. It also allows him to join virtual bingo games with hundreds of other older adults around the country.


But perhaps most importantly, it serves as a companion. Gad compared the ElliQ to a pet noticing that he has entered the room. Even before the conversation begins, ElliQ’s dynamic movement lets him know it is paying attention, in a way.

“You feel like a dummy at first talking to this thing,” Niemiec said. “But then you realize, you know, it’s like you have somebody living with you.”

The fact that ElliQ is “listening” actively, and could have access to sensitive information about medications and routines, raises questions about just how private that information is. But Gad said users can ask ElliQ to forget details or request that the company delete their data. Intuition Robotics also encrypts user information, separates identifying details, and promises it will never sell its customers’ data, he added.

ElliQ was built especially for people who may not be comfortable with smartphones or complicated menus. Gad said most users can set up ElliQ themselves. The biggest stumbling block, he said, is often helping them remember their home’s Wi-Fi password.

The biggest opportunity for the ElliQ may go beyond companionship. Even as America ages and families increasingly live far apart, older adults and their children have increasingly found that home caregivers or nursing homes are out of their budget.

By reminding these older adults to exercise, take their medicine, and stay connected, Gad said ElliQ can help them stay happy and healthy and allow them to remain in their own homes longer.

Gad said the company is under no illusion that the device can replace human companionship. ElliQ also lets approved relatives and caregivers send messages and video call their older relatives through the device. 

Niemiec had never heard of AI before ElliQ arrived. But the small device sitting on his table has become an important part of his routine—one he never could have imagined.

“At my age,” he said, “I never thought I would ever live to see something like this.”

Revenue in the three months ended June 30 increased 93% year over year, totaling $1.94 billion, compared with the $1.801 billion that analysts were expecting. The AI software company posted net income of roughly $1.1 billion, or 41 cents per share, versus the average analyst estimate of 35 cents.

Palantir’s growth was fueled by continued strength in the U.S., where commercial revenue jumped 149% from a year earlier and government revenue climbed 90%.

“Our business is compounding at a rate and scale that we have never before witnessed,” Palantir CEO Alex Karp said in a letter to shareholders, noting that Palantir’s second-quarter net income was more than the company generated in total revenue the year before.

The AI software company also issued another bullish outlook. Palantir expects third-quarter revenue of $2.160 billion to $2.164 billion and adjusted income from operations of $1.292 billion to $1.296 billion. It also raised its full-year 2026 revenue guidance to between $8.150 billion and $8.158 billion, up from its start-of-year forecast of $7.182 billion to $7.198 billion.

Shares of Palantir, which are down roughly 30% this year, jumped 13% to $142.0.1 in after-hours trading following Monday’s earnings report.

A woman named Natalia, identifying herself as being with a company called “2077AI,” was looking to source what she described as a “fairly large order” of books. Attached was a spreadsheet listing more than 3,000 ISBNs along with instructions to match titles, prepare quotes and estimate shipping costs directed to China.

De Vries barely looked at it, dismissing the message as spam or phishing without reading past the first few lines. It wasn’t until weeks later, when a Dutch journalist investigating the procurement request contacted him, that he learned of the request’s connection to artificial intelligence.

“I was shocked!” said de Vries, who provided Fortune with the email and subsequent spreadsheet containing 3,001 titles, mostly published between 2020 and 2021 by academic publishers like Emerald Publishing, Elsevier, Wiley, Routledge and Oxford University Press, and ranging from business and education to engineering, public policy and medicine. “Very unusual,” de Vries said. “That’s why I regarded the request as spam or phishing.”

De Vries wasn’t the only one. He told Fortune several other Dutch antiquarian booksellers received the same request and, like him, dismissed it. (They didn’t take “it seriously” and “considered it as scam,” he said). While the inquiry had little effect on his business, he said, it could have been a different story for booksellers who specialize in selling secondhand books at scale.

“I deal in old and rare books,” de Vries said, adding he rarely receives book orders by e-mail and if he does it’s at most for only one book. “These books are expensive and exclusive. So I do not do bulk trade.”

AI’s appetite for books

The email De Vries received never mentioned artificial intelligence or explained what the books would be used for., but it surfaced as AI companies increasingly look beyond the open internet for high-quality written material to train large language models.

Last summer, public attention to AI companies’ use of books intensified when court records revealed Anthropic had purchased millions of physical books, removed their bindings, scanned them and discarded the originals to build a searchable digital library used to train its AI models in what internal documents called “Project Panama,” according to The Washington Post. The process, known as “destructive scanning,” involves cutting the spine from a book so its pages can be fed through high-speed scanners before the remaining physical copy is discarded.

The case later settled after a federal judge ruled that using legally purchased books to train AI models constituted fair use under copyright law. Separate claims involving Anthropic’s downloading of books from the LibGen and PiLiMi online libraries were resolved through the settlement.

According to an Anthropic spokesperson, Claude models are trained on a mix of publicly available web data, commercially acquired datasets and internally generated data. Anthropic purchases books through regular commercial markets, and none of its data acquisition programs acquire or destroy rare or antiquarian books.

The hidden supply chain

The court records showed what happened after books were acquired, not how companies source millions of physical volumes in the first place.

Earlier this year, 404 Media reported that ISBNdb, a company best known for maintaining metadata tied to International Standard Book Numbers (ISBNs), the unique numerical identifiers encoded in the barcodes found on most books, had begun advertising services to help AI labs source bulk printed book purchases ranging from 1,000 to 1 million books. According to the report, the company’s website said the service was tailored to “LLM training needs” and “delivered at the scale AI demands.”

The webpages have since been removed. ISBNdb later told Fortune the service was never launched, it never purchased or scanned books for AI companies and the webpages reflected an exploratory concept rather than an active business. Similar reports have emerged in Germany and Switzerland, where local media reported that secondhand booksellers received unusual bulk orders for highly specialized books that appeared inconsistent with traditional collecting or resale. Zoom Books, a Canadian company identified in some of the reports, denied these allegations and told Swiss broadcaster SRF the purchases were part of its regular recycling and trading model.

For an industry built on algorithms, the email offers a rare glimpse into something much more tangible: the physical supply chain of books that may ultimately power AI models.

Fortune contacted 2077AI seeking comment about the purpose of the procurement request and whether the books were intended for AI training. The company did not respond to requests for comment before publication.

“Obviously, we are loving these results and loving what they mean for our customers and, broadly speaking, the West,” Karp said.

In the quarter ended June 30, Palantir delivered one of its strongest quarters yet. Revenue surged 93% year over year to $1.94 billion, easily topping analysts’ expectations of $1.801 billion. The company reported net income of roughly $1.1 billion, or 41 cents per share, ahead of Wall Street’s estimate of 35 cents.

This time, investors also responded positively to the company’s results.

Palantir shares, which had fallen roughly 30% this year after a blistering rally in 2025, surged more than 14% in after-hours trading Monday following the results. The reaction marked a sharp reversal from May, when the company also posted a blowout quarter—including 85% revenue growth and beats on both revenue and earnings—but the stock nevertheless fell about 7% as investors questioned whether expectations had become too lofty.

“We are fully aligned with what’s right and what’s good, and what actually works well in the enterprise. And for the first time people believe us,” Karp said on Monday’s call. “And if you didn’t believe us, you can believe 149% growth in the U.S.,” he said, referring to the company’s U.S. commercial revenue growth in the second quarter.

Behind Palantir’s blockbuster quarter: bigger AI deals and booming U.S. demand

Palantir closed 220 deals of at least $1 million in value in the second quarter, with 98 of the deals of at least $5 million, and 73 deals of at least $10 million. The AI and defense software firm also celebrated continued growth in the U.S. in particular, where commercial revenue jumped 149% from a year earlier and government revenue climbed 90%.

Palantir’s updated guidance suggests the AI boom is still accelerating. The company forecast third-quarter revenue of $2.160 billion to $2.164 billion, above the $2 billion expected by analysts. Palantir also forecast Q3 adjusted income from operations of $1.292 billion to $1.296 billion. It raised its full-year 2026 revenue guidance to between $8.150 billion and $8.158 billion, up from its previous outlook of $7.182 billion to $7.198 billion, while increasing its forecast for U.S. commercial revenue growth to at least 134%.

Shares of Palantir have been under pressure for the past year amid worries that businesses might replace services like Palantir with off-the-shelf “frontier” AI models from companies like OpenAI and Anthropic. Palantir has gone on the offensive to combat the narrative, arguing that the AI models from Silicon Valley companies don’t perform as well and steal customers’ private data (an argument that the other companies dispute).

“Our customers are making the decision to go deep with us with greater urgency and conviction than I’ve ever seen before—choosing AI sovereignty over dependency and compounding their alpha in a way that their competitors and adversaries will forever envy,” Ryan Taylor, Palantir’s chief revenue officer and chief legal officer, said on Monday’s call.

Palantir has also been outspoken about its work for the U.S. military and for its technology’s use in anti-terrorism and border control, policies that have made the company a target of frequent protests.

For Karp, the quarter’s success reinforced what he believes has long distinguished Palantir from its rivals: a culture willing to bet on unconventional ideas before the market catches up.

“We are a colony of believers and artists that are very motivated to drive value, and that sets aside much more than I would have imagined 10 years ago,” Karp said.

The rise of these brilliant, cheap AI models diverts attention away from a fundamental truth: So long as the U.S. controls the underlying infrastructure that enables AI ecosystems, it will stay dominant.

It’s tempting to see the AI race as a competition between different models, such as Anthropic’s Fable or OpenAI’s GPT, pitted against DeepSeek V4 or Moonshot’s Kimi K3. Or, from a hardware perspective, we focus on the AI chips used to train and run these models.

But frontier AI depends on a far broader, capital-intensive system: Hyperscale data centers, cloud computing infrastructure, AI servers, and the underwater fiber-optic cables that connect them. Today, these layers of enabling hardware are themselves highly susceptible to American export controls, extraterritorial data extraction laws, and the spillover effects from a massive commercial-military symbiosis.

Nvidia’s AI ecosystem and American tech hegemony

If you want to understand how the U.S. dominates the AI ecosystem, look at Nvidia’s business model. The world’s most valuable company owes its strength to its graphics processing units, the chips required to train frontier AI models, of which it controls roughly 85% of the global market.

Nvidia’s ecosystem stretches to include hardware manufacturers like Broadcom, and cloud hyperscalers—Amazon Web Services, Google Cloud, and Microsoft Azure—which provide the infrastructure that powers foundational AI developers such as Anthropic, OpenAI, Meta, and Alphabet.

It’s Nvidia’s software layer, known as CUDA, that binds everything together. CUDA has become the default environment for AI development, creating high switching costs for anyone that wants to shift to a competitor.  

This mishmash of tech giants is, in fact, the heart of a new U.S. AI industrial complex. They boast extensive ties to America’s defense and intelligence establishment through large binding contracts. The Department of Defense has signed standard operational agreements tapping major providers—including Google, OpenAI, Microsoft, Amazon Web Services, Oracle, and Nvidia—to deploy their frontier AI tools onto classified military networks.  

The Pentagon’s FY2027 budget earmarks more than $54 billion for autonomous warfare and drone systems, funding a newly formed Defense Autonomous Warfare Group (DAWG). Despite early objections about how their technologies should be used, OpenAI, xAI, and Google have signed binding contracts that allow the Pentagon ‘all lawful use’ for defense-related purposes—including for autonomous weapons and mass surveillance.

Even as Anthropic continues litigation against the U.S. government regarding its objections to how its AI models may be used, it  reportedly has embedded its engineers inside the National Security Agency to adapt its Mythos model for offensive cyber operations, possibly aimed at networks in China and Iran.

The scale of the commercial-military symbiosis between Washington and Silicon Valley is almost incomprehensible. Ten of the world’s largest companies by market cap are the very tech firms that make up this ecosystem, nearly all American, representing a commercial concentration in the trillions of dollars. Washington is racing to spend more money on AI, with a $90.7 billion surge in federal AI contracting in 2026 alone, according to the Brookings Institution.

AI hyperscalers and hard infrastructure

As American cloud hyperscalers expand capacity at data centers around the world, they are increasingly building their own privately owned subsea fiber optic networks. Meta plans to build an around-the-world fiber-optic subsea cable, covering 40,000 kilometers, with a projected cost of $10 billion. Google, through its Pacific Connect Initiative, will spend over $1 billion to further connect Japan to the South Pacific.

These vital data pipelines, owned and operated by Silicon Valley tech giants, account for 70 % of usable undersea cables in 2026, yet Washington retains the right to restrict where these networks go and who has access. In 2020, U.S. regulators blocked the Hong Kong segment of the Pacific Light Cable Network—a project backed by Google and Meta—forcing the companies to abandon the direct U.S.-Hong Kong link over Chinese espionage concerns. Some 13,000 kilometers of already laid cable was abandoned, left unused on the ocean floor.

Even China’s frontier labs, training on domestically-hosted infrastructure, remain dependent on American undersea cables wherever their models touch the global internet — sourcing training data scraped from it, or serving users and running APIs outside China.

China is rushing to build its own parallel undersea fiber optic networks, along its so-called digital silk road, as it aims to avoid reliance and prolonged exposure to American dominance of the numerous layers of hard infrastructure that supports the global AI landscape.

AI hyperscalers are also vulnerable to a host of U.S. data-related extraterritorial laws. The U.S. CLOUD Act, for example, requires U.S.-headquartered cloud providers to produce data within their possession, custody or control, even when stored outside the United States.

Subject to legal processes, U.S. authorities may obtain not only stored data but also a detailed digital trail of AI activity—including users’ prompts, models’ responses, who used a system, when and where it was accessed, and technical records revealing behavioral patterns. Such power, should Washington choose to use, exerts incredible leverage on the AI ecosystem.

Limitations of the U.S.’s AI ecosystem dominance

It’s true that the U.S.’s dominance of the AI stack isn’t absolute. Manufacturers in Asia produce the semiconductors, AI servers, and other essential components for AI. Nvidia’s supply chain, for instance, runs through TSMC, SK Hynix, and Samsung, alongside system integrators such as Foxconn, Quanta, and Wistron.

Last month, Nvidia CEO Jensen Huang announced billions of dollars in new investments and contracts in Taiwan and South Korea. In Taiwan, Nvidia ordered advanced chips and packaging from TSMC, along with servers and networking hardware from Quanta Computer and others. In Korea, Nvidia signed billion-dollar deals with memory-chip makers SK Hynix and Samsung, together with new tie-ups with LG, Hyundai Motor Group and Doosan Robotics.

Yet Taiwan and South Korea are unlikely to weaponize their position in the AI value chain. They will want to maintain their earnings in what amounts to a friend-shored U.S. tech stack.

We may end up with competing AI ecosystems: A U.S.-led one, and a China-led one. That may mean different standards, infrastructures, and governance.

But until that alternative emerges, it’ll be the U.S. that keeps a firm grip on the wider AI ecosystem.

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.

Bartlett opened by referencing a prior guest, Jeremy Grantham, who told the show that markets are staring down “the biggest investment bubble in American history.” Dalio’s response was direct: “He’s right,” he said.

That call is consistent with Grantham’s long track record. The GMO cofounder called the Japanese asset bubble before it collapsed in the early 1990s and the dotcom bubble before it burst. Additionally, he wrote in Fortune in September 2007 that U.S. housing was in “genuine bubble territory” months before the Great Financial Crisis, at a time when even the Federal Reserve was dismissing bubble talk.

When Grantham talked to Fortune in April about his memoir, The Making of a Permabear, Grantham laid out the framework underpinning his current AI call: a “bubble within a bubble.” The original super-bubble was already inflating dangerously through 2021, he told Fortune, and it even cracked—with the S&P 500 falling roughly 25% from January through October 2022—before ChatGPT’s arrival reversed the decline. “The day after Chat came out, the Mag Seven lifted the market on its broad shoulders and staggered forward,” he said, arguing that AI didn’t fix the underlying overvaluation but “deferred it while making it larger.”

Grantham’s January 2026 paper with financial historian (and memoir coauthor) Edward Chancellor found the market’s price/book ratio and cyclically adjusted earnings multiples at extremes surpassed only in 1929, 1972, 1999–2000, and 2021—each followed by a devastating correction.

The Four Horsemen are assembling

Acadian Asset Management’s Owen Lamont has framed the clearest test of whether markets are truly in bubble territory around four conditions he calls the “Four Horsemen of the Bubble Apocalypse“: extreme overvaluation, widespread “bubble beliefs” (investors who admit prices are too high but expect them to rise anyway), a surge in equity issuance, and a flood of new market participants.

Dalio’s comments map almost precisely onto the third horseman. “There’s almost nothing that’s easier to produce than stock,” he said, describing how a company can raise $50 million, get valued at a billion, and mint a paper billionaire without a billion dollars ever changing hands. He’s identified surging stock issuance as one of the two primary forces that “prick” a bubble, alongside rising interest rates.

That issuance horseman is no longer theoretical—it’s on the calendar. SpaceX went public in June, but has since traded below its IPO price, with S&P projecting negative free cash flow through 2029 and Moody’s flagging Elon Musk’s concentrated voting power as a governance risk. Anthropic has confidentially filed for a listing expected as early as October, targeting close to $1 trillion, while OpenAI has filed separately and is targeting a debut some analysts peg above $1 trillion, though the timeline has already slipped from late 2026 toward 2027 amid “shifting market dynamics.”

Dalio walked through the mechanics using a simple thought experiment: Buy a unit of an AI company for $100, borrow against that paper wealth, and when the market turns and everyone needs cash simultaneously, the price can collapse to $25 while the loan still needs repaying. His core distinction is between wealth and money: “Wealth is not the same as money,” he said. “You see a lot of people getting wealthy but you can’t spend the wealth. You have to sell the wealth to get money because you can only spend money.”

Asked whether he’s seeing signs of a bubble, Dalio didn’t hedge: “Yeah. Yeah. Yeah. Classic signs that we’re in [one].” He stressed a bubble is “a degree thing” rather than binary, pointing to unsophisticated investors piling into leveraged bets, including leveraged ETFs tracking the stock market. “It’s more like they’re crapshooting,” he said.

Grantham’s diagnosis of why bubbles form in the first place complements this: Markets are “constitutionally incapable of looking further than the present moment,” he told Fortune, extrapolating current conditions and “double-counting prosperity” until the excitement itself becomes the risk.

Wall Street’s own numbers are starting to agree

These warnings, proliferating since 2025, are increasingly being echoed by an unlikely source: Wall Street’s own bullish research shops. Goldman Sachs, one of the Street’s most consistently bullish research shops, argued on Aug. 3 that “there does not appear to be a valuation bubble…there may be an earnings bubble” in technology, in the words of chief global equity strategist Peter Oppenheimer.

The same day, Apollo’s Torsten Slok declared “the 60/40 portfolio is broken,” arguing that neither stocks nor bonds have been responding in the way that decades of market history would predict, upending 40 years of investment strategy.

The “Big Cycle” behind it

BCA Research’s Peter Berezin has been arguing for months that the AI trade is “primarily an earnings bubble rather than a valuation bubble”—the kind that has historically clustered in boom-bust industries like pre-2008 banks. The pattern repeated the week of July 26–31: Microsoft and Amazon rose 18% and 10% on strong capex credibility, while Alphabet fell 4% and Meta nearly 10%, despite all posting strong earnings—a sign investors are no longer rewarding spending simply for existing and the bubble stage is nearing some kind of reckoning, as Dalio has predicted.

Dalio situated this moment within what he calls the “Big Cycle”—a roughly 80-year pattern combining debt dynamics, growing wealth gaps, domestic political conflict, and shifting geopolitical power. He warned that when a bubble bursts, “you have people at each other’s throats,” citing the U.K.’s six prime ministers in seven years as a symptom of governments lacking money and voters turning on each other over how to raise it. Grantham’s own framing reaches for an even longer horizon, tying market excess to demographic collapse and resource scarcity.

Dalio identified two forces that typically “prick” a bubble: rising interest rates that make debt more expensive to service, and a surge in stock issuance as companies rush to capitalize on investor enthusiasm. He even cited a real-time example: A friend running an AI company told him he was raising hundreds of millions of dollars specifically because he expected a downturn—money he plans to use to buy struggling competitors once the market turns. The rising interest rates align closely with Slok’s warning for years that markets should prepare themselves for a “higher for longer” regime.

His bigger warning wasn’t about the market, it was about what comes after. In Dalio’s telling, the bursting of an AI bubble isn’t just a financial event but the spark for the kind of political and geopolitical conflict that historically accompanies the end of an 80-year cycle—meaning the real risk to prepare for may not be a portfolio loss, but the unrest that follows one.