Aug 5, 2025

Welcome back to The Hidden Layer. I’m Ian Krietzberg, waiting to see if the rumors of GPT-5’s unveiling will actually come true this week.

In today’s issue, I’m examining the steadily soaring capex commitments of the biggest Big Tech players. Their investments, along with the frenzy of the private markets, are driving the A.I. revolution—even as these same companies are also forcing user adoption by integrating their tech into their existing services, like Google Search and Microsoft 365.

Discussed in this issue: Meta, OpenAI, Petar Tsankov, Tesla, Joseph Weizenbaum, TikTok, Neil Dutta, Paul Kedrosky, and many more…

Let’s get into it…

Three Things You Should Know

A spoonful of regulation: Last weekend, after seemingly endless criticism from tech companies on both sides of the Atlantic, the European Union finalized its General Purpose A.I. Code of Practice. Tech companies can voluntarily sign the document, which is intended to streamline compliance with E.U.’s general purpose A.I. obligations now that the law is enforceable. The likes of OpenAI, Microsoft, Google, IBM, Anthropic, and Amazon have already signed on. Developers that have refused, like Meta, will have to create their own framework for compliance.

The code has three main chapters: Transparency, copyright, and safety and security. The first two are relatively brief and straightforward. They require transparency around model architecture and intended use cases, summaries of training data, and affirmations that companies will abide by European copyright law when it comes to data scraping. For developers whose models exceed 10²⁵ FLOP of compute, there are additional safety and security standards, including the creation of an internal risk framework, risk-mitigation strategies, and transparent incident reporting.

Dr. Petar Tsankov, the co-founder and C.E.O. of LatticeFlow, a company working to enable trustworthy A.I. systems, noted that the burden of compliance is relatively minimal: It only applies to big vendors, and can be achieved by a single dedicated employee. “It’s very non-bureaucratic,” he told me “And I think everybody who was pushing heavily against [this] on the industry side got quiet after this.” Developers whose models entered the market before the August 2 deadline will have two years to align with the new laws of the land.
Tesla’s liability: Last week, a Florida jury found Tesla partially liable for a fatal Autopilot crash that occurred in 2019, ordering the company to pay $329 million in damages. Tesla plans to appeal the decision. This is the first of about a dozen similar lawsuits to actually reach a verdict. In this particular case, the driver, who had just dropped his phone, was under the impression that Autopilot would stop the car if it detected an obstacle. As he moved to pick it up, he whipped through an intersection at 60 miles per hour and hit a parked car, killing a 22-year-old and severely injuring her boyfriend.

Tesla said that the verdict “is wrong and only works to set back automotive safety and jeopardize Tesla’s and the entire industry’s efforts to develop and implement life-saving technology.” Meanwhile, safety engineer and self-driving expert Dr. Missy Cummings, who was the A.I. expert on this case, called the verdict “not only a win for the victims, but all of automotive safety.” TeslaDeaths.com, which has been tracking Autopilot-related deaths since the service became available, has catalogued at least 58 deaths attributed to Full Self-Driving or Autopilot mode.

At the heart of many of the F.S.D. cases is a claim of false advertising. In 2024, the National Highway Traffic Safety Administration—which is also investigating Tesla’s self-driving efforts—warned the company to be more careful in online messaging about its so-called “self-driving” products. Since then, Tesla has added the word “Supervised” to tweets about Full Self-Driving. The hype from Elon, of course, has hardly slowed down, and shareholders are getting fed up; on Monday, Tesla shareholder Denise Morand filed a proposed class-action lawsuit against Tesla and Musk, accusing them of securities fraud for heavily over-playing the whole ‘robotaxi’ thing.

A MESSAGE FROM BLOOM ENERGY

As AI adoption accelerates, power has become the defining constraint—and opportunity—for data center growth. Our latest 2025 Mid-Year Power Report reveals a dramatic shift in how industry leaders are planning for the future.

Read the report.

The ChatGPT effect: ChatGPT isn’t the first chatbot that users have co-opted as a sort of digital therapist. The first, named “Eliza,” was designed by M.I.T. computer scientist Joseph Weizenbaum more than 50 years ago with that explicit purpose. For Weizenbaum, Eliza only perpetuated an “illusion of understanding”—but, much to his surprise, people very quickly bought into the illusion and developed emotional bonds with the machine. This phenomenon was later termed “The Eliza Effect.” (Troubled by the outcome, Weizenbaum became one of the first A.I. skeptics and spent years writing the book Computer Power and Human Reason: From Judgment to Calculation.)

Since then, of course, the number of people self-reporting using ChatGPT for therapy or life coaching or whatever has been on the rise. Studies have warned that L.L.M.s are not suited for this role, mainly because they tend to “encourage delusional thinking.” Alas, even the threat of A.I. psychosis hasn’t stopped people from using Chat for off-label purposes. OpenAI acknowledged the problem on Monday, saying that it has been working with medical experts to create tools and rubrics that “better detect signs of mental or emotional distress so ChatGPT can respond appropriately.” It’s also taking a page out of Netflix’s book, adding “gentle reminders during long sessions to encourage breaks.”

A week earlier, OpenAI deactivated a feature that allowed users to share their (often deeply personal) chats, which had resulted in Google indexing tens of thousands of name-redacted conversations that were nonetheless replete with personal questions and information. Though OpenAI started removing these chats from the internet, they’ve already been archived by services like the Wayback Machine, where they’ll presumably live forever.

Hallucination of the Week: Mad as Rabbits

If you happened across a video of a bunch (a gaggle?) of rabbits jumping on a trampoline, you are not alone; the original post on TikTok has been viewed around 230 million times. Too bad nothing about the video is real.

The clip was generated by A.I., something that becomes pretty obvious if you watch closely as one of the bunnies magically (or impossibly) merges with another. Still, the style of video—C.C.T.V.-type footage, nighttime, crickets in the background—is really convincing. Hate to burst your bubble (break your trampoline?).

And now for the main event…

A.I. Bubble Bust Theories

As the major A.I. hyperscalers prepare to spend some $325 billion in capex this year—enough to noticeably raise U.S. G.D.P.—economists are beginning to worry we’re in a bubble… even as they acknowledge there may not be any real alternative.

Ian Krietzberg

The quest for scale remains a preoccupation at almost every level of the A.I. pipeline: There’s the mind-boggling amount of training data required; the alarming quantity of water and electricity needed to run data centers; and the millions of G.P.U. chips that provide the necessary computing power. Then there are the extraordinary public market valuations for many of the largest A.I. players, which have buoyed the stock market for the past two years.

Just as notable, of course, is the unprecedented capex flooding the industry. In 2023, the four major hyperscalers (Meta, Google, Amazon, and Microsoft) reported a modest $140 billion in total capital expenditures, largely for A.I. chips and data centers. In 2024, that number crossed $220 billion. At the beginning of 2025, that cadre of tech giants forecasted total expenditures of $325 billion for the year. After this most recent round of earnings reports, the total is likely to come in above $360 billion.

And no one really expects this to stop anytime soon. Morgan Stanley analysts Vishwanath Tirupattur and Vishwas Patkar forecast in a recent note that an additional $3 trillion could be invested in data centers globally through 2028. The analysts expect roughly half of that amount to be financed through debt, something that’s already starting to happen—Elon Musk’s xAI has raised more than $5 billion in debt, and is reportedly in talks to raise a further $12 billion; CoreWeave has billions of dollars in debt, something that its C.E.O. has called the “fuel for this company”; Meta is reportedly in talks to raise $26 billion in debt to fund data centers; and OpenAI last year secured a $4 billion line of credit. Historically, a concentration of bank debt tends to make the difference between a crash that everyone walks away from (like in ’02), and a crash that has broader economic impacts (like in ’08).

A MESSAGE FROM BLOOM ENERGY

AI growth is outpacing power infrastructure, making energy strategy a board-level concern. The 2025 Mid-Year Power Report highlights grid delays, rising onsite generation, and the strategic importance of power access.

Get the report.

The scale of investment is so significant, according to Neil Dutta at Renaissance Macro Research, that A.I. capex contributed more to G.D.P. growth than consumer spending did across the first half of 2025. Dr. Paul Kedrosky, a researcher and investor, estimated that A.I. spending alone accounted for nearly half of the 3 percent growth in the U.S.’s latest G.D.P. report. Tejas Dessai, the director of research at Global X, called A.I. the “economy’s new engine” in a recent note.

So far, so good, right? After all, these companies keep exceeding analysts’ expectations as they reinvest their profits into their future. But the role that A.I. plays in their future business units is less transparent than the overwhelming financial narrative. We’re dealing with massive corporations, each of which already has enormous established businesses, making it even more difficult to measure the impact of all their investments in A.I. Deployment of A.I. might lead to, e.g., more (or more valuable) ad sales, or cheaper revenue, or operational efficiencies. But we don’t have a whole lot of visibility into that, and more-direct revenue streams remain decidedly uncertain.

The Four

It’s worth putting the scale of these capital expenditures in context. Amazon, for instance, raked in $167.7 billion in revenue for the quarter. Meta handily beat expectations with $47.5 billion in revenue. Microsoft also beat expectations with $76 billion in revenue. Meanwhile, Google earned $96.4 billion in quarterly revenue. Despite their extraordinary investments in A.I., these earnings largely derived from core and well-established business units.

And yet the A.I. narrative, which has contributed some $5 trillion in market cap to the so-called Magnificent Seven since April, buttressed the value of their capital allocation as retail investors plowed hundreds of billions into the stock market, and specifically into tech and A.I. companies. J.P. Morgan expects retail investors to contribute $360 billion, of a total of $500 billion, in equity inflows for the second half of the year. This frenzy of investment just keeps driving valuations higher, pushing the tech weighting of the S&P 500 to around 34 percent, just about double its historical average. Kedrosky told me the valuations are “highly irrational.”

Indeed, corporate investment into A.I. data centers is becoming comparable, as a percentage of G.D.P., to the railroad-building fever of the late 1800s, according to Kedrosky, which culminated in the Panic of 1873. (A.I. capex comprises about 1.3 percent of G.D.P. today; railroads once contributed 6 percent.) The critical difference is that A.I. chips depreciate a lot faster than railroad infrastructure. “A.I. data centers are short-lived, asset-intensive facilities riding declining-cost technology curves, requiring frequent hardware replacement to preserve margins,” Kedrosky warned.

On some level, it seems obvious that we’ve entered an A.I. bubble. Yes, The Four are going to be fine and make more money than ever, but others won’t be as lucky in this Manichean, winners-take-most economy. At the same time, there’s no clear definition for what constitutes an asset bubble, especially when the potential returns are speculative (it’s much easier to call it a bubble after it pops). Winston Dou, a professor of finance at Wharton, articulated the two main schools of thought: One could argue these investments are entirely rational, and the valuations justifiable, if there is a sufficiently high degree of certainty that the bet will pay off. But if expectations fall short of reality, the entire market will have to grapple with valuations coming down across the board. The trouble right now is a total lack of clarity, a blank space that investors have filled with evidence to support their preferred narrative, all centered on upside uncertainty, which seems to have unlimited potential. This is a kind of uncertainty that Dou said “investors want to embrace,” largely because the alternative is gruesome.

As for Meta, Microsoft, Amazon, and Google… well, Dou doesn’t think they have much of a choice. A failure to effectively compete poses significant displacement risks, and when investors keep driving up valuations over excitement around A.I., now certainly seems like a good time to go all-in, even if we’re still waiting to see the flop. But even if it is rational, Dou said that this current environment is just “not sustainable.” At some point, all this uncertainty gets resolved into clear losers and winners, and we’ll see a rapid decline in the market valuation of the entire sector as the losers tank, something that’ll look an awful lot like a crash. This, he said, “is just the nature of the game.”

When the market does correct—or revert to the mean, if you prefer—Kedrosky expects states like California and Virginia, whose economies have benefitted from massive amounts of A.I. spending, to suffer revenue shortfalls, and for funds that are unexpectedly exposed to A.I. to underperform. Mainly, though, he thinks the “price of tokens will collapse, as data centers ‘dump’ supply onto the market, much like OPEC. This will reduce the cost of replacing humans in many occupations, causing accelerated job losses in areas like customer service.” The operative question now is how much patience investors can summon as all this uncertainty plays out.

That’s all for today. I’ll see you Thursday.

Ian

Join Emmy Award-winning journalist Peter Hamby, along with the team of expert journalists at Puck, as they let you in on the conversations insiders are having across the four corners of power in America: Wall Street, Washington, Silicon Valley, and Hollywood. Presented in partnership with Audacy, new episodes publish daily, Monday through Friday.

A professional-grade rundown on the business of sports from John Ourand, the industry’s preeminent journalist, covering the leagues, players, agencies, media deals, and the egos fueling it all.

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006