Sam Altman의 1조 달러 거품: 왜 OpenAI가 AI 업계의 '리먼 브라더스'가 될 수밖에 없는가?

Original article from Ed Zitron

Translation & Editing | Odaily Planet Daily Qin Xiaofeng (@QinXiaofeng 888 )

Editor's Note: OpenAI secretly filed for an IPO with the SEC in June 2026 and is gradually approaching an IPO, with the current timeline potentially pushed back to 2027; if successful, it would become one of the largest AI IPOs in history.

Recently, independent tech commentator Ed Zitron published a lengthy critique of OpenAI. He argues that the true AI bubble is essentially the "OpenAI bubble" and that if OpenAI fails, it will become the "Lehman Brothers" of the AI era, potentially triggering a repricing of data centers, AI infrastructure, and tech stocks. He emphasizes that OpenAI's business model is unsustainable, relying on subsidies and circular financing, and lacks genuine demand. (Odaily note: Ed Zitron is known for his frank, financially detailed style of dissecting tech bubbles. This long-form article is the latest and most aggressive in Zitron's series criticizing the AI bubble, following previous pieces like the exclusive report on OpenAI's losses and the Silicon Valley Bubble series.)

The original article is lengthy. Odaily Planet Daily has compiled and translated it as follows. Enjoy～

——————————————

Today's article is one of the biggest free newsletters I have ever written, compiling the work of the past six months.

And it all starts with one question: How much do you trust Sam Altman? The stock market, and to some extent the global economy, depends on your answer.

OpenAI has become one of the biggest liabilities in modern economic history. You can argue that OpenAI is no longer the focus of the AI bubble—you can talk about open-source models or Anthropic or any other element—but without OpenAI, the AI industry would not exist, and the justification for trillions of dollars in capital expenditure would disappear.

The AI bubble does not stem from any actual return on investment—whether purely monetary, like revenue or profitability, productivity gains, or anything tangible or measurable. Instead, it's a cult-like psychotic episode infecting the minds of some of the most powerful and wealthy individuals and institutions, where the powerful mythology of one company inspired—and was used to inspire—the greatest capital misallocation in history.

Although this will annoy some people, I am fully convinced that the only reason this has lasted so long is that OpenAI has not yet collapsed. Its failure would be a watershed moment—the Lehman Brothers of the AI bubble, an event that will define the end of one era and the beginning of another, shaking those infected out of their delusion. Without this wake-up call, NVIDIA continues to sell GPUs, the coffers of the semiconductor industry continue to swell, and more spending commitments are made.

OpenAI plans to burn through over $852 billion by the end of 2030. It accounts for $748 billion of Microsoft, Amazon, and Oracle's remaining performance obligations, in addition to at least $70 billion in RPOs from Cerebras, CoreWeave, Nebius, IREN, Lambda, and Nscale (according to Kakashii), and plans to spend an undetermined tens of billions on Broadcom's "Jalapeno" chip. It plans to spend $50 billion or more on compute this year alone, which I estimate is over 50% of global AI compute spending (OpenAI accounts for over 50% of all AI compute infrastructure).

OpenAI can afford these expenses only because of its latest round (presumed fully completed) of $122 billion in funding, of which at least $50 billion has been received, including $20 billion from SoftBank (totaling $30 billion, with the third tranche due on October 1, 2026). NVIDIA mentioned in its latest quarterly earnings that it "estimates that one AI research and deployment company, by purchasing cloud services from its customers, contributed a significant share to [its] revenue for the first quarter of fiscal 2027," referring, of course, to OpenAI.

The AI Bubble is the OpenAI Bubble—Heading Towards an Inevitable End

OpenAI is the reason everyone is paying attention to AI. In March 2019 (according to JustDario), NVIDIA acquired a company called Mellanox, which produces the high-speed networking technology needed to create AI GPU clusters. Four months later, Microsoft invested $1 billion in OpenAI and began purchasing AI GPUs and building AI infrastructure for it. By March 2020, NVIDIA shipped its A100 GPU, and in May 2020, Microsoft announced it had built a supercomputer for OpenAI with "over 285,000 CPU cores [and] 10,000 GPUs."

ChatGPT launched in November 2022, perfectly timed for a tech industry that had run out of ideas and was teetering on the brink of a prolonged downturn. The IPO market had collapsed, interest rate hikes ended the zero-interest-rate era, pandemic-era over-hiring was ending with some of the worst layoffs in industry history, global venture capital shrank after historic over-investment in 2021, and tech stocks were hit hard.

For the first time, the tech industry was forced to live within its means—something it has historically been unwilling to do. Big tech companies were unpopular with investors and the public. The excesses of the past decade—along with growing frustration over "tech exceptionalism" (the idea that the rules governing the rest of the world don't apply to Silicon Valley)—tested the patience of regulators and lawmakers. And, lacking "the next big thing"—a sensational, game-changing product category—it no longer had an excuse for profligate spending or regularly breaking the unwritten and written rules governing society.

OpenAI's existence justified an era of frenzy and extravagance. Hyperscalers, lacking new hyper-growth ideas, pointed to ChatGPT's "fastest-growing user base ever" and the Microsoft "supercomputer" that built it, telling investors they would be left behind if they didn't invest. Amazon, Meta, and Google announced their own vague "supercomputers" in 2023.

By the end of 2023, NVIDIA had sold 500,000 A100 GPUs, and the only reason it could do so was ChatGPT's rapid growth. Sam Altman's brief ouster only further inflated the AI bubble—adding a layer of dreary palace intrigue to a tech industry lacking imagination and personality—and further solidified Microsoft's role as OpenAI's paternalistic funder, ensuring Altman's return to the helm.

To be clear, by "rapid growth," I mean OpenAI reached 100 million weekly active users by the end of 2023, with monthly revenue of around $108 million. Microsoft invested another $10 billion that year, most of it in the form of Microsoft Azure credits.

OpenAI is also the reason Anthropic exists—not only because several of its founders came from the company, but also because Google and Amazon both agreed to provide a total of $6 billion to it in 2023 as a means to "compete" with Microsoft's new darling, giving both further reason to spend hundreds of billions more to "not miss out on AI."

When you take the word "AI" out of the equation, it all seems a bit absurd. $16 billion in equity investments, plus over $150 billion in capital expenditure by the end of 2023, all essentially justified because one website became very popular.

And the only reason these two companies could grow was that hyperscalers funded their entire infrastructure.

In the fourth quarter of 2023, global venture capital fell to its lowest level since the third quarter of 2016, with US startups accounting for $183.6 billion of the year's total investment. Venture capital itself could not—and would not—truly support the scale of infrastructure that OpenAI or Anthropic needed. Without the hyperscalers inflating the value of both companies, there wouldn't be any hunger from hyperscalers or those providing debt financing for data centers—and this is almost entirely due to OpenAI's success.

If you remove OpenAI from the period between 2020 and 2024, the AI bubble would never have inflated. No other major AI company showed any signs of life—whether being peddled by hyperscalers, funded by venture capitalists, or launched by other tech companies.

The only reason any hyperscaler's AI efforts can generate revenue—and it's quite meager outside of OpenAI and Anthropic!—is because they know they just have to sit there and keep saying "AI is the future" until customers eventually give in and try it… largely because everyone is talking about ChatGPT.

Anthropic wasn't considered a serious contender until early 2025, and it could only sustain funding because people wanted to invest in the next OpenAI. Anthropic's initial funding rounds and infrastructure buildout were justified solely on the basis of competing with OpenAI.

The $178.5 billion in US data center debt transactions in 2025? Almost entirely justified by OpenAI's growth and its insatiable demand for compute power, because, aside from OpenAI (and later Anthropic), no one else was using massive clusters of tens of thousands of GPUs, and demand for compute at this scale didn't seem to materialize for months or even years afterward.

The biggest consumers of compute remain Microsoft (for OpenAI), Google (for Anthropic), Amazon (for OpenAI and Anthropic), CoreWeave (for OpenAI and Anthropic), Meta (imitating other hyperscalers), and Oracle (for OpenAI). Otherwise, there is little evidence—and I have looked carefully—that demand for AI compute exceeds a few tens of billions of dollars, which is being generous.

All of this investment—whether in AI startups or data centers—exists to fund the next OpenAI, or to become the landlord of the next OpenAI.

The assumption—because no one ever seriously thinks about it—is that since one OpenAI exists, more OpenAIs will spring forth. Since one big compute customer exists, the template for future compute-intensive startups has been established… and, again, because no one ever seriously thinks about anything, no one realizes that a second OpenAI hasn't appeared because OpenAI and Anthropic are a financial psychological warfare operation orchestrated by the world's largest software companies.

OpenAI and Anthropic Are Hyperscale Psychological Warfare Experts Tailored for Silicon Valley's Monoculture

The brutal truth is that you can't fund an AI lab through venture capital. Although OpenAI and Anthropic have raised nearly $300 billion in the past few years, their actual infrastructure costs—the GPUs and data centers powering their services—are entirely funded by hyperscalers, who have probably spent another $250 billion in the process, given that Microsoft said it had spent $100 billion on its relationship with OpenAI by early 2026.

However, the true cost isn't just financial; it includes the experience and industrial knowledge required to execute a massive infrastructure bailout. Besides Google, Microsoft, and Amazon, no other company has the scale or experience in building the kind of AI clusters that OpenAI (and later Anthropic) needed.

We know this for several reasons. First, because before 2023, almost no companies were building the type of AI compute clusters that OpenAI or Anthropic needed. The closest might have been cryptocurrency mining companies, and it's telling that many of today's new cloud service providers (most notably CoreWeave) originally operated warehouses filled with ASIC chips to mine Bitcoin and Ethereum.

Second, according to conversations with data center industry insiders, the entire Overton window of what constitutes a "large" facility has shifted. Previously, a 50MW data center would have been considered a significant (even noteworthy) development. These were exceptions, not the rule; most data centers were much smaller. The only companies with experience building at this scale were mostly the hyperscalers themselves.

By presenting OpenAI as a "venture-backed startup," hyperscalers created the illusion that it was the next big company, which would in turn create the next huge demand center for cloud computing—except the only reason these companies exist is that the hyperscalers themselves are facilitating their existence, funding them with massive amounts of money, and allowing them to burn cash at will.

This is why the idea that OpenAI will continue to grow indefinitely is central to the AI bubble myth. The existence of one OpenAI allows others—no matter how illogically—to imagine the existence of more OpenAIs, which in turn means those OpenAIs will need as much compute as OpenAI.

The foolish investors who believe this nonsense can rationalize it through countless buy-side analysts or bought-off media figures who talk about "insatiable demand for compute," pointing to capacity constraints (caused by slow data center construction and OpenAI/Anthropic consuming most of the world's compute) and rising GPU prices as evidence of massive actual demand, without ever truly thinking deeply.

The greatest trick the hyperscalers ever pulled was never backing down. By pouring over a trillion dollars into AI capital expenditure without showing a dollar of profit, they provided justification for anyone to invest in AI data centers, with the logic being "the world's largest companies can't all be wrong"—even though the reason they did it was to expand capacity for OpenAI and Anthropic, two companies they themselves incubated.

The hyperscalers spending this much money on AI infrastructure is fundamentally illogical and insane, and the reason so few are willing to say so is that, until recently, suggesting it was a waste of money was considered radical—almost entirely because of OpenAI's existence and continued growth.(Note: I realize Anthropic has garnered significant attention and grown quickly in the past year, but it could do so A) because of the OpenAI myth, and B) because it, too, was incubated and allowed to operate at massive losses.)

Whether you derive utility from LLMs or not is irrelevant, because that's not the actual basis for data center investment. While accelerated progress in code generation (which itself is only possible with massive subsidies) may have helped Anthropic grow, the vast majority of data center capital expenditure is chasing the phantom of what AI might become, rather than being tied to any company's overall revenue or economic situation—aside from their compute spending, of course.

This is the underlying greed driving this wasteful, reckless, and destructive era: the belief that there will be another OpenAI, and, as I said, the opportunity to become the landlord of the next OpenAI. And because the media and analysts rarely have original ideas, everyone rationalizes (and continues to rationalize) this waste using the same clichés, saying it's "like Uber (it's not!)" or "like Amazon Web Services (Amazon spent $29.7 billion in capital expenditure between 2003 and 2015, adjusted for inflation)."

Like any great investment bubble, the more money that pours in, the stronger the FOMO, the more dollars can be rationalized, and the more complex and distorted the myth becomes—which is why you see prominent venture capitalists claiming AI labs have "inference margins of over 90%," a completely unsubstantiated claim that AI enthusiasts cling to and repeat so often it becomes accepted as truth, likely because they don't want to face the fact that you can burn through $14,000 worth of tokens on a $200/month ChatGPT subscription.

This myth only grows in an environment deliberately starved of good information. We've been stuck in this terrible bubble for four years and still have no consensus on the actual cost of large language models, which is testament to the industry-wide effort to suppress this information.

OpenAI, Anthropic, Microsoft, Google, and Amazon have gone to great lengths—according to discussions with sources familiar with their infrastructure—to obfuscate the true underlying costs of their operations, and Silicon Valley, an industry supposedly composed of free thinkers and individuals, is all too happy to accept any convenient myth that might sustain its dreams.

Ultimately, they all became useful idiots for the hyperscalers. Their obsessive attachment to OpenAI—and by extension, Anthropic—appears to be decisions made in the name of "democratizing powerful AI," when in reality every dollar flows to Microsoft, Google, Amazon, or Oracle, which then pass the money to NVIDIA or Broadcom, which then pass it to TSMC, SK Hynix, Samsung, or Micron.

Investing in AI startups? They pay an AI lab, which then pays a hyperscaler. Investing in AI infrastructure companies? Money flows to NVIDIA, then upstream to semiconductor companies. Ultimately, whether they die or get acquired (because none of them will go public), all value ends up in the hands of the hyperscalers who created this ephemeral era and then inflated it into something very dangerous.

Why the AI Bubble Cannot Survive Without OpenAI

Yet the problem is that this industry cannot survive without OpenAI under any circumstances.

When people discuss the possible collapse of OpenAI, they act purely cowardly, either saying "it won't be that bad" or vaguely calling it "too big to fail."

If OpenAI—the company with the most funding, the most infrastructure, the most attention, and the most AI talent—collapses, it will likely happen after AI data center debt and venture capital funding are almost completely exhausted.

You see, Goldman Sachs' Jeffrey Papai recently noted that replicating the hundreds of billions of dollars raised by hyperscalers over the past four years would be "very difficult"—$244 billion in 2026 alone, if you count NVIDIA and SpaceX—which is a problem because, as of Q3 2026, they can no longer fund their data center capital expenditure with cash flow.

And to be clear, hyperscaler capital expenditure doesn't have to stop completely for NVIDIA to stumble. It just needs to slow