With OpenAI moving closer to an IPO, a roughly 15,000-word blog post has pushed the AI bubble debate back to the center of market discussion. Ed Zitron, a longtime AI bear with a large technology-sector readership, made one of his sharpest calls yet: the real bubble is not AI in the abstract, but an “OpenAI bubble.”
In Zitron’s framing, if OpenAI were to fail, it could become the AI era’s version of Lehman Brothers. He argues that such an outcome would not only break the logic behind current AI investing, but could also trigger a broad repricing across data centers, AI infrastructure and global technology stocks.
The argument quickly drew attention from financial media. The key issue, as the coverage describes it, is not whether AI has value at all. It is whether OpenAI has a business model strong enough to support the current AI capital cycle. If the answer is no, then the financing structures, compute investment plans and capital spending programs built around OpenAI could all face knock-on pressure.
That view is far from universal. Investors including Howard Marks, co-founder of Oaktree Capital Management, have recently said they now place more weight on AI’s long-term value as a general purpose technology than they did when they were more skeptical and saw a greater risk of a pure bubble. In that camp, the industry is still in an early phase of commercialization.
From an AI bubble to an OpenAI bubble
Zitron’s argument differs from broader claims that the entire AI sector is overheated. He narrows the concern to one company. In his view, OpenAI has become the “credit anchor” of the generative AI era since ChatGPT broke out in late 2022.
He says a series of market assumptions now rest on OpenAI continuing to grow quickly: that AI will change the world, that hyperscale data centers are worth building, that GPU demand will remain on a steep growth curve, that large model companies will eventually become profitable, and that AI startups will create enough end-user demand to justify the spending.
Under this logic, OpenAI has done more than define the current AI boom. It has also shaped how capital markets value the broader AI supply chain. If that central assumption fails, the shock could spread well beyond a single unicorn. Zitron’s description is that OpenAI now looks less like one company and more like a systemically important institution inside the AI investment cycle.
Why he sees flaws in the business model
Zitron’s criticism is built around three main points.
Inference costs are still high
First is inference. As ChatGPT’s user base expands, every query adds GPU, power and server expense. If a large share of users remain on low-priced or free tiers, and enterprise revenue growth does not rise fast enough to absorb those costs, then scale can come with wider losses rather than better operating leverage.
Capital spending is outrunning cash-flow improvement
Second is the gap between spending and cash generation. The article says the biggest outlays in AI are no longer centered on model training alone. They now include inference compute, GPU procurement and global data center construction. OpenAI and its partners are backing data center investment worth tens of billions of dollars or more, and projects of that size often take years to pay back. If AI demand grows more slowly than expected, utilization across that infrastructure could weaken.
Reliance on external financing may persist
Third is funding dependence. Zitron argues that OpenAI will likely need to keep raising capital for years to cover model development, compute purchases and infrastructure buildouts. If market risk appetite falls or financing conditions tighten, that reliance could become a more serious vulnerability.
The report makes clear that these are Zitron’s own judgments and have not been endorsed by OpenAI. Even so, they reflect a live market debate over AI return on investment, or ROI.
Why Oracle, CoreWeave and data center operators are in focus
Zitron is even more concerned about leverage effects across the supply chain than about OpenAI alone. Over the last two years, the U.S. technology sector has gone through a data center construction surge on a historic scale. Microsoft, Google, Meta and Amazon have all raised capital spending, while companies such as Oracle and CoreWeave have taken on a larger role in building AI compute capacity.
Those projects rely heavily on long-term leases, project finance, private credit, corporate debt and large capital commitments. If demand from major customers such as OpenAI comes in below expectations, or if investors reassess AI ROI, then data center utilization, lease structures and financing capacity could all come under pressure.
Media coverage of the essay says Zitron sees Oracle and CoreWeave as particularly exposed if OpenAI suffers a major setback, because a large part of their premium valuations has rested on the assumption that AI infrastructure demand will keep surging.
At the same time, major technology companies including Microsoft, Meta and Alphabet are still increasing AI capital spending and continue to say that infrastructure investment fits their long-term strategy. There is no clear sign yet of a broad pullback in capital expenditure.
Anthropic and SoftBank are also part of the debate
Zitron also turned to Anthropic. His reasoning is that while Anthropic and OpenAI are following different paths, both require sustained spending on model development and compute procurement, and both depend on large technology companies for computing resources and financing support. If AI commercialization moves more slowly than expected, both could face profitability pressure.
SoftBank has also been mentioned repeatedly. In recent years, SoftBank has returned to the front line of major AI investment, taking an active role in financing AI infrastructure, chip and model companies. If the sector moves into a valuation adjustment cycle, SoftBank’s broad AI asset portfolio would naturally come under closer watch. For now, though, SoftBank remains firmly committed to AI’s long-term development and treats it as a key direction for the next wave of technology change.
Wall Street is still split on whether the AI trade is overheated
The broader argument over whether AI has entered bubble territory has already been running on Wall Street for more than a year.
- The bearish side says infrastructure investment is growing much faster than revenue, the business model for large language models is still not fully proven, data center capital spending has reached record levels, and market valuations depend more and more on growth projected years into the future.
- The more constructive side argues that AI is a classic general purpose technology. In that reading, it resembles earlier shifts such as the internet and electrification, where upfront investment often ran far ahead of near-term returns before creating new industries and business models.
Howard Marks recently said he has moved away from his earlier concern that AI might be just another bubble and now gives greater weight to its long-term value. He said modern AI shows reasoning, contextual understanding and interaction capabilities with characteristics that are difficult to compare directly with past speculative manias.
Some academic research has reached a more balanced conclusion. It suggests the current AI market contains both real technological progress and pockets of overheated valuation and front-loaded capital spending. On that view, the environment looks more like a technology revolution with local bubbles than a simple speculative frenzy.
The real question is cash flow and profitability
Whether or not investors agree with Zitron’s conclusion, the questions he raised are gaining traction: when will AI spending turn into stable cash flow?
Over the past year, capital markets largely treated higher AI capex as a good in itself. More recently, investors in chip stocks, server makers and cloud companies have started paying closer attention to another set of measures: AI revenue growth, paid conversion for AI products, the pace of inference cost declines, data center utilization and the payback period on AI investment.
If those metrics keep improving, current spending levels may eventually be seen as a forward-looking investment cycle similar to the internet era. If commercialization keeps lagging behind investment expansion, the valuation logic behind the AI trade may need to be recalibrated.
That is why the essay’s impact goes beyond the headline question of whether OpenAI will become the next Lehman Brothers. It has put the core issue of the AI era back in front of investors: after capital spending has repeatedly hit new highs, can cash flow and profitability keep up? The answer may do more than settle an argument. It may help shape the direction of global AI trades over the next few years.