In 2025, OpenAI spent $34 billion and brought in $13 billion in revenue, netting a loss of $39 billion. For every dollar earned, it spent two dollars and sixty cents. These numbers alone are enough to make anyone ask: is this a bubble?
The problem is that “bubble” is not one thing. The dot-com bubble was one kind. The 2008 subprime crisis was another. They burst in completely different ways with completely different social costs. The AI industry right now carries three different kinds of bubble risk simultaneously, each with its own way of bursting. Mixing them together produces nothing but a vague anxiety about “will it crash.” Looking at them separately tells you what to worry about and what indicators to watch.
The most intuitive bubble is “too much money spent, can’t get it back.” But this AI cycle has one crucial difference from the 2000 dot-com bubble: the people losing money are different.
During the dot-com bubble, companies could operate with a website. Fixed assets were almost zero. Funding came from equity — venture capital and retail investors. When the bubble burst, stock prices dropped 90%, investors took their losses and left. The damage stayed locked with shareholders and didn’t spread.
AI is different. Building an AI data center requires land, buildings, GPUs, power grid connections, and cooling systems. Every item is hard physical infrastructure. NYU’s valuation professor Aswath Damodaran calls this the largest infrastructure expansion in his career, comparable in scale only to the rise of the automobile industry a century ago. Where does the money come from? Not all equity. The largest cloud providers have been collectively issuing debt: Alphabet issued about $85 billion in bonds, Oracle nearly $50 billion, Amazon $37 billion, Meta $25 billion, and even Nvidia issued $20 billion. Alphabet stopped its stock buyback for the first time in a decade, pouring all its capital into AI infrastructure.
What’s the difference between an equity bubble and a debt bubble? When an equity bubble bursts, shareholders lose money and that’s the end of it. When a debt bubble bursts, companies can’t repay, defaults spread outward along the lending chain — from the banks that lent to the funds that bought the bonds to the entire credit market. Damodaran puts it bluntly: the social cost of debt that can’t be repaid is far greater than a stock price dropping 90%.
When should you worry? Not about whether Nvidia’s stock is still rising. Watch whether the bond spreads of smaller, weaker borrowers like Oracle and Super Micro are starting to widen. A Yahoo Finance analyst called them the canary in the coal mine. The canary hasn’t fallen yet, but the cage is already down in the mine.
The second risk doesn’t require waiting for a crash. It’s happening right now. It has nothing to do with valuations and everything to do with capital concentration.
In 2025, Amazon announced a $50 billion investment in OpenAI. Shortly after, Amazon’s film studio completed production on a biographical movie about OpenAI and Sam Altman, then decided not to release it. The stated reason was “another distributor would be a better fit,” but multiple media outlets directly pointed out: you just put $50 billion into this company, releasing a movie that portrays its boss as a comedic character isn’t great timing. A tech giant using its capital relationship to determine whether a movie about another tech giant can reach audiences — that itself is a form of bubble. It’s not a valuation bubble; it’s capital relationships growing large enough to distort things that should be independent.
Something similar happened in chips. An AI startup called Odyssey, which builds world models, had its Series A led by Nvidia four months ago, using Nvidia GPUs. Four months later in its next funding round, Nvidia exited, Amazon and AMD came in, and the chips switched from Nvidia GPUs to Amazon’s own Trainium chips. The switch wasn’t because Trainium performed better — it was because taking Amazon’s money meant using Amazon’s chips. Chip selection shifted from “whose is better” to “whose investment terms are tighter.”
This funding round also included In-Q-Tel, the CIA’s venture fund. World model technology now carries national defense implications. When national security technology procurement starts embedding into tech stacks locked by big tech capital relationships, the boundary between commercial competition and geopolitical competition blurs.
Put these events together and the pattern is: capital injections lock in technology choices, technology choices lock in narrative control, narrative control locks in defense pathways. You don’t need to wait for a crash to judge this kind of bubble — you watch whether capital relationships have started redistributing power in media, chips, and defense, areas that should be governed by market competition or public decision-making. It has already begun.
The first two risks address “what if AI fails” and “even if it doesn’t fail, there are problems.” The third asks: “what if AI succeeds?”
Microsoft CEO Satya Nadella wrote a long post on X raising a question most people don’t think about when discussing AI bubbles: suppose a few AI models actually succeed, absorbing all the professional knowledge from every industry, concentrating value from thousands of sectors into a handful of model providers — what happens then?
His answer: society won’t tolerate it.
His analogy is the manufacturing offshoring of the late 20th century. GDP numbers looked healthy at the time, but the hollowed-out industries, declining communities, and political polarization left behind took twenty years to erupt and are still festering. Nadella’s concern is that AI could do the same thing with professional knowledge. The difference is that globalization took twenty years to manifest; AI’s pace of concentration is much faster.
This kind of bubble isn’t a valuation bubble — it’s a value concentration bubble. It doesn’t burst through stock corrections or debt defaults. It bursts when regulation, antitrust action, or social backlash externally interrupts the concentration. The trigger for this backlash is precisely AI being too successful: the stronger the models, the more professional knowledge they absorb, the broader the industries they replace, the greater the political and social pushback.
What’s striking is that Microsoft and OpenAI, the closest allies in AI, have placed completely opposite bets on this risk. Nadella’s strategy is to let every customer own their own AI capabilities, keeping value on the customer side rather than concentrating it in a few models. Meanwhile, Microsoft has started testing DeepSeek as a replacement for some OpenAI models. OpenAI has gone the opposite direction: concentrating value in a few frontier models, making every industry dependent on them through APIs. Allies betting in opposite directions at the same moment signals that this risk has grown too large for even the closest partners to maintain consensus.
Three risks, three ways of bursting, three timelines, three sets of indicators.
Type one (debt transmission) is a near-term risk after failure. Watch whether Oracle and Super Micro’s bond spreads are widening, and whether big tech is cutting data center projects due to capital constraints. Type two (capital relationship distortion) is happening now. Watch chip-locking terms in investment deals, transparency of capital relationships between content industries and AI giants, and whether defense procurement is starting to depend on capital-locked tech stacks. Type three (value concentration backlash) is a medium-to-long-term risk after success. Watch antitrust actions against AI concentration, industry-level AI usage restriction legislation, and signs of political mobilization in labor markets.
Judging whether AI has entered late-stage bubble territory requires looking at all three simultaneously, not just one. Currently, all three signal chains are still in the signaling phase, not the bursting phase. But Oracle’s bond spreads, Amazon’s movie cancellation, KPMG’s fabricated case studies, and Microsoft’s strategic pivot — these signals appearing in dense clusters within the same six-month window carry more weight than any single signal. Whether you bet right on whether AI will succeed matters less than how many indicators you’re tracking at once.