A Three-Year-Old Holding a Quantum Computer: Testing the Big Tech AI Bubble Against Real Revenue

To state the conclusion first: today's AI market has a huge gap between an extreme, supply-side performance race and a demand-side focus on practicality. But neither unconditional bubble theory nor unconditional rosy forecasts can stand up to the numbers. Based on the audited financial statements and earnings disclosures released in the first half of 2026, let us verify who is earning and who is burning, and where the money is actually flowing.

1. OpenAI: Earns $13 Billion and Loses $20.9 Billion

The audited financials leaked in June 2026 (obtained by Ed Zitron, verified by the FT) shattered the myth.

Three things are key. First, the single R&D line exceeded total revenue for two consecutive years. Second, $10.59B of that R&D was paid to Microsoft. Third, the $7.5B cost of revenue is inference cost — a variable cost that follows you as users grow. It is a structure that burns more the more it earns.

In 2026 it cuts both ways. Annualized revenue reached $40B (July, Sacra estimate), the enterprise share exceeded 50%, and advertising reached $1B annualized. But it burned $3.7B in the first quarter alone, and annual cash burn is projected at $27B in 2026 and $63B in 2027. Break-even is in 2030. It raised $122B in March, recognizing a valuation of $852B. With a denominator of 900 million weekly users and 50 million paid subscriptions, the market is still putting in money.

2. Anthropic: From $900 Million to $65 Billion in Seven Months

In Q2 2026 it posted $11.5B of revenue with an adjusted operating profit, filed confidentially for an IPO in June, and is being talked about at $2 trillion for a fall listing, having raised $6.5B in a Series H in May at a $965B valuation. 80% of revenue comes from enterprise customers, there are over 1,000 customers spending $1M+ a year, and the single product Claude Code is at a $2.5B annualized run rate. Because coding ties directly to payroll, enterprises do not skimp. According to a Menlo estimate, 54% of enterprise spending on coding models goes to Anthropic.

But three things need a discount. First, Anthropic reports on a gross basis while OpenAI reports on a net basis, so a direct comparison is impossible. Second, 79% of its customers overlap with OpenAI contracts, creating churn risk when budgets are consolidated. Third, the SpaceX Colossus lease runs at about $1.25B a month ($15B a year). Judgment should be withheld until the S-1 is public.

3. xAI: $500 Million in Revenue, $230 Billion of Value, a 460x Multiple

The star of the most extreme gap is xAI.

The paid conversion rate is dismal. Of 117 million monthly users, paid users of the advanced model number 1.9 million. And yet there is a twist. In May 2026 Anthropic signed a contract to lease xAI's Colossus 1 capacity for $1.25B a month. The total value over the contract term exceeds $40B. In other words, a competitor's data center earns more money than the competitor's model. xAI's real business may be GPU leasing, not Grok. With SpaceX's capital and X's distribution behind it, it will not collapse, but among AI labs it has the loosest financial discipline.

4. China: 15-20% of the US Scale, But at Twice the Speed

According to a September 2026 Rhodium Group analysis, China's AI CapEx doubles this year to 932 billion yuan ($139B) and exceeds 1.2 trillion yuan in 2027. That is 15-20% of the US level (about $800B).

China suffers from the same disease. The combined free cash flow of Alibaba, Tencent, and Baidu swung from 170 billion yuan in 2025 to minus 16 billion yuan in the first half of 2026. The US big five also plunged from $191B to $14B, and the shortfall is covered with debt (from $90B to $163B). The funding line of the AI boom is the bond market.

5. The Inference Price War: Token Prices Fell 40%

The effective enterprise token price (per million tokens), which was $1.15 in March 2026, fell 41% to $0.68 in September. OpenAI cut GPT-5.6 by up to 80%, and Chinese open source (the DeepSeek family) adds downward pressure. The corporate response is a flight from the flagship. The flagship share fell from 53% in early August to 45% in September, and per-seat spending at the top 1% of spenders dropped 10% in a single month.

6. The Inference Paradox: Tokens Got Cheaper but the Bill Tripled

Even as unit prices fall, the total rises. This is because agents use 10-100 times the tokens of a chatbot. In a McKinsey survey, 93% of companies exceeded their AI budgets, and in an MIT analysis, 95% of corporate pilots had no measurable profit-and-loss effect. Gartner calls this the inference paradox. Agentic inference costs providers more than 5x what chatbots do, and it expects that to rise 5x more by 2028.

The implication for individuals is clear. The same task varies 5-100x in cost depending on routing. Haiku is 5x cheaper than Opus, GPT-5 mini 6.7x, and DeepSeek 100x. The skill is not in using an expensive model but in picking out the work that a cheap model can handle.

7. The Flow of Money: The Truth of Stargate and the Power Bottleneck

Dig into Stargate (OpenAI, SoftBank, Oracle; targeting $500B by 2029) and the committed capital is only the initial $100B, with the remaining $400B to be raised. It is a ceiling, not a floor. The real bottleneck is neither money nor chips but power. 10GW equals ten nuclear plants, and Microsoft's $80B of unfulfilled Azure orders are also because of power. That is why nuclear PPAs and even gas plants are being mobilized.

8. Synthesis: Three Scenarios

Optimistic: demand eats supply, break-even comes in 2028, and today's CapEx remains like the fiber-optic cables of the 1990s. The evidence is backlog (Google $240B, Oracle $523B) and capacity that is consumed immediately.

Neutral: a price-normalization phase. Falling token prices and rising agent consumption offset each other, and only those who get unit economics right (Anthropic, Google) survive. It resembles the digestion period of the 1999 telecom bubble.

Pessimistic: a 2027 correction as the ROI gap, debt maturities, and a GPU glut (bullwhip) converge. The trigger is enterprise budget consolidation and the default of a neocloud.

Whichever it is, the direction is the same. From a parameter race to a unit-economics race, from high performance to substance. The value-for-money practical agent that protects the user's security, costs less, and handles the drudge work is the protagonist of the next two years. The bubble does not pop so much as substance takes over the job the bubble was doing.

9. The Business Model Gap: So Everyone Just Shouts "Coding"

Take apart each AI company and there is not really a business model. Subscriptions and API billing — those two are everything. So the words every company chants in unison are "coding." Claude Code, Codex, Copilot, Cursor — all of them package coding automation as though it were the greatest product ever.

But let us ask. How many people actually code? No nationwide coding transition has taken place, yet they advertise as though the whole population will use it. The coding market is the exclusive preserve of a limited population called developers. Even generously, the global developer population is around 30 million, and of those, the number who will pay $20 a month for a paid coding tool is a fraction. Stacking hundreds of billions of dollars of valuation on this narrow market is the current landscape.

What is more serious is that practicing engineers do not use it because of company confidentiality. A company's core source code is a security asset. Pasting company code into an external cloud AI is a security violation at most companies. Finance, defense contractors, and large corporate research labs are closed networks by default. In other words, the enterprise customers who could pay the most are precisely the ones who cannot use cloud coding AI. This is not saying the market is small. It is saying the profitable segment is blocked off.

10. The Second Product, Image Generation: Curiosity Is Not Payment

After coding, the next thing companies push is image and video generation. At first it is novel. A few lines of text produce a plausible picture, so everyone tries it once. The operator himself used it at first because it was novel. But now he does not use it at all. This is precisely the trajectory of the public.

The demand for image generation is curiosity, not payment. An ordinary person who is neither a company staffer nor a designer has no reason to pay monthly to churn out images. You make a profile picture once, make a few cover slides for a presentation, and that is it. Video generation is worse. Compute is expensive, so companies cannot release it for free, and users play with it once or twice and leave. It is a product to judge by churn rate, not retention.

The numbers prove it. The paid conversion rate of image generation services stays in the single digits, and the resubscription rate is even lower. One viral moment accumulates sign-ups, but it does not translate into billing. Curiosity traffic only eats away at server costs.

11. The Shelf Life of a Craze: Novelty Lasts Three Months

Look at the technology adoption curve. The explosion right after ChatGPT's launch, the virality of image generation, the fervor over coding agents — they are all the same pattern. The peak is three months after launch, and after that only real users remain. Two groups stay: those who attach it to their work and make money, and those who play with it as a hobby. Everyone in the middle churns out.

Today's AI industry revenue leans on the curiosity payments of that middle layer. Curiosity is not refillable. No one is amazed twice by the same thing. So companies must keep manufacturing stronger novelty, and that cost is the hundreds of billions of dollars of CapEx we saw earlier. It is a structure that makes up for the inflation of novelty with the inflation of cost.

So how long does this craze last? The operator's answer is simple. Until the substantial tools remain. Coding or images — only the uses that produce results worth paying for survive. The rest evaporate at the speed at which curiosity cools. A bubble collapse does not arrive suddenly. It quietly goes out in proportion to how much curiosity cools.