Not a Bubble: The AI Boom Has Two Customers

Microsoft's fiscal 2026 annual report contains a line the company had never published before. Revenue from its OpenAI partnership: $24.1 billion.

That is the first hard number anyone outside the two companies has had. It is also, on the available analysis, roughly 70% of everything Microsoft books as AI revenue. The largest software company in the world has an AI business, and most of that business is one customer.

What the filing actually says

The $24.1 billion is not one line item. It bundles Azure capacity that OpenAI itself buys, revenue-sharing payments, and other commercial agreements between the two companies. Against Microsoft's $331.8 billion of FY26 revenue that is about 7% of the whole business - meaningful, not existential.

The AI-specific share is where it gets interesting. Analysis of the disclosures puts OpenAI at roughly 70% of what Microsoft reports as AI revenue. Strip out the partner and the AI business is a fraction of the number that gets quoted in earnings coverage.

The filing also shows about $6 billion of accounts receivable from OpenAI as of 30 June 2026. That is money booked but not yet collected, from a company that is itself spending heavily to grow.

None of this is hidden or improper. It is a large commercial relationship disclosed exactly as it should be. It is simply more concentrated than the phrase "AI boom" suggests.

The compute contracts underneath

Anthropic's arrangements make the same point from the other direction.

In May 2026 Anthropic signed a compute deal with SpaceX - which now owns xAI - for the full capacity of the Colossus 1 data centre in Memphis: over 300 megawatts and more than 220,000 Nvidia GPUs. The price is $1.25 billion a month, locked in through May 2029, totalling over $40 billion.

In August it signed a second, six-year, $10 billion agreement with Volta Infra Holdings for 121 megawatts of Nvidia Vera Rubin capacity at Bitdeer's Tydal campus in Norway. Volta was founded in January 2026. The deal is backstopped by $1.3 billion of credit from JPMorgan affiliates and a second institution.

Read the Volta deal slowly, because the structure is the story. Nvidia is an investor in Volta and the chip supplier to the site. Dell is an investor in Volta and the hardware supplier. A bank underwrites the counterparty risk of a company seven months old, so that a model lab can commit to capacity that does not exist yet - first phase targeted for the end of 2026.

Every participant in that chain books real revenue. The chip vendor, the hardware vendor, the data centre operator, the cloud reseller, the bank collecting fees on the backstop. Anyone who has costed AI infrastructure at any scale will recognise the shape: capacity is bought years before it is used, and the commitment is fixed long before the demand is. All of it is legitimate. Almost all of it traces back to demand originating at two companies.

Why "bubble" is the wrong word

A bubble is too many buyers paying too much for something. Housing in 2007. Dot-com equities in 1999. The defining feature is a crowd.

This is close to the opposite. There is no crowd. There are two companies whose growth underwrites a very long chain of committed spending, and a set of infrastructure providers who have taken multi-year positions against that growth continuing.

That distinction matters because the two failure modes look nothing alike. A bubble deflates when sentiment turns - broadly, visibly, and with plenty of warning in the price. Concentration risk does not deflate. It transmits. One buyer changes its capacity plans and the effect arrives at the far end of the chain intact, at a company that never spoke to that buyer and has no idea why its forecast just moved.

Note what is not being claimed here. Nobody serious argues the demand is fake. Anthropic's revenue has grown fast enough that the compute commitments look defensible against its own trajectory. The question is not whether the demand exists. It is what happens to a chain of multi-year, take-or-pay commitments if growth at the origin merely flattens - not reverses, flattens.

What this means if you are building on AI

Most businesses reading this are not exposed to any of the above directly. You are not signing gigawatt contracts. The exposure arrives through your vendors, and it arrives as three specific risks.

Your unit costs are set by someone else's negotiation. Per-token pricing has fallen steadily, and most teams have quietly built financial models that assume it keeps falling. That assumption is a bet on capacity staying abundant. Anyone who costed an AI agent at last year's prices and shipped it at a fixed monthly fee has already discovered what happens when the assumption slips.

Model availability is not a contract. Models get deprecated, rate-limited, re-tiered and re-priced. If an automated workflow depends on one specific model version behaving one specific way, you have a dependency you did not write down. This is the same lesson the LiteLLM supply chain attack taught about credentials: the risky dependency is the one nobody listed.

Your vendor's roadmap is not your roadmap. When a platform's AI features are largely a reseller relationship with a lab, that platform's pricing and capability changes are not really its own decisions.

Three things worth doing this quarter

Not a strategy. Three concrete pieces of engineering that cost days, not months.

Put a layer between your code and the model. One interface, provider behind it. Then swapping a model is a config change instead of a refactor. Teams doing serious LLM integration build this on day one, and everyone else builds it under pressure later. The cost of doing it upfront is roughly an afternoon.

Run the second-choice model once, for real. Not a benchmark - your actual prompts, your actual evaluation set, your actual outputs. You will find the differences that matter, and you will find them at a time of your choosing. Most teams discover their prompts are more portable than they feared, and their output parsing is far less portable than they assumed. Retrieval systems tend to travel best of all, because the expensive part is your index rather than the model reading from it.

Know your cost per unit of work, and what it does at double the price. Not per token - per resolved ticket, per drafted proposal, per enriched lead. If the answer is "the product is still profitable, margin drops", you are fine. If the answer is "we do not know", that is the thing to fix, and it is a spreadsheet, not a project. This is the first question we ask in any AI consulting engagement.

Everything above is portability work, and portability has a real cost: an abstraction layer is a thing to maintain, and the second model needs re-evaluating whenever either provider ships. We would not do it for a prototype. We would do it for anything a business depends on.

What we are not saying

We are not predicting a crash. The demand is real, the products work, and the revenue at the origin of this chain is growing rather than shrinking.

We are not saying avoid the frontier labs. Their models are the best available and using them is the correct call for most work. Nothing here argues for a worse model on principle.

We are not claiming to know the timing. Concentration risk is a statement about structure, not schedule. It can persist for years. The buildout that looks reckless from outside frequently looks obvious in hindsight.

What we are saying is narrower. The AI market has fewer real buyers at its origin than the coverage implies, most of the committed spending traces back to them, and the sensible response is not to sit out. It is to make sure that if the chain moves, the thing you built can move with it.

That is mostly ordinary engineering discipline - abstraction, evaluation, knowing your costs. It is the same discipline that survives a vendor being acquired, an API being deprecated, or a price list changing on thirty days' notice. AI has not introduced a new category of risk here. It has raised the stakes on an old one, and compressed the timeline for noticing.

If you want a view of where your own stack is exposed, that is a conversation worth an hour. It usually turns up two or three dependencies nobody had written down.