$725 Billion of AI Capex This Year. Only Half of This Market Is a Bubble.
Microsoft, Alphabet, Amazon, and Meta are on pace for roughly $725 billion in combined capital spending in 2026 — up 77% from last year. Add Oracle and the figure reaches about $750 billion, equal to 38% of those companies' combined revenue.
Michael Burry is short semiconductors. Moody's has flagged roughly $662 billion in data center lease commitments that have been signed but not yet commenced. Free cash flow across the hyperscalers is being consumed by capex fast enough to push Amazon negative and compress the other three simultaneously.
The word "bubble" is now unavoidable in any conversation about this market. We think the word is being applied too broadly, and that the imprecision is causing founders to make bad decisions in both directions.
Separate the two layers
The AI market is not one market. It is an infrastructure layer and an application layer, and they are in completely different conditions.
The infrastructure layer has textbook bubble characteristics. Spending is growing far faster than attributable revenue. The capital is being committed on multi-year horizons against demand forecasts that assume a continuation of the current adoption curve. Participants are spending because competitors are spending — the arms-race logic where standing still is read by the market as surrender. Lease commitments signed but not commenced are a particularly clear tell: that is capacity being reserved against demand nobody can yet name.
The application layer — companies selling AI-powered software to customers who pay for it — does not look like that. Those businesses have revenue, churn, gross margin, and sales cycles. Some are excellent and some are bad, in roughly the proportions software businesses have always been excellent or bad. There is froth in their valuations, but froth in valuations is not the same thing as a structural bubble in the underlying activity.
Conflating the two leads founders to conclude that "AI is a bubble" means their AI application company is a bubble company. That does not follow, and acting on it — slowing down, under-raising, hedging the positioning — is how you lose a market to someone who read the situation more carefully.
The strongest argument against a bubble
Intellectual honesty requires engaging with the other side, and the other side is not stupid.
JPMorgan Asset Management points at a data center vacancy rate of roughly 1.6%. That is not a market with excess capacity; that is a market where everything built is immediately absorbed. The bear case requires demand to stop, and demand has not shown any sign of stopping.
The second argument is subtler and better. AI accelerators run on a three-to-four-year refresh cycle. In a conventional infrastructure bubble — railways, fiber in 1999 — the overbuilt asset sits there for a decade as stranded capital, depressing prices for everyone. A GPU does not do that. It becomes obsolete fast enough that overcapacity converts into obsolescence rather than a decade-long glut. The downside case is therefore shorter and shallower than the fiber analogy implies, even if it is sharp when it comes.
Our honest position: we do not know. Anyone who tells you confidently how this resolves is selling something. What we are confident about is that the shape of the risk is concentrated in the infrastructure layer, and that the exposure of a typical application company to that risk is indirect and mostly manageable.
What a correction would actually do to you
This is the useful exercise, and it takes twenty minutes.
If infrastructure capex corrected hard, the first-order effect on an application company is that compute gets cheaper, not more expensive. A capex correction means overcapacity means falling inference prices. For a company whose COGS is dominated by tokens, a bust in the infrastructure layer is a gross margin event in the right direction.
The second-order effects are where the damage is. A hard correction in AI infrastructure would take the venture market with it, because the venture market's enthusiasm for the entire category is downstream of the same narrative. Your next round gets harder regardless of your own numbers. That is the exposure to plan against — not your compute bill.
Which produces a short, unglamorous checklist:
- Know your runway without a next round. The correction scenario is a financing-market event for you, not an operations event. Default-alive is the hedge.
- Do not architect around one provider's pricing. Keep the model layer swappable. This is cheap to do early and expensive to retrofit, and it protects you in both the bust case and the price-hike case.
- Measure gross margin per customer with tokens in COGS. A surprising number of AI companies cannot answer what they earn on a customer after inference. In a sentiment correction that number is the first thing a board asks for.
- Sell the outcome, not the AI. Budgets allocated to "AI initiatives" are the first cut in a downturn. Budgets allocated to a function that works are not.
Where we are putting our effort
We build and operate companies rather than allocating capital, so this resolves into a concrete bias: we are building in the application layer and treating the infrastructure layer as a utility whose price we expect to fall.
Nothing in our portfolio requires owning compute. Nothing requires the $725 billion to keep compounding at 77%. The companies we are building need inference to be available and cheap, and essentially every resolution of the current debate — continued buildout or sharp correction — produces available and cheap inference on a two-year view. One path gets there through abundance and the other through distress, but they arrive at the same place.
That asymmetry is the whole thesis. If you are building something whose unit economics improve when compute gets cheaper, the bubble question is mostly somebody else's problem.
The companies with a genuine problem are the ones whose business model requires compute to stay expensive, or whose valuation requires the narrative to hold. Be honest about which you are.
