The chip makers supplying AI's infrastructure are printing money, but the math underneath is starting to look like a financial ouroboros eating its own tail.
The Summary
- ASML raised its 2026 sales forecast and plans to ramp chip-making machine production to meet AI data center demand, while Nvidia commits another $2B to European deals
- Fortune flags the divergence between AI capital expenditure and actual revenue as a signal we're approaching peak boom
- New capital flows into US AI infrastructure are creating circular deal structures that may introduce systemic financial risk
The Signal
ASML makes the machines that make the chips. When they raise guidance and promise to build more capacity, it means someone downstream is still spending. That someone is Nvidia, among others. Jensen Huang just committed $2 billion more to European infrastructure. The equipment layer is confident. The question is whether the confidence is justified or self-referential.
Fortune's analysis points to the numbers that should make anyone pause: AI capex is still climbing faster than AI revenue. Companies are spending billions to build the capability to run inference at scale, train larger models, and deploy agents across enterprise workflows. But the actual dollars coming back from customers who pay for AI services haven't caught up. This is the gap that defines late-stage booms.
"AI capex is climbing faster than AI revenue, the classic late-boom divergence."
It's not that the revenue isn't real. It is. OpenAI, Anthropic, and every enterprise software vendor bolting LLMs onto their products are seeing growth. But the rate of infrastructure investment implies a level of demand that doesn't exist yet, or at least hasn't been proven at the scale these capital commitments require. The equipment companies and chip designers are betting on a future that has to materialize, soon, or the math breaks.
Meanwhile, Bloomberg notes new financial structures emerging: circular deals where the capital funding AI infrastructure is itself tied to AI valuations or future revenue streams. Sovereign wealth funds, private equity, and tech giants are creating investment vehicles that lend to each other, invest in each other's suppliers, or buy capacity from data centers they partially own. The money moves fast, but the loops are tight.
This isn't inherently dangerous. Circular capital flows happen in any infrastructure boom, from railroads to telecom. But when the circle is small and everyone is betting on the same growth curve, a single weak link, a customer that doesn't materialize, a use case that underperforms, can propagate quickly. The risk isn't that AI isn't real. It's that the financial structures built around it assume linear adoption and revenue that matches the exponential capex.
Key numbers to watch:
- ASML's actual shipment volume vs. their raised forecast in Q4 2026
- Nvidia's Europe deal deployment timeline and who the anchor customers are
- Enterprise AI spend as a percentage of total IT budgets through 2027
The Implication
If you're building in this space, pay attention to where the money is actually being spent and by whom. Nvidia's European push suggests they see regulatory clarity and sovereign compute demand as real. ASML ramping production means the belief in multi-year GPU demand is still strong. But if Fortune's thesis holds, the window to prove product-market fit at revenue scale is closing faster than the build-out timeline suggests.
For operators: lock in infrastructure deals now while capacity is being added. For investors: start asking what happens if AI revenue growth slows to 30% annually instead of 100%. The answer to that question determines whether this is 1999 or 2009.