AI, Compute & Infra: Is There a Bubble?

Executive summary — free edition

Finnacl Research · September 2026 · The full report covers 12 core companies in depth, 25+ more across the supply chain and power layer, eleven ETFs, and a quarterly-scored early-warning scorecard.

The answer

Not one bubble — three different markets wearing one name.

1. The cash-flow core is expensive, not bubbly. The companies doing the building and the companies selling to them are earning real money at a pace with no precedent: Nvidia's data-center revenue is up 117% year-on-year at a 75% gross margin; Microsoft's AI business runs at $37B and is growing 123%; Amazon Web Services is growing at its fastest rate in 18 quarters. And the stock market is not paying bubble prices for it — Nvidia trades at about 29 times the earnings it has actually reported; Cisco, measured the same way, peaked at 218 times in March 2000.
2. The financing edge is where bubble mechanics are forming. Nvidia has put more than $40 billion of its own money into its customers this year. Chip buyers are being paid in supplier stock. The largest independent GPU cloud carries $35.6 billion of debt against $6.9 billion of cash and borrows five-year money against three-year contracts, at interest spreads that widened every time it came back to market in 2026. Two of the five biggest buyers now spend more on capacity than their operations generate in cash. The Bank for International Settlements warns of "the same asset being pledged multiple times." This is the layer where the 2000 telecom bust lived — and the telecom analogy indicts the edge, not the core.
3. The narrative perimeter is already frothy. A public "AI ETF" whose fourth-largest holding is a private company; a chip IPO at 73 times sales that doubled on day one. We name it, and we don't cover it.

The constraint has a clearer answer than the bubble

The scarce thing is no longer chips. It is electricity — getting a large new load connected to the grid takes three to seven years in the main US hubs, heavy gas turbines are quoted five to seven years out, and one regional capacity market has cleared at its legal price cap three years running. Chip constraints (advanced packaging, high-bandwidth memory) are real but cyclical — capacity investment relieves them on a two-to-three-year cycle. Power is a decade-scale constraint, and that is where the theme's duration lives.

The three numbers that matter

What would change our mind

The full report scores eight measurable early-warning markers every quarter — capex affordability, the revenue gap, chip-rental prices, supplier backlogs, the power gap, lender behavior, depreciation accounting, and breadth of demand — each with the number printed and a stated threshold. Today's reading: one green, one red (the power constraint, which supports the thesis), six amber — including chip-rental prices, which are firming rather than falling. We publish the scorecard so readers can check us, not trust us.

What's in the full report

The value-chain map (who keeps what of each dollar) · four constraint chapters with the numbers · 12 core-company profiles with what each price assumes · the supply-chain, power, and landlord layers · the ETF purity analysis (a "diversified" trio of popular funds is 47 percentage points of a single stock) · exposure routes by scenario · the client Q&A · full sourcing on every figure.

Disclosures. Impersonal research, identical for every purchaser and tailored to no one; not individualised advice, and Finnacl Research is not an investment adviser (publisher's exclusion, Advisers Act §202(a)(11)(D)). The tiers differ in seats and redistribution rights, not in content. No statement about what any reader should hold, no recommended allocation, no forecast. The author holds no position in any company covered in this report, nor in any security named in its fund tables. No issuer payment, no advertising, no trading against published views, and no trade in any security named here in the thirty days before or after publication. Full disclosures in §13 of the report.