The Loop That Eats Itself

On October 8, two institutions not given to panic—the Bank of England and the IMF—issued matching warnings: AI-inflated valuations are approaching dot-com altitude, and risk of correction has increased. When narrative outruns cash flow, gravity resumes command.

Call it “strategic partnerships.” Better: call it circular finance.

What to do now.

1. Require plain-English disclosure of reciprocal deals—equity, prepayments, and capacity commitments among the same few firms.

2. Publish concentration maps showing which chipmakers, clouds, and labs rely on counterparties they themselves finance.

3. Stress-test vendors whose forward revenue depends on customer equity they underwrote. Sunlight first, then discipline.

Here is the emblem. Reports describe a letter of intent in which Nvidia commits up to $100 billion to OpenAI for data-center buildouts designed—by specification—to run on Nvidia systems. The money enters OpenAI as equity or committed funding. It leaves as purchase orders for Nvidia hardware and multi-year service contracts. Nvidia recognizes revenue as systems ship and support accrues. OpenAI records expense and long-dated obligations. The circle flatters both sides in the short run, but the costs that follow—power, operations, migration risk—do not vanish, and the supplier’s dependence on a single buyer deepens. Supplier funds buyer. Buyer redeems supplier. The loop tightens.

Arithmetic is undefeated.

Defenders answer that this time is industrial, not financial. They point—rightly—to real productivity gains: code assistants that raise throughput and cut defects; customer-service copilots that reduce handle time; back-office agents that digest documents in minutes, not days. These are tangible wins. The question is scale and timing. Do today’s earnings from these uses service the capex being pre-booked against tomorrow’s hopes? Productivity anecdotes do not pay power bills.

History offers one clean rhyme. In 1999–2001, telecom equipment vendors financed carriers to lay fiber the carriers could not carry on their own balance sheets. The network was built; the revenues were not. When funding snapped, vendor finance turned from lubricant to accelerant, and the unwind was brutal. The lesson: when suppliers manufacture demand with their own money, overcapacity is not a risk—it is an inevitability.

Call it what it is: leverage—of expectations, of capex, of credibility. If vendors must keep writing checks so customers can keep buying, belief becomes a condition of solvency. Miss a revenue ramp. Slip a model cadence. Meet a rival who ships comparable capability at lower compute intensity. Each step reprices not one firm but a chain—chips, clouds, land, and power—lashed to the same story.

What happens if belief blinks? Three paths, each costly.

Path one: consolidation by patron. A deep-pocketed platform absorbs its champion lab. Continuity is preserved and service persists, but independence is gone and competition is dulled. The center of gravity shifts to the acquirer’s cloud and roadmap.

Path two: negotiated slowdown. Timelines stretch; “phased activation” becomes policy; and “optimization” is the lingua franca of delay. Picture a 300-megawatt campus built for year-one occupancy that idles at a 40% load factor while contracts are amended. Racks are installed, but half remain dark. Power draw is capped below interconnect capacity. Minimum-spend floors are quietly converted into rolling credits. Press releases celebrate efficiency and carbon cuts. The asset is stranded; the return slides to the right.

Path three: a snap. Credit tightens, new money vanishes, and orders are canceled or deferred at scale. Chip stocks gap down; AI-optimized clouds scramble to backfill a marquee tenant; downstream startups renegotiate or shutter. Not apocalypse—regime change.

Investors should stop grading press releases and start grading cash conversion. Three tests suffice. Contract reality: Are enterprise deals paying their own way without vendor-financed coupons? Self-financing: Can platform capex be carried by platform gross margin, not by suppliers’ checkbooks? Efficiency: Are gains driven by more capability per watt, or merely by more watts per model?

The bull case deserves its due. If operating income from real deployments can shoulder the buildout, circularity resolves into ordinary commerce and the system hardens. But belief is not a business model, and vendor finance is not free.

Unchecked circularity breeds fragility. Deterrence—in finance as in statecraft—works best before the test.