The Trillion-Dollar Circle

A private valuation is not a bank balance. It is not a public share price, a signed purchase order or a pile of cash waiting to be withdrawn. It is an estimate attached to a transaction in which a limited number of shares change hands. That distinction becomes urgent when the companies buying artificial-intelligence infrastructure are also investors in, cloud partners of, landlords to, or suppliers for one another.

The concern is not that every partnership is fraudulent or that every large valuation must collapse. The concern is concentration: if a small number of firms finance the customers whose spending supports their own revenues, the same growth story can appear in several places in the accounts. It may be genuine demand, but it is not independent evidence repeated many times.

What a valuation says

When a private company raises capital, the new share price can be multiplied by all its shares to produce a post-money valuation. The calculation is useful. It gives investors and employees a reference point for pricing ownership. But the new round may involve only a fraction of the company. The resulting figure is not an appraisal of every share in a saleable market, and the company does not receive the headline valuation as cash.

Anthropic’s February 2026 announcement provides a concrete distinction. The company said it had raised US$30 billion in a Series G round at a US$380 billion post-money valuation. Anthropic’s announcement documents both the financing and the valuation. The company received the new capital, subject to the round’s terms; US$380 billion was the implied value of the company after that round, not the amount invested.

Later reports of a higher valuation, or conversations about a future fundraise, are different evidence. A signed financing round has investors and documents. A report that a company is “in talks” describes a possibility, not cash received. An announced commitment may still be conditional. These labels matter because a valuation can rise before the company’s revenue, cash flow or delivered products have caught up.

The same care applies to spending commitments. A cloud contract, a lease for future data-centre capacity, a chip order and a guarantee are not interchangeable. A contract can contain timing and volume conditions. A lease can cover a long future period. A guarantee creates contingent exposure if another party fails to pay. None should be added to a funding round as if every dollar were already spent, delivered and earned.

How the loop could form

The mechanism is straightforward. An investor puts equity into an AI developer. The developer agrees to buy computing capacity from a cloud provider. That provider finances or builds data centres and buys equipment from chip and power suppliers. Those suppliers invest in the developer, sign partnerships with the cloud provider, or make financing and capacity available to the same ecosystem. If the developer succeeds, the sequence may be commercially sound: capital builds a product, customers pay for useful work, and suppliers earn revenue.

The loop becomes a fragility when each company’s apparent demand depends on another company’s financing rather than on independent end customers. One dollar can support several narratives: a large investment demonstrates confidence in the developer; the developer’s cloud commitment supports the provider’s growth plan; the provider’s equipment orders support the supplier’s revenue outlook. The dollar is real, but the stories it is used to support are not separate confirmations of final demand.

The links between investment and supply are not hypothetical, but they need careful description. The Federal Trade Commission’s staff study of major AI partnerships documented arrangements between leading AI developers and cloud providers, including Microsoft–OpenAI, Amazon–Anthropic and Google–Anthropic. It examined access to compute, switching costs, exclusivity and information rights. The FTC report documents commitments requiring developers to spend a large share of their cloud partner’s investment on that partner’s services. This links financing to purchases directly; it does not establish fictitious revenue or unlawful conduct.

An investor can rationally finance a customer whose purchases increase demand for the investor’s products. The test is what happens after the investment. Are systems delivered? Does the customer use them? Does revenue arrive in cash on time? Can the buyer pay for the capacity without another capital raise? If the project is delayed, who bears the cost of idle equipment and power capacity? The answers distinguish productive financing from a chain that depends on the next round to sustain the last one.

Real revenue, contingent promises

The commercial scale is not imaginary. Nvidia’s filing for its second quarter of fiscal 2027 reported revenue of US$96.2 billion, including US$89.0 billion in Data Center revenue, up 117 per cent year over year. Those are company-reported financial results in its SEC filing. Revenue from delivered products is stronger evidence than a valuation rumour. Still, one supplier’s revenue does not tell us how concentrated the buyers are, how much credit financed their purchases, or whether end users will earn enough from the equipment over its life.

The difference between sales and promises is central. A reported order may be cancellable or extend over years. A data-centre lease may not produce capacity immediately. A supplier’s “purchase obligations” can include future inventory, components and manufacturing capacity; they are not all current-period costs or guaranteed sales. A guarantee can protect a lender or landlord without transferring cash unless a borrower defaults. The SEC filing should be read for its definitions and contingencies, not mined for one large number and treated as a demand forecast.

There is a real countercase to the circle thesis. Capital-intensive industries often rely on long-term contracts, anchor customers and supplier finance. Aircraft, telecom networks and power plants are not built by spot orders alone. Long commitments can lower financing costs and make a new technology available sooner. If customers can use the capacity and pay for it, overlapping investments and contracts may reflect coordination that creates value rather than a bubble.

The other countercase is that AI demand has many sources. Public companies, governments, research institutions and small firms are testing products. Computing can be useful even when some expensive projects fail. A few dominant suppliers may reflect economies of scale and the difficulty of manufacturing advanced accelerators, not a secret agreement to inflate one another’s books. The existence of concentration is a reason to inspect terms and dependencies, not a verdict on motive.

What the regulator’s concern is

The Federal Trade Commission’s 2025 staff report on AI partnerships and investments focused on deals involving major AI developers and cloud providers. It described questions about access to compute and data, switching costs, exclusivity and how partnerships might affect competition. The report is a staff study of selected agreements, not a finding that every partnership is illegal or that financial statements are false.

Competition and financial stability are related but distinct questions. A cloud partnership may limit a startup’s ability to move workloads without creating immediate credit risk. A supplier investment may support a new product without making it anticompetitive. But if a few firms control the chips, data centres, cloud platforms and access to customers, a shock to one can travel through the network. A delayed data centre can weaken a cloud provider’s growth plan; that can affect equipment orders, power contracts and the valuation of firms relying on the same buildout.

The risk is not necessarily a sudden crash. It can appear as lower returns, renegotiated contracts, stranded facilities, higher financing costs or a slower rollout. Those costs may be distributed across shareholders, lenders, workers, customers, electricity users and local communities. Public attention often focuses on the investor’s loss; infrastructure built for an optimistic demand forecast can also leave utilities or municipalities facing unused capacity and unrecovered costs.

A checklist for reading the next announcement

Start with the transaction type. Is it a completed equity round, a stated intention, a purchase order, an agreement to negotiate, a lease, or a guarantee? Who pays whom, when, and under what conditions? Is the announced value gross of future obligations or cash received? Does the company report revenue recognized on delivery, or a booked commitment for future years?

Then ask whether the buyer has independent customers. A cloud provider’s sales to an AI developer matter, but so do the developer’s revenue from outside customers and the cash it collects from them. If a project is funded primarily by investors who are also its suppliers, the reader should not count each related announcement as a separate vote of market demand. Look for customer concentration, receivables, capital spending, cancellation rights and financing costs in filings.

Finally, identify the physical constraints. Compute requires chips, buildings, electricity, cooling, network links and skilled maintenance. The timing of those pieces rarely aligns perfectly. A signed capacity plan is not the same as a powered facility. A delivered accelerator is not the same as a useful workload. A useful workload is not automatically a profitable service.

There is no single “trillion-dollar circle” that proves the sector is either sound or doomed. There are transactions with different legal forms, timelines and risks. Some capital will build productive capacity; some projects will fail to find customers at the price needed to justify their cost. The work is to follow the cash, the commitments and the customer base separately. A valuation is a signal. It is not the same thing as a completed sale, an enduring business or a guarantee that the next link in the chain will pay.

An NVIDIA GPU photographed 10 February 2023. Historical hardware image illustrating the chip supply chain; it is not a photograph of the accelerator models or data-centre contracts discussed in the article. Credit: Mickael Courtiade, via Wikimedia Commons. Image source · CC BY 2.0 Generic