The most revealing number in Anthropic's IPO prospectus is not the valuation it seeks, nor the revenue it reports. It is the $5 billion that AMD has pledged to spend buying Anthropic's own stock — disclosed in the same filing where Anthropic commits to purchase over $20 billion of AMD's Instinct MI450 compute. Contractors, in the ordinary course of business, render services for payment. They do not route that payment back into their client's equity. When they do, the transaction has a different name, and it is worth being precise about what that name is.
The Architecture of the Arrangement
Start with the timeline, because the sequencing matters more than any single disclosure. In July, AMD announced a "strategic partnership" with Anthropic, framed around securing a marquee customer for its AI accelerators. The language was conventional — deepening collaboration, frontier compute, mutual commitment. What the July announcement did not foreground was the direction of capital. AMD was not merely selling chips. It was committing up to $5 billion to purchase equity in the buyer.
The September IPO prospectus made the structure legible. Anthropic committed to acquire AMD compute capacity valued in excess of $20 billion, covering up to two gigawatts of Instinct MI450 deployments beginning in the first half of 2027. In the same document, AMD's reciprocal commitment appears: up to $5 billion in Anthropic stock purchases, "subject to certain contingencies," expected to be deployed through fiscal year 2028.
AMD's own quarterly filing confirms the figure — "investment commitments of up to $5.0 billion, subject to certain contingencies, which are expected to be made through fiscal year 2028" — while conspicuously omitting Anthropic by name. The precision of the match is not coincidental. A supplier does not commit exactly $5 billion to an unnamed counterparty in the same quarter it announces a $5 billion partnership with a named one.
Strip away the vocabulary and the mechanics resemble a rebate executed through share buybacks. Anthropic pays AMD for compute; AMD returns a portion of that consideration by purchasing Anthropic equity, which supports the valuation at which Anthropic sells its shares to the public. The cash flows in a circle, and the circle closes just in time for the IPO.
Why the Circle Exists
The Wall Street Journal reported that hyperscalers are capping Anthropic's access to AI compute. This is the constraint that explains everything else. Compute access is not a procurement detail for a frontier lab; it is the binding input. Without capacity, there is no training run, no inference margin, no product. Anthropic co-founder Tom Brown said as much in July: "Access to compute is central to keeping Claude at the frontier and meeting demand from our customers." That sentence is the honest summary of the company's strategic position.
But compute is expensive, and Anthropic's financials are not those of a company that can fund two gigawatts of MI450 deployment from operating cash flow. In 2025 the company ran an operating loss above $8 billion on just $4.6 billion of revenue. The loss-to-revenue ratio is the kind of figure that invites a second look at any valuation premised on future profitability. A company burning nearly two dollars for every dollar earned needs capital, and it needs that capital to arrive without repricing its equity downward.
This is where the circular structure earns its keep. A conventional vendor financing arrangement — AMD lending Anthropic the money to buy AMD chips — would appear on the balance sheet as debt and invite scrutiny of leverage. A rebate would reduce reported revenue for AMD and require disclosure as a pricing concession. The equity commitment does neither. AMD books a customer commitment and an investment; Anthropic books compute access and a supportive shareholder. Both companies report what looks like growth. The circularity is distributed across two sets of financial statements, and neither statement, read alone, reveals it.
The Precedent Nobody Wants to Name
The pattern is not new. During the 2021 cycle I spent three weeks auditing the compound interest mechanics of Aave and MakerDAO, watching billions in total value locked flow into yield farms that manufactured returns through token emissions rather than economic activity. The instruments were different; the architecture was identical. Value was created by moving it in a circle and booking each pass as growth. When the emissions stopped, the TVL evaporated, and the protocols that survived were the ones with real cash flow underneath the incentives.
The AMD-Anthropic arrangement is the institutional-grade version of that architecture. The participants are serious, the technology is real, and the chips will physically ship. But the demand signal is contaminated. When a supplier funds its customer's purchase, the resulting revenue measures the supplier's willingness to invest, not the customer's willingness to pay. Anthropic may genuinely need two gigawatts of MI450 capacity. The question is whether it would commit $20 billion for that capacity if AMD were not simultaneously committing $5 billion to buy its stock. The filing does not answer that question, and the structure is designed so that it does not have to.
Nvidia's arrangements with its own customers have drawn similar scrutiny — the practice of investing in AI startups that then purchase the investor's hardware is now a recognizable feature of the compute economy. What distinguishes the AMD-Anthropic deal is the synchronization: the investment commitment and the purchase commitment appear in the same prospectus, at matching scale, timed to an IPO. That is not an industry pattern. That is a transaction.
The Contrarian Reading
The charitable interpretation deserves a fair hearing, because the cynical one is too easy. AMD is competing against a dominant incumbent for the loyalty of the most important AI labs. Locking in Anthropic as a design partner for MI450 — a chip architecture that needs real-world deployment to mature its software stack — has strategic value that exceeds the nominal size of the equity commitment. If Anthropic's workloads stress the hardware in ways that improve ROCm, AMD's entire accelerator business benefits. The $5 billion is, on this reading, an R&D subsidy dressed as an investment.
There is also a legitimate argument that equity commitments align incentives in ways that purchase contracts cannot. A supplier with an ownership stake has reason to prioritize its customer's success. Vertical alignment of this kind is common in semiconductor history.
But alignment arguments do not survive contact with the disclosure structure. If the arrangement were a straightforward strategic investment plus a straightforward supply agreement, there would be no reason to omit Anthropic's name from the quarterly filing that discloses the $5 billion commitment. There would be no reason for the July announcement to emphasize partnership while the September prospectus reveals the reciprocity. And there would be no reason for the two figures to match so precisely. Liquidity is a mirage; only settlement is real — and the settlement here is that AMD's investment supports the valuation at which Anthropic's IPO will clear, which means public market investors are buying into a company whose largest customer commitment is partly funded by the vendor itself.
That is the blind spot in the bull case for AI infrastructure. The sector's revenue growth is being measured at the point of sale, but an increasing share of those sales are financed by the seller. Each circular deal makes the aggregate demand picture look stronger than the underlying end-market demand. When the AI capex cycle eventually cools, the unwinding will not be visible in any single company's filings. It will appear as a sudden, unexplained deterioration across the sector, because the financing that sustained the revenue was distributed across counterparties who all reported growth simultaneously.
What to Watch
The two-gigawatt MI450 deployment begins in the first half of 2027, which means the arrangement has roughly eighteen months to prove whether the demand is real or manufactured. Three signals will settle the question. First, whether Anthropic's post-IPO filings disclose the AMD equity commitment as a related-party transaction or bury it in investment income. Second, whether AMD's investment commitment is drawn down on schedule or quietly renegotiated as contingencies fail. Third, whether Anthropic's compute spending continues at the committed pace once the IPO capital is raised and the valuation is locked.
There is a fourth signal, quieter than the others. Watch whether other AI labs announce similar structures in the coming quarters. If the chips-for-shares model spreads, it will mean the compute economy has discovered that circular financing is the only way to sustain the growth expectations embedded in current valuations. If it does not spread, it will mean AMD and Anthropic found a bespoke solution to a bespoke problem — a company that needed compute it could not afford, and a vendor that needed a customer it could not otherwise win.
Either way, the prospectus has already told us something the press release did not. In the AI infrastructure buildout, the distinction between a customer and an investment is collapsing. The companies selling compute are increasingly the companies financing its purchase, and the resulting revenue figures should be read with that circularity in mind. Hype is a liability when the ledger is circular. The chips will ship. The shares will settle. What remains uncertain is whether the demand they represent was ever independent of the capital that made it possible.
