How much of current AI infrastructure spending reflects independent end-user demand versus interlocking investments among a small group of chipmakers, cloud providers, and model labs, and what would a demand shock actually look like?

Advocate A’s weakest move is treating every intercompany commitment as if it were already a reciprocal purchase order booked into Nvidia revenue. That is the crux of the telecom-vendor-financing analogy, and it is not yet established here. Nvidia’s Q1 FY2027 10-Q says existing partner facility lease guarantees have a maximum gross exposure of $3.5 billion, were received in exchange for warrants, and are reduced as partners pay lessors; the same filing separately reports deferred revenue and customer advances, so guarantees, equity investments, and revenue are not the same accounting object. A reported $250 billion OpenAI backstop after quarter-end would be economically important and would require scrutiny, but the $3.5 billion disclosure is not a “cap” proving hidden exposure; it is the exposure under agreements then in force. A’s “same dollar recorded as investment, revenue, and backlog” claim needs transaction-level evidence that Nvidia financed a customer’s chip purchase and immediately recognized the financed sale. The sources cited show talks and concerns, not that accounting result.

The better economic frame is scarce capacity underwritten by large, observable demand signals. OpenAI’s own CFO states that available compute rose from 0.2 GW in 2023 to about 1.9 GW in 2025 while ARR rose from $2 billion to more than $20 billion, with consumer subscriptions, workplace subscriptions, and usage-based APIs tied to production workloads. That is company-reported and should be discounted for self-interest, but it is a falsifiable demand trajectory, not a shell customer with no retail revenue. Microsoft independently reported Microsoft Cloud revenue of $54.5 billion, up 29%, said calendar-2026 capex would be roughly $190 billion, and still expected to remain capacity constrained at least through 2026. A telecom CLEC analogy is weakest when the buyers are Microsoft, Google, Amazon, Meta, Oracle, and OpenAI-scale platforms rather than thinly capitalized carriers dependent on speculative local-loop revenues.

The SSI deal also does not prove circular revenue. Reuters reports a $5 billion equity investment and hardware access to Vera Rubin; equity risk in a pre-product research lab can be criticized as expensive strategic optionality, but absent a disclosed reciprocal purchase obligation it is not the same as sham customer financing. The proper concern is concentration and counterparty risk, not accounting circularity.

Finally, the July 28 selloff is weak evidence of a current demand shock. AP attributes the KOSPI halt and Samsung/SK Hynix declines to sustainability fears and Chinese competition after CXMT’s roughly 466% debut, not to cancelled cloud orders, rising GPU inventories, falling utilization, or price cuts. A real demand shock would look like hyperscalers reducing capex guidance, cloud AI utilization falling, Nvidia deferred revenue/customer advances reversing, or model labs cutting committed capacity—not merely semiconductor equities repricing after an extreme run-up.

A’s rebuttal concedes the decisive accounting point: there is no transaction-level evidence that Nvidia has already financed chip purchases and booked those financed purchases as revenue. That concession narrows A’s case from “circular capital” to a legitimate but different claim: AI infrastructure has counterparty concentration, duration mismatch, and execution risk. Those risks matter, but they do not establish the telecom-vendor-financing mechanism unless the supplier’s capital is shown to be creating revenue that would not otherwise exist. A’s new evidence still does not close that gap.

The CoreWeave contract is A’s strongest remaining fact, and it should be taken seriously. CoreWeave’s SEC filing says Nvidia’s $6.3 billion order gives Nvidia access to residual unsold cloud capacity through April 2032 if CoreWeave’s own customers do not fully utilize capacity. But that is a cloud-capacity backstop, not a disclosed commitment to fund CoreWeave’s GPU purchases and book a reciprocal GPU sale. It is also bounded: $6.3 billion over roughly seven years, compared with Nvidia’s $81.6 billion of total revenue and $75.2 billion of data-center revenue in Q1 FY2027 alone. Nvidia’s 10-Q separately discloses $30 billion of multi-year cloud-service commitments “primarily” for R&D, customer concentration, liquidity, and revenue categories. That is not proof of sham demand; it is evidence the market can audit the risk.

A’s “demand shock” evidence is also weaker than claimed. Reuters reported that Oracle, CoreWeave and other AI-linked stocks fell after a WSJ report that OpenAI missed internal user and revenue goals and that CFO Sarah Friar worried about future computing-contract affordability if growth did not accelerate. That is evidence investors rapidly reprice OpenAI-linked duration risk. It is not evidence that Azure or OpenAI cancelled capacity, that Nvidia’s customer advances reversed, that GPU utilization collapsed, or that hyperscaler capex guidance was cut. Market beta is not the same as physical demand destruction.

The remaining evidence still favors the scarcity explanation. OpenAI reports that available compute rose from 0.2 GW in 2023 to about 1.9 GW in 2025 while ARR rose from $2 billion to more than $20 billion, and says usage-based APIs and workplace subscriptions tie spend to production use. Microsoft independently reported $54.5 billion of Microsoft Cloud revenue, Azure growth of 39% constant currency, $190 billion of 2026 capex, and capacity constraints through at least 2026. The July 28 KOSPI halt and CXMT debut fit a valuation and competitive-supply shock: AP attributed the selloff to AI sustainability fears and Chinese competition, while CXMT’s revenue reportedly surged on AI demand. A has shown concentration risk; they have not shown that current capex is primarily manufactured rather than consumed.

A’s strongest case is that economic circularity can exist even when GAAP categories are formally distinct: Nvidia can buy Oracle capacity, Oracle can buy Nvidia GPUs, and OpenAI can sign long-dated cloud contracts, causing backlog and strategic investment to reinforce one another before end-user revenue fully catches up. A is also right that Oracle’s $638 billion RPO, only about 12% recognized over the next twelve months, creates real duration, counterparty, financing, and technology-obsolescence risk if AI demand disappoints.

My position is still stronger because A has moved from evidence of risk to proof of circular financing without showing the necessary causal link. Oracle’s RPO schedule is not, by itself, evidence of fake demand; Oracle’s 10-K defines it as contracted revenue to be recognized over future periods and lays out a multi-year recognition pattern. A long recognition tail is exactly what one would expect from multi-year data-center and cloud-capacity contracts, not uniquely from round-tripping. Likewise, Oracle’s official disclosure that Q2 RPO growth included commitments from Meta, NVIDIA, and others proves Nvidia is an Oracle customer, but it does not prove Nvidia’s capital is funding OpenAI’s or Oracle’s GPU purchases in order to create Nvidia revenue. A reciprocal commercial relationship among hyperscalers, chipmakers, and model labs is not automatically sham vendor financing unless the supplier is underwriting the buyer’s purchase and recognizing the sale without independent demand.

The independent-demand evidence remains more concrete than A allows. OpenAI reports ARR rising from $2 billion in 2023 to more than $20 billion in 2025 while available compute rose from 0.2 GW to about 1.9 GW, and describes monetization across subscriptions, workplace products, and usage-based APIs. That is not proof the next trillion dollars of commitments will earn adequate returns, but it is direct evidence that compute additions have coincided with end-user monetization. Microsoft independently reported Azure growth around 39% constant currency, roughly $190 billion of calendar-2026 capex, and capacity constraints through at least 2026; that is hard to reconcile with a claim that demand is primarily manufactured by Nvidia’s financing rather than constrained by supply.

A’s Sequoia-style “revenue gap” is useful as a valuation stress test, not as forensic evidence of circular accounting. It assumes a required payback structure and margin stack; it does not measure actual GPU utilization, customer churn, order cancellations, or whether workloads are warehoused. The strongest conclusion the evidence supports is concentration and execution risk in AI infrastructure. It does not yet support the telecom-collapse analogy, where vendor loans directly enabled insolvent customers to buy the vendor’s equipment and inflate supplier revenue.