When compute became a receivable

I.The biggest quarter, and the thinnest cash

On 26 August 2026 Nvidia reported revenue of $96.2 billion for the quarter ended 26 July 2026, up 106% year over year, with data centre revenue of $89.0 billion and a 75.0% gross margin. Guidance for the current quarter is $108.0 billion. Then, on the call, the company did something it had never done before: it guided a full year ahead, telling investors to expect roughly 70% revenue growth in fiscal 2028 — an implied $690–700 billion against a sell-side consensus nearer $570 billion. CFO Colette Kress framed the number as "constrained by supply", not demand.

The line that matters, though, is further down the cash flow statement. Nvidia's own CFO commentary puts operating cash flow at $24.1 billion for the quarter. One quarter earlier, on $81.6 billion of revenue, it was $50.3 billion. Nvidia booked 18% more revenue and collected less than half the cash. Accounts receivable went from $40.7 billion in April to $63.1 billion in July; inventory went from $25.8 billion to $31.6 billion. Days sales outstanding — receivables divided by revenue, annualised — moved from about 45 days to 60. That is not a rounding difference. It is a change in what a sale is.

II.Two inversions arriving at once

The first inversion is financial. Nvidia is now extending working capital to its buyers. Kress attributed the 60-day DSO to extended terms for large investment-grade customers with shipments spanning multiple quarters, and reporting on the filing indicates terms running from 90 days out to a full year for some accounts, with five customers holding roughly 70% of the receivable balance. Separately, Nvidia disclosed financing platforms built with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR that are expected to raise more than $500 billion, plus take-or-pay commitments on facility capacity to help neocloud operators reach project finance.

The second inversion is physical, and less discussed. Nvidia's gross margin is going down — 75.0% in Q2, 74.0% guided for Q3, a trough of 71–72% guided for the January quarter, settling at 72–73% next year once price increases land. The cause is memory. Kress described "extreme pricing conditions"; TrendForce had already forecast on 9 July 2026 that server DRAM contract prices would rise 13–18% quarter over quarter in 3Q26, with RDIMM bit supply growing only 15–20% year over year against faster server growth, and with long-term agreements shielding some hyperscalers so that everyone else absorbs the increase. On 2 June 2026 the same firm projected HBM contract prices rising by multiples in 2027.

For fifteen years the working assumption in every AI business case has been that the cost of a unit of compute falls. In this cycle the binding constraint has moved from logic to memory, and memory is a commodity market in shortage. The cost of a frontier training or inference rack is rising, and Nvidia has chosen to eat roughly three points of margin before passing it on.

III.What 60 days does and does not prove

The honest read is that a single quarter of stretched receivables is not evidence of anything fraudulent, and the loudest version of the bear case overreaches. Nvidia has faced this argument before — when DSO hit 53 days in late fiscal 2026, comparisons to Lucent-era vendor financing circulated, and the company responded that its DSO sat near long-run averages and that it does not use special purpose entities. The strongest counter-argument is comparative: suppliers selling into the same cloud customers — Arista, Celestica, Vertiv — routinely run 60–70 day DSO. On that benchmark Nvidia has moved from unusually fast collection to merely normal collection.

What the quarter does establish is directional and it is not ambiguous. Receivables absorbed roughly $22 billion of cash in three months. Concentration rose rather than fell: five names now carry about 70% of the balance, against 56% a year earlier. And the terms are not incidental — Kress addressed the criticism head-on rather than denying the structure, saying the company knows some will call this circular financing and characterising the exposure as low-risk, high-reward. That is an argument about credit quality, not an argument that credit is not being extended.

IV.Who actually funds the next $700 billion

Work the arithmetic forward with stated assumptions. Fiscal 2027 lands near $406 billion if the $108 billion Q3 guide holds and Q4 comes in around $120 billion. Seventy per cent growth on that base is roughly $690 billion of fiscal 2028 revenue. Nvidia's customers must therefore find something on the order of $690 billion for Nvidia alone, before power, land, buildings, networking outside Nvidia's stack, and the memory premium they are also paying. Set against hyperscaler capital expenditure already running near $600 billion across Alphabet, Amazon, Meta and Microsoft in calendar 2026, this is not a marginal increment. It is a second buildout stacked on the first.

The four hyperscalers can largely fund that from operating cash flow and investment-grade debt, and they have been issuing heavily to do it. The other half of Nvidia's business cannot. Kress put the enterprise, neocloud and sovereign segment at about $40 billion in the quarter, growing 138% year over year, and Huang described non-hyperscale customers as roughly half the business growing at 100% annually. Those buyers — neoclouds especially — are thinly capitalised relative to the assets they are buying, which is precisely why Nvidia is arranging $500 billion of third-party financing and offering take-or-pay backstops. The fastest-growing half of the demand base is the half that needs someone else's balance sheet to buy.

V.The second-order effect: token prices may stop falling

Most enterprise AI business cases written in 2025 and 2026 assume the price of inference keeps collapsing, so that workloads uneconomic today become economic on a two-year view. Two things in this quarter cut against that. First, Nvidia has told the market it will raise prices on upcoming systems to recover memory costs. Second, memory shortage is a supply-side problem with multi-year lead times; TrendForce's own supply arithmetic has RDIMM bits growing 15–20% against faster demand, and HBM tightening into 2027.

If the delivered cost per unit of frontier compute is flat to rising through 2027, the effect is not that AI adoption stops — it is that the marginal workload does not clear. Batch-scale, low-value-per-token applications get deferred; high-value-per-task agentic work, where a dollar of inference replaces many dollars of labour, keeps clearing easily. That is a composition shift, and it is the mechanism by which a compute shortage becomes visible in the real economy: not as an AI slowdown, but as a narrowing of which AI use cases are worth running. For anyone modelling AI-driven margin expansion in a forecast period, the assumption to test is the cost-per-token glide path, not the adoption rate.

VI.The precedent, and where it breaks

The comparison bears reach for is telecoms in 1999–2001, when equipment vendors financed carriers to buy their own gear and the receivables turned out to be the earnings. The structural rhyme is real: vendor-extended terms, concentrated counterparties, capacity commitments made years ahead of the revenue they serve. The differences are also real, and this is a judgment rather than a fact. Nvidia's largest counterparties are among the most cash-generative companies in the world, not leveraged start-up carriers; the assets being bought are producing revenue today rather than lighting dark fibre for future demand; and Nvidia is funding commitments from a business generating tens of billions in quarterly profit, having still returned $26 billion to shareholders in the quarter. The exposure that resembles 2000 is not the hyperscaler book. It is the neocloud and sovereign tail, and that tail is growing fastest.

VII.How it plays out: three scenarios

Timeframes run to end-2027. Probabilities are analytical judgment, not market-implied, and sum to 100%.