Nvidia wrote down H200 chips it could not sell while one customer drove 16% of revenue — the cost of a $279 billion supply bet

Nvidia's fiscal second quarter, ended July 26, 2026, was as strong as the headline numbers suggested: revenue of $96.2 billion, up 106% from a year ago; data center revenue of $89.0 billion, up 117%; and guidance of $108 billion for the October quarter. But the quarterly report filed the same day carried two disclosures the growth smoothed over. Nvidia took a $0.4 billion charge during the first half of fiscal 2027 for excess H200 inventory and purchase obligations — its own words, "as the demand for H200 products diminished" — and it disclosed that one direct customer alone represented 16% of the quarter's revenue. Neither figure moves the arithmetic of the quarter. Both matter to what the stock is priced for.

The charge

To see why Nvidia held chips it could not sell, follow the H200 through one year of export policy. The H200 is Nvidia's Hopper-generation flagship, introduced in late 2023 — the architecture that Blackwell and then Vera Rubin, now in full production, superseded. Outside of China, H200 barely appears in the demand story anymore. Inside China, it became the whole story.

In early December 2025, President Trump announced the United States would allow H200 exports to China while collecting a 25% fee on sales — a reversal of the controls that had all but closed the market to Nvidia. Chinese buyers reportedly ordered more than two million chips, and Nvidia pushed TSMC to add production. Then the ground shifted. Beijing restricted the chip's entry, and in January 2026 the formal U.S. green light for H200 exports arrived with new conditions, including third-party review and a tariff Nvidia says it cannot pass through to customers. By May, the Commerce Department had approved about ten Chinese firms to buy H200s — each allowed up to 75,000 chips, a combined potential near 1.5 million — but actual deliveries stayed minimal.

The gap between that promise and reality is the charge. Analysts put roughly 500,000 H200 chips in Nvidia's inventory, the bulk earmarked for China, while the first cleared sales are a trickle — reportedly 10,000 chips each for Tencent and ByteDance. The 10-Q states the cost precisely: a $0.4 billion charge over the first half of fiscal 2027, with the licensed shipments Nvidia has actually made worth less than 1% of data center revenue.

This is the second time the mechanism has produced a charge in two years. The first was H20, the cut-down chip Nvidia built specifically for the Chinese market: when Washington required an export license for H20 in April 2025, Nvidia initially warned of a charge of up to $5.5 billion, and the actual charge in the first quarter of fiscal 2026 came to $4.5 billion for H20 excess inventory and purchase obligations. H200 is the repeat — a market opened, then closed, by governments in the middle of a product generation, with the inventory Nvidia had built stripped of value.

The bet underneath

By scale, the H200 charge is noise: about 0.2% of the $177.8 billion Nvidia recorded in the first half of fiscal 2027, against a 75.0% gross margin for the half (the year-earlier half, scarred by the H20 charge, printed 66.6%). Inventory provisions clipped roughly a percentage point off gross margin this time. The charge matters less for what it cost than for what it exposes about the machinery producing the growth.

Nvidia does not build to order; it pre-commits supply years ahead of sales. Inventory at the end of July stood at $31.6 billion, up roughly 48% in six months, and the company's supply-and-capacity commitments more than doubled in one quarter — from $119 billion to $279 billion as of July 26, 2026 — according to its own report. That is a $279 billion position in chips, systems and capacity that Nvidia expects to sell in later quarters, and the H200 write-down is the small, visible case of what a commitment becomes when the demand estimate behind it is wrong, whether because a product generation turned over or a licensing regime did.

The 16% customer

The second disclosure is the mirror image, on the demand side. One direct customer represented 16% of total revenue in the July quarter — a single buyer worth roughly $15 billion of revenue in one quarter. A year earlier, the same disclosure showed two direct customers at 23% and 16% of revenue; for the first half of fiscal 2027, three direct customers represented 16%, 15%, and 13%.

"Direct" is doing quiet work here. Nvidia's direct customers are the firms that sign its invoices — cloud providers, AI model makers, server builders, distributors — as opposed to the end users who buy through those channels. That gap is where the real dependence hides. Nvidia itself states that one AI research and deployment company contributed "a meaningful amount" of revenue by buying cloud services from Nvidia's customers. In plain terms: a frontier AI lab that never appears in the direct-customer disclosure still shows up inside Nvidia's revenue. The disclosed 16% is a floor on concentration, not a ceiling.

The same structure, read from two sides

Put the two disclosures together and they are one structural fact. Nvidia's growth is built on enormous forward commitments of supply, matched to a demand funnel that narrows to a few enormous capital programs. The H200 charge is the supply side failing in one channel; the 16% customer is the demand side concentrating. Neither changed what happened in the July quarter — net income was $59.7 billion on a 75.0% gross margin. Together they describe where a wobble would come from: a single hyperscaler pausing its build is a swing of many billions of dollars, and Nvidia's revenue is the downstream echo of that spending. The four biggest hyperscalers spent $166 billion on capital equipment in the June quarter, up 87% from a year ago — Nvidia's quarter was, to a large degree, already determined by that number before it reported.

The market has noticed the tradeoff in its own way: after years of a rich multiple, Nvidia's forward earnings multiple has compressed into the mid-20s even as revenue sets records, a market capitalization near $5 trillion backing it. Paying a lower multiple per dollar of earnings is another way of saying the market trusts the compounding less than it once did — and the fine print of the 10-Q is precisely where its doubts would find evidence.

What to watch is specific. Management's own guide for the October quarter assumes no data center compute revenue from China at all — the $108 billion target excludes the very chips it just wrote down. Watch whether Chinese data center revenue returns, whether inventory keeps climbing against the $279 billion of commitments, what the customer-concentration section says in each 10-Q, and whether gross margin holds near 75% as the mix shifts from one generation to the next. The H200 charge was barely a tenth the size of the H20 charge a year earlier, while the supply bet behind it more than doubled. That ratio — smaller charges against a larger commitment — is the quiet signal in the same document, and the one worth watching.