Nvidia's China problem

Nvidia reported revenue of $46.7 billion for the second quarter of fiscal 2026, another very strong result. The awkward part was the outlook. Its guidance left out any revenue from China.

Data-centre sales made up $41.1 billion of the total. The company forecast third-quarter revenue of $54 billion, plus or minus 2 per cent, on the assumption that it ships no H20 chips to China, a forecast that still implies sequential growth of more than 15 per cent. It sold none there in the second quarter either, despite this month's licensing deal under which it would pay the US Treasury 15 per cent of such sales. The board added $60 billion to its share buyback authorisation, a measure of how much cash the rest of the business generates.

Three years of shrinking access

The H20 was itself a product of export controls. In October 2022 Washington barred sales of Nvidia's top data-centre chips to China; the company responded with slowed-down versions, which a second round of rules in October 2023 also caught. The H20 was designed to fit under those thresholds, and for a while it sold well to Chinese cloud and internet groups.

In April this year the administration told Nvidia that the H20, too, would need a licence. The company took a $4.5 billion charge on inventory and purchase commitments in the quarter to April, and warned that the rule would cost it about $8 billion of second-quarter sales. The summer's reversal, which promised licences in exchange for a cut of the revenue, has so far produced no shipments at all.

China taken out of the numbers

Faced with shifting export rules, licensing delays and a revenue-sharing requirement, Nvidia's management removed Chinese data-centre demand from its baseline forecasts. Analysts regard this as an unusual step for a company to take. China has in the past accounted for over twenty per cent of Nvidia's data-centre revenue worldwide. Leaving it out of forward estimates concedes that geopolitical division has shrunk the market Nvidia can dominate, probably for good.

The obstacles sit on both sides of the Pacific. Washington's licences came with a price attached and no formal rule behind them, while China's internet regulator summoned the company in July over alleged security risks in the H20. The finance chief, Colette Kress, said H20 shipments of $2 billion to $5 billion were possible in the third quarter if those issues were resolved. Excluding them from guidance turns China from a forecast into an option, which flatters the chance of a beat but hides how much of the old franchise has gone. Every quarter of absence also gives Huawei and other Chinese chipmakers time to win customers that would once have defaulted to Nvidia.

Beijing has reasons to keep the door half shut. January's release of DeepSeek's R1 model, trained on restricted hardware, persuaded Chinese officials that domestic firms can get far with less, and it knocked more market value off Nvidia in a single day than any company had lost before. Discouraging purchases of a deliberately weakened American chip pushes Chinese buyers towards Huawei's Ascend line, which is the industrial policy Beijing wants anyway.

Four customers carry the load

Without China, Nvidia's future growth depends almost entirely on the capital budgets of four American hyperscalers: Microsoft, Meta, Alphabet and Amazon.

For now those budgets are rising. In July Alphabet raised its capital spending forecast for this year to $85 billion and Meta narrowed its range to $66 billion to $72 billion, both pointing to AI infrastructure. The spending is funded mainly from operating cash flow, which makes it steadier than the debt-financed telecoms build-out of the late 1990s. That episode is still the relevant warning. When carriers cut spending in 2001, Cisco's revenue fell and its shares, briefly the world's most valuable in 2000, have never regained their peak.

The arithmetic runs from their income statements to Nvidia's. Every dollar a hyperscaler spends on servers depreciates over several years, so the cost lands on its earnings gradually while the revenue lands on Nvidia's at once. As long as AI services bring in enough to cover those rising depreciation charges, boards will keep signing off. If they do not, the pressure to slow orders will build first in the budgets of the companies with the thinnest AI revenue.

Beyond the big four

The counterargument is that the customer base is broadening. Sovereign buyers such as Saudi Arabia's Humain, specialist cloud providers such as CoreWeave and model builders such as OpenAI are all buying in volume. Many of those buyers, however, rely on contracts or funding that lead back to the same hyperscalers, and none is yet large enough to offset a pause at Microsoft or Amazon.

The concentration cuts both ways for shareholders. As long as the four keep competing to build the largest AI clusters, Nvidia enjoys pricing power and gross margins that few hardware makers have ever seen. Once any one of them decides it has enough capacity, or shifts more work to the in-house chips that Alphabet, Amazon and Microsoft are all developing, the order book shrinks with little warning.

A valuation priced for continuity

The market already treats that concentration as tolerable. Nvidia became the first company to pass $4 trillion in market value in July, and the shares slipped only modestly after the results. A valuation of that size assumes years of data-centre growth, not one strong quarter, which is why the buyback matters less than the order books behind it. Repurchases return cash; they do not replace a missing market.

The strongest case for the bulls is that China was always going to be lost, and that Nvidia has absorbed the loss while still guiding to record revenue. That is fair as far as it goes. The flaw is that the same political forces that closed China can reach elsewhere, through licence requirements on sales to the Gulf or tariffs on semiconductors that the White House has already floated.

What to watch next

The next signals come from the hyperscalers' own results in late October, when capital spending plans for next year start to take shape, and from any sign that Beijing lets the H20 back in. Dropping China from its guidance protects Nvidia's forecasts from regulators' whims, but it leaves the shares as an unhedged bet that those four companies keep raising capital spending, and any pause in their budgets will hit the stock directly.