Nvidia reports a $96.2 billion quarter, guides to $108 billion, and forecasts a full year ahead for the first time — with zero China revenue in the numbers

Nvidia has spent four consecutive quarters beating its own guidance and watching its stock fall the next morning anyway. The pattern had become its own kind of commentary: the numbers were never the problem, the question was always whether the customers buying them could keep it up. On Thursday the pattern broke. Shares rose 7.2 per cent in premarket trading after the company reported second-quarter revenue of $96.2 billion, up 106 per cent from a year ago and 18 per cent from the previous quarter, and guided the current quarter to $108 billion.

What moved the stock was not the quarter. It was a sentence from the chief financial officer that Nvidia had never said before.

The first annual forecast in the company's history

Colette Kress told investors on Wednesday that Nvidia expects revenue growth of about 70 per cent for fiscal 2028 — the year running from February 2027 to January 2028. Nvidia does not forecast a year in advance. Jensen Huang said so explicitly on the call, adding that the company now has "a lot greater visibility" across its supply chain than it has ever had, which is why it was willing to try.

He then went further, in the direction that matters. Demand, he said, "is much greater than 70 per cent." The forecast is not a demand estimate. It is a supply estimate. Nvidia is telling the market what it thinks it can physically build, and the gap between that and what customers want to buy is now the single most important unmeasured number in the semiconductor industry.

The constraints are the familiar ones and they are getting worse rather than better. Taiwan Semiconductor Manufacturing Company, which fabricates essentially everything Nvidia sells, remains capacity-constrained at the advanced nodes. High-bandwidth memory, the other bottleneck in an AI system, is in shortage across the industry. Neither is a problem Nvidia can solve with its own balance sheet inside a twelve-month window.

The quarter itself

Data centre revenue was $89.0 billion, up 117 per cent year on year and 18 per cent sequentially — which is to say the segment is now larger than the entire company was two quarters ago. Gross margin came in at 75.0 per cent on both a GAAP and non-GAAP basis, a tenth of a point better than the prior quarter and 2.6 points better than a year ago, which is a remarkable thing to achieve while shipping a brand-new rack-scale platform into volume.

Operating income was $63.7 billion, up 124 per cent. GAAP diluted earnings per share came in at $2.46 against $1.08 a year ago; on a non-GAAP basis, $2.22 against $1.01. Operating expenses grew 55 per cent to $8.4 billion — fast, but a third of the rate at which the top line grew. For the first half of the fiscal year, Nvidia has now booked $177.8 billion of revenue against $90.8 billion in the same period last year.

The company returned about $26.0 billion to shareholders during the quarter through buybacks and dividends and finished with roughly $99.0 billion still authorised for repurchase. The next quarterly dividend of 25 cents a share is payable on October 1 to holders of record on September 10.

The customer-concentration answer

The bear case on Nvidia has never really been about demand in the aggregate. It has been about whose demand. A business in which a handful of hyperscalers account for the majority of revenue is a business whose growth rate is a decision made in four or five boardrooms. Nvidia has been trying to demonstrate otherwise for a year, and this quarter it put a number on the effort: customers it classifies as AI clouds, industrial and enterprise — the ACIE bucket — accounted for $40.3 billion of sales, up 138 per cent year on year.

That is a growth rate faster than the company overall, from a base that is now more than 40 per cent of revenue. It is the strongest evidence Nvidia has produced that the buyer base is genuinely broadening rather than simply concentrating further.

This time last year, one lab alone was driving the buildout. Today, we have a golden age of new AI labs and startups, multiple frontier labs scaling in parallel, a thriving open-model ecosystem and physical AI coming online.

Huang's framing of the moment was blunter than usual: AI has "reached its inflection point," he said, and "compute is revenue." The claim underneath the slogan is that inference has become a business with unit economics rather than a research cost centre — that the tokens being generated are, in his words, productive and profitable. If that is true, the capital expenditure cycle stops being a bet and starts being a supply chain.

What the guidance quietly assumes

The $108 billion outlook for the third quarter carries a footnote worth more than most of the release: Nvidia is not assuming any data centre compute revenue from China. The company that once counted China as a mid-teens share of data centre sales has now written it to zero in its own forecast and is still guiding to 12 per cent sequential growth. Gross margin is guided down slightly to 74.0 per cent, plus or minus 50 basis points, with operating expenses rising to about $9.2 billion.

Zeroing out China is defensive accounting, and it is also a strategic statement. Any policy thaw is now pure upside to the model rather than an assumption that has to be defended. The cost is that Nvidia has stopped arguing for a market it spent two years trying to keep.

The threat the analysts actually named

The note of caution on Thursday came from a different direction than usual. Several analysts pointed to custom silicon — the inference accelerators being designed by hyperscalers and by the AI labs themselves — as the real medium-term threat to Nvidia's near-monopoly on frontier training and inference hardware. That threat is no longer theoretical; purpose-built inference parts have been announced by multiple large buyers within the last month.

Nvidia's answer, visible in the quarter's product disclosures, is to make the rack rather than the chip the unit of competition. The Vera Rubin platform is ramping into full production with systems running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. Spectrum-6 switching, supporting both pluggable and co-packaged optics, is shipping into gigascale sites. Vera, a CPU designed specifically for agentic workloads, has been announced with broad partner adoption. A competitor can build a better inference die. Replacing an integrated rack, its networking fabric and its software stack is a different order of problem.

Reported separately on Wednesday, and not addressed in the earnings release, was a deal that fits the same logic: Nvidia has agreed to acquire the open-source model platform Hugging Face for $12.9 billion, according to The Information, with Business Insider reporting that talks had been under way. If it closes, the most widely used distribution point for open-weight models would sit inside the company that makes the hardware they run on.

There is also a product line in the release that would have been unthinkable eighteen months ago: Nvidia says the Groq 3 LPX interactive inference accelerator is now in full production under its own banner. Buying the competitive answer to your own weakest position — low-latency, low-cost inference — and shipping it as part of your platform is the move of a company that has decided it would rather cannibalise itself than let somebody else do it.

The read-across

The rest of the complex moved with it. Micron rose 4.5 per cent in premarket trading, Marvell 5.7 per cent, Arm 4.7 per cent, Intel 3 per cent and Advanced Micro Devices 1.7 per cent. The neocloud names, whose entire business model is a leveraged bet on Nvidia's roadmap holding, moved hardest: Nebius up around 7.5 per cent and CoreWeave around 6 per cent.

That is the same group that lost roughly $1 trillion of combined market value during July's AI selloff before recovering. The volatility is the point. What Nvidia bought itself on Thursday is not a rerating so much as a longer fuse — a supply-based forecast that extends visibility through to January 2028, delivered by a company with enough manufacturing intelligence to make it credible. Paul Meeks of Freedom Capital Markets put the resulting time horizon plainly: he does not see a slowdown risk before 2028 at the earliest.

The obvious rejoinder is that a supply-limited forecast is only conservative while demand exceeds supply, and every number in this release depends on that continuing to be true. Nvidia has now told the market it can see twelve months out. The next four quarters will establish whether that was visibility or confidence.

Figures are from Nvidia's second-quarter fiscal 2027 results release for the quarter ended July 26, 2026, and from the company's outlook statement. Share moves are premarket on Thursday, August 27, as reported by CNBC. The Hugging Face acquisition has been reported but not confirmed by Nvidia.