Nvidia published its second-quarter fiscal 2027 report, for the quarter ended 26 July. Revenue came to $96.2 billion.
But the question everyone asked before the report was not what happened last quarter - it was whether demand is cooling. And the answer sits in a single line of guidance.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
What This Company Sells Today
Nvidia is no longer a graphics card maker that also sells to data centres. The ratio has inverted completely.
92.5% of the quarter's revenue came from Data Center - that is, the company is a compute infrastructure supplier to clouds and AI labs, and what it sells is not a single chip but a complete system: accelerators, CPUs, network switches, optical cabling, software and the design of an entire rack.
| What entered this quarter | |
|---|---|
| Vera Rubin | The next-generation platform, in full production - installed at CoreWeave, Google Cloud, Azure, Oracle and Nebius |
| Spectrum-6 | Network switches supporting both pluggable and co-packaged optics |
| Vera | A CPU built specifically for AI agents |
| Groq 3 LPX | An interactive inference accelerator, in full production |
And this explains why it is hard to compare against a rival: selling an accelerator means competing on a product; selling a whole rack means competing on an architecture. I set out the three physical bottlenecks of this infrastructure separately - power, cooling and light.
The Quarter
| $ millions | The quarter | Prior quarter | A year earlier | Y/Y change |
|---|---|---|---|---|
| Revenue | 96,221 | 81,615 | 46,743 | +106% |
| of which Data Center | 89,023 | +117% | ||
| Gross margin | 75.0% | 74.9% | 72.4% | +2.6 pts |
| Operating expenses | 8,408 | 7,621 | 5,413 | +55% |
| Operating income | 63,734 | 53,536 | 28,440 | +124% |
| Net income | 59,688 | 58,321 | 26,422 | +126% |
| Diluted EPS | $2.46 | 2.39 | 1.08 | +128% |
| Non-GAAP EPS | $2.22 | 1.87 | 1.01 | +120% |
The consensus stood at about $92.1 billion of revenue and $2.09 per share. The company itself guided to $91.0 billion.
The operating margin stands at 66.2%. For comparison, a year earlier it was 60.8%. A company of this size expanding its operating margin by five percentage points in a year is a rare thing.
And the Line That Answers the Question
Third-quarter guidance: $108.0 billion, plus or minus 2%.
That is growth of 12.2% in a single quarter - on a base of $96 billion.
And above it sits one sentence that changes the meaning:
"NVIDIA is not assuming any Data Center compute revenue from China in its outlook."
That is, those $108 billion are without the Chinese market entirely. Not "on a conservative assumption" and not "partially" - zero.
Anyone who asked whether demand is cooling has an answer: it is not. Not in the quarter reported, and not in the quarter guided - and not because a market reopened, but despite a market being taken out of the equation.
And in fairness, the caveat needs its own caveat: the release notes that shipments of Data Center Hopper products to China in the quarter were less than 1% of Data Center revenue. That is, China is already almost absent from the base - so assuming zero is not a large concession, but a statement of what is already the case.
And Jensen Huang put it this way in the release: "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue."
And 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."
That claim is testable, and that is what distinguishes it: if demand were concentrated in one buyer, customer concentration in the annual report would show it. If it has genuinely broadened, concentration will fall. That is a published number, and it can be followed.
The Figure That Proves - or Disproves - the Demand Claim
Huang argued that demand has broadened from one lab to many. That is a testable claim, and the company publishes the number itself.
| $ millions | This quarter | Prior quarter | A year earlier | Q/Q | Y/Y |
|---|---|---|---|---|---|
| Data Center | 89,023 | 75,246 | 41,096 | +18% | +117% |
| of which Hyperscale | 48,710 | 43,050 | 24,168 | +13% | +102% |
| of which AI Clouds, Industrial & Enterprise | 40,313 | 32,196 | 16,928 | +25% | +138% |
| Edge Computing | 7,198 | 6,369 | 5,647 | +13% | +27% |
The category that is not the hyperscalers is growing at almost twice their rate - and that is the answer.
25% in the quarter against 13%. 138% in the year against 102%.
And translated into share:
| That category's share of Data Center | |
|---|---|
| A year ago | 41.2% |
| Prior quarter | 42.8% |
| This quarter | 45.3% |
That is, four percentage points moved in a year from the giants to everyone else.
And what makes the comparison valid: the company notes that this quarter it reclassified one company out of that category and into Hyperscale - "and recast the prior period revenue associated with this company" accordingly. So the comparison is clean, and it is even conservative: the company removed from the smaller category went to strengthen the larger one.
And who sits in that category, per the release: "end-demand from AI natives, enterprises, and sovereign customers, as well as hyperscalers utilizing AI clouds."
"Sovereign customers" are states building their own national compute infrastructure - a category that did not exist two years ago.
And What Does Deserve Attention
Two lines pull the other way.
The first - the gross margin is guided down. This quarter it is 75.0%. Guidance for next quarter is 74.0% plus or minus 50 basis points.
That is a full percentage point, on a base of $108 billion. The company does not detail the reason in the release. In a business growing at this pace, a falling margin can indicate a different product mix, a higher component cost - memory, advanced packaging - or pricing. The release does not say, and I will not guess.
And the second - inventory. It rose to $31,575 million, from $21,403 million at fiscal year end. An increase of about 47%.
Growing inventory is a double-edged thing: at a company where demand exceeds supply, it is a build-up ahead of future shipments - and that is positive. At a company where demand is cooling, it is the first sign. What decides between the two readings is precisely the guidance - and $108 billion supports the first.
And on the margin, there is one hint in the release worth noting: the company explains that the margin rose year over year on improved mix from Blackwell Ultra, and that it was roughly flat sequentially "as our Blackwell architecture remains the vast majority of our revenue".
Vera Rubin is entering full production now. A transition to a new platform typically begins at a lower margin, which rises as manufacturing matures. That is a possible explanation for the percentage point - but the company did not say so, and it is therefore my inference and not a fact.
And the Line That Confirms Another Report Published the Same Hour
Nvidia's Edge Computing brought in $7.2 billion, up 27%. And the explanation in the release is interesting:
"The increases were driven by strong sales of Blackwell workstations, partially offset by slower consumer PC sales that were tempered by elevated memory and systems prices."
That same evening HP reported the complementary opposite: PC revenue rose 18%, and the number of units sold fell 16% - meaning the average price jumped about 40%, and management there explicitly mentioned memory supply.
Two companies that do not compete with each other, reporting the same hour, pointing at the same thing: memory prices are making PCs more expensive and shrinking volumes.
And for anyone following the infrastructure chain, this is the detail that links them: the same memory required for AI accelerators is the memory missing from the laptop - and when industrial demand takes the supply, the consumer pays the price.
And the New Thing in This Report
Nvidia announced partnerships to establish independent compute financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR - aimed at mobilising over $500 billion of third-party capital for the AI infrastructure buildout over time, subject to definitive agreements.
And this detail looks technical and is not.
The logic is simple: a data centre is a capital-heavy asset built years before it generates revenue. Whoever buys accelerators has to finance them, and as the sums grow, financing capacity becomes a constraint in its own right.
And that is exactly what I wrote a few days ago in the piece on the three vectors of AI infrastructure - that the fourth constraint, after power, cooling and light, is the debt market. Order cover on hyperscaler issues fell there from five times in February to below two in July.
What Nvidia is doing here is bringing alternative institutional capital into the equation - infrastructure funds and private equity instead of, or alongside, the corporate bond market.
And the balance sheet points the same way: the company's non-marketable securities rose to $51,157 million, from $22,251 million at fiscal year end. The company is investing in its own customers and partners at more than twice the previous scale.
That strengthens the ecosystem. It also ties the fate of its revenue to the fate of its investments - a question worth following, neither dismissed nor accepted automatically.
Two More Lines From the Balance Sheet
| $ millions | The quarter | Fiscal year end |
|---|---|---|
| Inventories | 31,575 | 21,403 |
| Non-marketable securities | 51,157 | 22,251 |
| Total assets | 320,272 | 206,803 |
And on returns to shareholders: about $26.0 billion in the quarter through buybacks and dividends, with about $99.0 billion remaining under the authorisation. The next dividend, $0.25 per share, is payable on 1 October.
What I Will Check Next Quarter
| The gross margin | Whether 74.0% is a point or the start of a trend |
| China | Whether anything enters the outlook - the assumption today is zero |
| Customer concentration | Whether the "multiple labs" claim shows up in the number |
| Inventory | Whether it keeps growing faster than revenue |
| The $500 billion | Whether the definitive agreements are signed, and on what terms |
| Non-marketable securities | The pace at which the company invests in its own chain |
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
The demand question has an answer, and I want to break it into three parts - because they are not the same thing.
The first part: does demand exist today. Yes, and there is no argument about it. $96 billion in the quarter, $108 billion guided, and without China. That is measured, reported, and testable next quarter.
The second part, and the one that interests me far more: who it comes from. A year ago that had a troubling answer - a small number of enormous buyers. This quarter's figure changes the picture: the non-hyperscale category grew 25% in the quarter against 13%, and its share rose from 41.2% to 45.3% over the year. AI natives, enterprises, and sovereign customers.
And the difference between those two pictures is a difference in the type of risk, not its size. Concentration among five buyers is decision risk - it takes two of them slowing a budget. Dispersion across hundreds of buyers, states and enterprises is cycle risk - it does not stop on one decision, but when it turns, it turns for everyone at once.
I am not sure the second is preferable. I do think it is different, and that anyone still analysing this company through "what will the hyperscalers do" is analysing a market structure that has already changed.
The third part, and to my mind the most significant for the long run: what Nvidia has become.
It is customary to call it a chipmaker. Look at what entered a single press release:
| CPU | Vera - the first CPU built for AI agents |
| Networking | Spectrum-6, including co-packaged optics |
| Storage and security | Vera BlueField-4 STX - storage processing with in-silicon security |
| Software | CUDA-X libraries, PhysicsNeMo, BioNeMo |
| Measurement | MLPerf, and AgentPerf - the first industry benchmark for agentic infrastructure |
| Land and power | Land, power and shell capacity at an Ohio campus, partnered with SB Energy |
| Finance | Platforms to mobilise over $500 billion with six funds and banks |
| Equity | Non-marketable securities that jumped to $51.2 billion |
This is not a chip company that expanded. This is a company that built itself an entire chain, and then began building its own demand as well.
And that is the thing I am most torn about.
On one hand, it is entirely logical. If the bottleneck is not the chip but power, cooling and light - and after them financing - then a company that wants to sell more accelerators has to open those bottlenecks itself. Securing land and power in Ohio is not business diversification; it is removing an obstacle.
On the other hand, that structure blurs an important line. When a company invests $51 billion in its own customers, and arranges $500 billion of financing for its buyers - part of its revenue comes from money it helped put in place.
I am not arguing that this is improper, and it is not new historically - telecoms equipment vendors did exactly this in the late 1990s. I am arguing that it changes what has to be measured. At such a company, "revenue growth" is no longer an independent gauge of external demand.
So the three lines I will look for going forward are not revenue.
The first - whether the non-hyperscale share keeps rising. If it does, the broadening is real. If it stalls around 45%, what we saw was a single wave.
The second - whether the $500 billion is signed, and on what terms. An announcement of intent is not an agreement, and the release itself qualifies it as "subject to definitive agreements". The terms will say who carries the risk.
And the third - the gross margin. 75.0% today, 74.0% guided. In a business growing 106% a falling margin is not troubling in itself, but it does ask who is paying for the growth - and we will see that answer once Vera Rubin has been in full production for a quarter or two.
And finally, the reservation I hold for myself: this entire picture - unlimited demand, inventory building, financing arranged externally for buyers - has appeared in every major infrastructure cycle in history, before the turn. That does not mean it ends the same way this time. It does mean that the claim "this time is different" has to rest on data rather than enthusiasm - and that is why I have tried to bring the data here.
(It is important to stress: this is my personal opinion only, and nothing herein constitutes a recommendation to take any action.)
