Nvidia's 106% Growth Is the End of One Cycle. Inference Is the Reason It Keeps Going

Nvidia reported a fiscal second quarter on Aug. 26 that most companies could not produce: $96.2 billion in revenue, up 106% year over year, with gross margins of 75%. Five days later the stock is up only a few percent and still about 7% below the record it set in May. A muted reaction to a doubling quarter is itself the signal — the market has stopped being impressed by growth it already expected. The open question is what the company is becoming, and the answer inside the numbers is that Nvidia has already crossed from the build phase of AI to the run phase.

The distinction is the whole story. Training is the expensive, one-time process that turns data into a capable model. Inference is the recurring cost of running that model — every answer, every generated token, every autonomous step an agent takes is inference consuming compute. Nvidia conquered training through CUDA, the software layer that locked in developers for more than a decade. That phase built one of the most valuable companies on Earth. The market's error is to keep pricing the phase that is ending.

Management said plainly that the center of gravity has moved. "AI has reached its inflection point," Jensen Huang told analysts. "It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue." Behind the slogan is a measurable change, and it is already visible in the revenue mix. Nvidia split its $89 billion data center quarter into two buckets: hyperscale — the big public clouds — at $49 billion, up 13% sequentially, and everything else, which it calls ACIE, at $40 billion, up 25% sequentially and 138% year over year. The result is nearly a reversal of roles: the customer base that defined the last two years is now the slower-growing half, and the fastest-growing customer base is the one that makes money by running AI rather than the one that only builds it. Sovereign AI revenue roughly tripled from a year ago, and data center networking revenue rose 138%.

The economics explain why this matters. Training demand arrives in bursts and can saturate. Inference is a meter that runs forever, and agentic workloads — AI that chains many steps to finish a task instead of answering one question — can consume 15 to 100 times the compute of a person using a tool. In that world, an operator's revenue is a direct function of the compute it can bring online. Huang put the escalation in dollar terms: revenue per gigawatt of deployed capacity has climbed from about $18 billion in the Hopper generation to $25 billion for Grace Blackwell and $40 billion for the new Vera Rubin platform, now in full production with purchase orders from every major customer.

That is why management still sees headroom. The current quarter is guided to $108 billion with no China data center compute revenue assumed, and Vera Rubin is expected to be about a fifth of data center revenue. For the following fiscal year, Nvidia laid out roughly 70% revenue growth and called it supply-constrained — the limit is how fast it and its memory, power, and foundry partners can physically build, not whether customers still want the product. Applied to the roughly $400 billion Nvidia will record this year, 70% implies revenue on the order of $700 billion in the fiscal year ending January 2028.

The caveat is that the market has heard grand order figures before. At GTC in March, Huang said Nvidia sees $1 trillion in orders for Blackwell and Vera Rubin through 2027 — double the $500 billion figure he cited a year earlier. But that figure describes a pipeline, not a result: firm purchase orders, letters of intent, and multiyear forecasts, with roughly 40% of the value expected to be recognized as revenue by the end of 2027. Commitments are not bookings. The conversion rate is the line to watch.

Those same commitments are Nvidia's largest current risk, and they are made of memory. A high-end AI rack now spends roughly a quarter of its build cost on DRAM and high-bandwidth memory, up from a small share two years ago, and conventional DRAM contract prices surged 90% to 95% in the first quarter of 2026. Supply commitments have swollen to $279 billion, mostly memory for Vera Rubin. This is the cost that drags gross margin from 75% toward a guided 71%-72% trough this fiscal year before settling near 72%-73%. Nvidia's answer is pricing: it has reportedly told customers to expect more than 15% increases on Vera Rubin and Grace Blackwell systems shipping early next year, and it has locked up multi-year HBM supply with SK hynix and Micron. If those increases stick, the trough is temporary and the pricing-power story holds. If customers push back, or if inference demand grows more slowly than the order pipeline implies, the commitments and the margin floor become the same problem from two directions.

So the useful judgment is not "buy the growth." Nvidia at roughly $5.3 trillion and about 17-18 times trailing sales is not cheap on any historical basis, and its near-term path is now determined by two numbers: how deep the margin trough goes and whether the $1 trillion pipeline converts the way management implies. The reason to keep considering it is the more durable one — the company now grows because its customers earn revenue from the compute they buy, not because a training boom is forcing them to. That is the difference between a cycle that ends and a cycle that compounds. Whether it justifies a position at this price is the allocation question every investor has to answer against the alternatives — and it is the only question that matters.