Nvidia is bracing for 70 percent growth, and one system now costs 8.5 million dollars

Jensen Huang turned up at the Goldman Sachs Communacopia and Technology conference to explain why he thinks Nvidia is not slowing down. "Most people think Nvidia builds a chip. You need airplanes to carry what we build," the chief executive said, in remarks quoted by TechCrunch.

His sharpest comparison was about price. "One GPU today is not 399 dollars. It is 8.5 million dollars. One GPU, everything connected by NVLink, 2 million parts, 250,000 kilowatts," Huang said. One specific system, in which 36 Grace processors link up with 72 Blackwell chips, is drawing 27 percent more orders every month.

That is not just marketing talk. Second-quarter results, published alongside the announcement of an expanded deal with Amazon, showed 96.2 billion dollars in sales revenue. Data centers brought in 89 billion dollars, up 117 percent from a year earlier. The company forecast 108 billion dollars in revenue for the third quarter, partly on the strength of its new-generation Rubin chips.

Demand shows up in the size of the orders. Amazon doubled down: after ordering more than a million Nvidia chips for 2026, it is adding another 2 million GPUs for AWS data centers in 2027 and 2028, including Blackwell Ultra, Rubin and Rubin Ultra. The American company is also building its own chips. In its latest quarterly report it said its in-house chip business passed 25 billion dollars in annualized revenue, resting on 225 billion dollars of commitments from AI labs such as Anthropic and OpenAI.

Huang names the competition outright: hyperscalers, meaning Amazon, Microsoft and Google, building their own chips; AI labs doing the same; the already listed Cerebras and startups such as Etched. His answer comes down to the scale of the system rather than a single chip, and that is the heart of this year's narrative about money in AI. Customers are not buying silicon. They are buying entire installations, ready to work, drawing hundreds of kilowatts.

Nvidia's answer to questions about competition is financial, not technical. The in-house systems of the hyperscalers and the specialized chips of startups handle individual inference tasks, but they do not replace a whole training installation. If a customer is buying not a card but a ready system drawing hundreds of kilowatts, a growing number of units ordered translates into revenue at a pace no maker of inference chips can match. That is why the 70 percent growth forecast does not sound like a boast in this narrative but like a consequence of the price list: the bigger the installation, the bigger the bill, and the bill today is paid in billions of dollars.