Nvidia's $60bn per-gigawatt math makes power the AI prize

Five years ago, Nvidia counted $30 billion of revenue for each gigawatt of data center it supported; on Wednesday CFO Colette Kress put the figure at $60 billion, a doubling that tells the market more about the future of the AI buildout than the $96.2 billion in quarterly revenue it came with.

The headline numbers are staggering: second-quarter revenue of $96.2 billion, up 18 percent quarter over quarter and 106 percent year over year, including $89 billion in data center revenue. Hyperscalers contributed $49 billion, while the company's "AI clouds, industrial and enterprise" segment, ACIE, added $40 billion, a 25 percent jump Kress attributed to neocloud capacity additions for enterprises, AI startups and sovereigns, as well as hyperscalers buying capacity to supplement their own builds.

The guidance is just as bracing, with Kress saying Nvidia expects fiscal 2028 revenue growth of 70 percent, above market expectations, and that the company is on track to double revenue next year while remaining supply constrained. "AI has reached its inflection point," said founder Jensen Huang. "It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue."

That per-gigawatt metric captures a shift in where the value of AI infrastructure sits: five years ago a gigawatt of data center generated roughly $30 billion in Nvidia revenue, today it generates $60 billion. The company's expansion up the stack — from GPUs to CPUs to networking to the newly acquired Groq hardware — means Nvidia extracts more from every facility it touches, and the industry's bottleneck has shifted from land and chips to electrons: a data center's power contract is now the highest-leverage asset in the deal.

The demand across customer types is what makes that per-gigawatt figure possible: neoclouds adding capacity for enterprises, AI start-ups, and sovereigns while hyperscalers buy Nvidia gear to supplement their own builds. Huang is careful to distinguish this moment from last year's single-lab boom — "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," he said — and Vera Rubin, now in full production, was built for exactly this. Kress said Nvidia remains supply constrained even as it guides to double revenue next year.

Nvidia is also behaving like an infrastructure investor, using its balance sheet to back customers including OpenAI's massive SoftBank data center project in Ohio. Kress said Nvidia expects those customers to become "the largest technology companies in history," and that demand from AI labs it backs could contribute roughly a quarter of Nvidia's business next year. This publication previously reported that Nvidia weighs a $3 billion SB Energy investment behind the Ohio campus — the guarantee, more than the equity, is what would make the $500 billion project financeable.

The balance-sheet strategy changes the capital hierarchy: Nvidia's stake in Nebius supported $5.75 billion in debt with a four-point coupon spread that priced the near-term buildout as infrastructure and the long haul as a risk asset. The same logic applies here: Nvidia is supplying collateral for its own future revenue, and if Nvidia makes $60 billion per gigawatt, deploying a few billion dollars of balance sheet to secure that gigawatt is cheap capital. But Kress's claim that the risk is "limited" deserves scrutiny. The bet concentrates Nvidia's revenue on a handful of AI labs; if one fails, Nvidia's chip sales and its balance sheet take a hit at the same time — the nature of vertical integration in reverse, Nvidia writing the investment case for its own downstream.

For infrastructure investors, the upshot is that Nvidia-linked assets now price as near-infrastructure while the underlying equity is still risk capital — the distinction between hyperscaler-anchored assets and everything else. The practical consequence is that the race for power and grid access will only intensify, because Nvidia's numbers put a price tag on electrons: $60 billion of chip revenue per gigawatt, up from $30 billion five years ago. That premium will flow into power rights, interconnection queues, and land with permits, where operators that secure the power first will command the returns — and Nvidia's balance sheet will be behind the best-connected of them.

The next thing to watch is whether Nvidia's $3 billion guarantee on the Ohio project becomes a final investment; if it does, chipmakers will underwrite power projects to protect their revenue per gigawatt. The market should price power accordingly — and the PPA, not the chip, becomes the most valuable document in the capital stack.