Nvidia Projects 70% Growth by 2028 and Admits Demand Exceeds Capacity

Blackwell Ultra Drives Data Center to $89 Billion

Nvidia reported $96.2 billion in revenue for the second fiscal quarter of 2027, ended July 26, an increase of 18% from the previous quarter and 106% year-over-year. This figure exceeded the upper end of its own guidance of $91 billion set in May by $5.2 billion, during which the company was still operating under regulatory uncertainty concerning exports to China.

The Data Center accounted for 92% of consolidated revenue and grew 117% on a year-over-year basis, reaching $89 billion. Hyperscale revenue generated by AWS, Azure, Google Cloud, Oracle, and Meta more than doubled in twelve months. Colette Kress, CFO, attributed the sequential increase of 18% to the ramp-up of Blackwell Ultra, the architecture that succeeded GB200 and began contributing to sales mix starting in May. The gross margin was 75% GAAP.

For the third quarter, the company guided revenue of $108 billion, still maintaining a 75% margin. The estimate does not incorporate any revenue from H20 chips in China, despite obtaining an export authorization in July. According to Kress, Nvidia "has no orders at this time" from the Chinese market, and any local revenue within the quarter would be considered upside to the guidance.

"Compute is Revenue," Says Huang to Skeptics

Jensen Huang utilized the call to reposition the discussion regarding the return on capex from big techs. "AI has reached its inflection point. It is doing useful work, tokens are productive and profitable. Now, computing is revenue," said the CEO. This statement sounded like a direct rebuttal to the public skepticism of Michael Burry, who reported a short position in AI infrastructure stocks that same week, and Ed Zitron's memos on the AI bubble.

Huang predicted 70% revenue growth for the fiscal year 2028, a figure significantly above Wall Street's consensus, which hovered around 45%. Following that came the most relevant caveat of the call. "Our demand is far above 70%," he stated. The bottleneck, according to the executive, has shifted to the schedules of TSMC and SK Hynix, not to orders. The company is working with suppliers to expand capacity, without forecasting when the imbalance will close.

Sovereign AI Turns the Tide with 100% Growth

Kress introduced a new segment in the report, named ACIE, which aggregates direct corporate customers, neoclouds such as CoreWeave, and sovereign projects. This bloc already accounts for half of Data Center revenue and is growing at 100% year-over-year. The company estimated the TAM for the Vera Rubin generation, expected in the second half of 2027, at $40 billion, compared to $25 billion for Blackwell and $18 billion for Hopper.

Huang characterized sovereign projects as the toughest vector in the portfolio, with strong multi-year contracts. "This year we have a golden age of new AI labs and startups, with multiple frontier labs scaling in parallel and a vibrant ecosystem of open models," the executive contrasted, compared to 2025, when a single customer concentrated the majority of training demand.

The Bottleneck Is Now Distribution

For the CIO in Frankfurt and the CTO in Seoul, the relevant number from the call was not the top line, but the admission that Vera Rubin is already committed to Western sovereign clients and three American hyperscalers. The European corporate client that planned its own H100 capacity at OVH or T-Systems is likely to run workloads under third-party allocation, not through direct purchase.

In Brazil, the effect impacts two immediate channels. The first is the unit price of H100 in the secondary market, pressured as Blackwell Ultra concentrates new contracts with hyperscalers. The second is the pent-up demand from banks and consultancies that placed medium-sized clusters on the roadmap for 2026, including Itaú, Bradesco, and local operations of Accenture and TCS. With Vera Rubin committed to sovereigns until the end of 2027, the most likely scenario for these institutions is to migrate workloads to instances in Miami and Ashburn, where AWS and Oracle still maintain contracted allocation.

The clinical reading is that Nvidia's problem has shifted from selling to distributing. Demand exists, hyperscale capex exists, and Huang has a ready response to the ever-skeptical critic. What the call did not address is how the company will cater to the long tail of the corporate market, which signed AI pipelines last year and is now entering a queue that starts in Washington, Berlin, and Tokyo.