Nvidia starts shipping H200 chips to China, but forecasts zero data center revenue

Nvidia confirmed the first shipment of H200 data center processors to China under a new United States licensing scheme, ending a block that had lasted months. The information was disclosed on an earnings call and reported by South China Morning Post. The detail that matters: despite the symbolism of the unblocking, the money involved is practically irrelevant to the company.

The number that undercuts the narrative

According to Nvidia, H200 sales to China accounted for less than 1% of the $89 billion in data center revenue in the second fiscal quarter, ended July 26. To put that in perspective: 1% of that total would already be less than $890 million, and the company says it came in below that.

The disclosure marks a change from the previous quarter, when Nvidia said it had made no shipments of Data Center Hopper products to China. In other words, it moved off absolute zero, but to a level that is still marginal.

More revealing is what the company projects going forward: the $108 billion revenue forecast for the third quarter still assumes zero data center computing revenue from China. In practice, Nvidia is planning its next few months as if the Chinese market didn't exist for that product line.

Why the H200 barely reaches China

The path to this point was blocked on both sides. Washington approved H200 exports in January. Beijing, however, initially restricted purchases while promoting domestic manufacturing alternatives, and only started allowing selected Chinese AI companies to buy limited quantities starting last month, according to the SCMP.

It's a game of two overlapping restrictions:

| Side | Stance | |---|---| | US | Approved H200 export in January, under licensing | | China | Restricted purchases, encouraged domestic chips, only opened up to selected buyers and limited volumes |

The result is a narrow, politically sensitive flow, in which Nvidia cannot (nor does it project to) turn China into a revenue pillar.

The optimistic tone comes from elsewhere

Despite China's irrelevant contribution, Nvidia was emphatic about global demand. The company said it expects revenue to grow around 70% in fiscal year 2028, describing that projection as limited by supply, not demand.

CEO Jensen Huang was direct in saying that actual demand is "much greater than 70%," and that the number reflects available supply capacity, not market appetite.

The forecast without constraints would be much higher.

>

-- Jensen Huang, CEO of Nvidia

Huang added that the company has secured significant capacity, but still needs "a lot more." He also broke down the composition of that demand: hyperscalers account for only about half of the opportunity, with the rest coming from enterprises, neo clouds, and sovereign AI projects.

What changes for those building software in Brazil

The Brazilian reading of this earnings report doesn't run through China, it runs through the supply queue. When Nvidia says its growth is limited by supply, not demand, and that it still needs "a lot more" capacity, the indirect message for those far from the big clients is clear: cutting-edge accelerators remain a scarce, contested resource.

Some concrete points worth watching closely:

• Allocation priority. With Huang saying hyperscalers are only half the opportunity and the rest comes from enterprises, neo clouds, and sovereign AI, the fight for GPUs is no longer just among big techs. This matters for those procuring training and inference capacity through cloud providers in Brazil, who depend on how these accelerators are distributed globally.
• Sovereign AI enters the radar. Huang explicitly cited sovereign AI projects as a significant share of demand. This is the category that national AI infrastructure initiatives fall into, something Brazil has been discussing. The mention shows that Nvidia treats governments as major clients, not an afterthought.
• Cost and availability. Tight supply tends to keep GPU access prices under pressure. For engineering teams planning to train or run larger models, the lesson is to plan capacity in advance and design architectures that don't assume abundant, cheap cutting-edge GPUs in the short term.

What remains open

The H200 episode shows that the Chinese market today is more a piece on a geopolitical chessboard than a revenue source for Nvidia, so much so that the company itself zeroes it out in its projections. What remains undefined is whether and when China will once again factor into revenue, given that Beijing keeps pushing domestic alternatives and releasing purchases in dribs and drabs.

For Brazilian developers, the most actionable fact isn't the unblocking itself, but the confirmation that the AI race remains strangled by chip manufacturing capacity, not by a lack of projects wanting to run. Anyone who depends on this infrastructure, directly or indirectly, does well to treat accelerator access as a project constraint, not a guaranteed commodity.