One of the biggest stories this week in the HPC-AI world involves — surprise! — NVIDIA. Actually, make that two or three of the biggest stories. One is NVIDIA’s stellar quarterly earnings announcement in the face of concerns in recent months about disruption of chip exports of GPUs to China.
The company shrugged off what it said a few weeks ago would be a $5 billion charge and charged ahead with revenues ($44.1 billion — including $39.1 in the data center market — up from $26 billion a year ago), and earnings ($18.8 billion, up 26 percent YoY).
That the company mostly exceeded expectations didn’t come as a complete surprise considering the enormous and increasing demand for AI compute to supply mammoth AI factories being stood up around the world, including OpenAI’s AI data center planned for Abilene, TX, which is slated to house up to 400,000 NVIDIA top-of-the-line Blackwell chips.
Still, challenges exist for NVIDIA, one of which is the China AI market, which has become increasingly walled off to U.S. technology vendors by the U.S. government. In an interview last night with CNBC’s Jim Cramer, NVIDIA CEO Jensen Huang made the case for continued sales of his company’s GPUs to China, where he said 50 percent of AI developers reside. And he wants them to develop their AI applications on NVIDIAs CUDA development platform to run on NVIDIA hardware.
“American technology stacks will run AI the best all over the world, and so this is the most important strategic reason to be in China, because there are so many developers there,” Huang said. “And because the world is going to adopt technology from one country or another, then we prefer to be it at the American technology stack.”
Earlier this year, exports to China of NVIDIA’s pared-down H20 AI GPU, previously allowed under U.S. trade rules, were closed to exports by the Trump White House, forcing NVIDIA to take the $5 billion charge. This raises another major story emerging this week, that NVIDIA and AMD are working on developing AI chips for the China AI market that comply with U.S. trade restrictions.
According to a story in DigiTimes, the two companies have “quickly adjusted their designs to reduced specifications, and it is expected that starting in July 2025, they will launch a new wave of downgraded, but compliant, AI GPUs that can be sold to China.”
According to the story, NVIDIA has tentatively named its China AI chip the B20, while AMD is moving ahead with the Radeon AI PRO R9700 and other products. “Both can support the operation of models such as DeepSeek,” DigiTimes reported.
NVIDIA’s expected B20 adopts the Blackwell architecture but with major alterations in memory and computing power performance, according to DigiTimes. AMD’s Radeon AI PRO R9700 and other products are designed for AI workstations to accelerate local inference, model fine-tuning and other data-intensive workflows, according to the publication.
As of 2021, NVIDIA’s market share in China was approximately 90 percent, but it has now dropped to 50 percent.
