Last year, NVIDIA’s GB200 Grace Blackwell Superchip artificial intelligence chip was marketed as the most powerful artificial intelligence (AI) training chip on the market. It combines two NVIDIA B200 GPUs with a single NVIDIA Grace CPU, allowing it to train a model with 1 trillion parameters.
Today, The Information reported that Microsoft, Google, Meta, and Amazon are among customers who have held back GB200 purchases due to overheating issues and a bug in the chip’s connection to the browser rack it uses. The issue has been known about since August, with NVIDIA saying it has been resolved.
But the delays that occurred while the issues were resolved caused customers to hold off on buying them in the first place. OpenAI, for example, opted to buy the older H100 chip rather than wait. NVIDIA reportedly received $10 billion in GB200 orders from customers after the chip launched in March last year.
The report comes on the same day the United States announced a ban on the sale of high-powered AI chips to countries that are not close allies, such as Malaysia, Singapore, and Israel.