Highlights:

  • Many generative AI applications in use today are powered by Nvidia’s wildly successful and in-demand H100 chips, which the Blackwell GPUs replace.
  • If the delay is real, it is anticipated to have a major impact on Nvidia’s clientele, many of whom have early next year plans to launch new AI data centers.

Nvidia Corp.’s Blackwell B200 GPUs will be delayed by at least three months. The chipmaker informed Meta Platforms Inc., Microsoft Corp., Google LLC, and other of its customers.

One of Nvidia’s major manufacturing partners has found a design issue that will prevent the chips from releasing on schedule. Anonymous sources disclosed that the production process found the problem unusually late.

Many generative AI applications in use today are powered by Nvidia’s wildly successful and in-demand H100 chips, which the Blackwell GPUs replace. They were unveiled in March and promise to reduce energy usage by up to 25% on certain workloads while providing a performance gain of up to 30 times over the H100.

The news that Nvidia’s Blackwell-based products will be available from its partners later this year was announced just three months ago, so the delay was unanticipated. They were scheduled to be the first significant release in a series of AI chip improvements from semiconductor companies, with competitors like Advanced Micro Devices Inc. scheduled to follow suit with their own products in the months ahead.

Reportedly, the issue pertains to the CPU die that joins two Blackwell GPUs on a single Nvidia GB200 Superchip—the first product line to use the new GPUs. The issue was reportedly discovered by Taiwan Semiconductor Manufacturing Co. (TSMC), which is responsible for the mass production of chips for Nvidia. The revelation reportedly compelled Nvidia to alter the die’s design. Before it can begin mass-producing the chips as intended, it will need to work with TSMC to conduct production tests for a few months.

If the delay is real, it is anticipated to have a major impact on Nvidia’s clientele, many of which have early next year plans to launch new AI data centers. For instance, it’s estimated that Google ordered around 400,000 GB200 chips in a transaction worth over USD ten billion. While Microsoft hopes to have 55,000 to 65,000 GB200 chips ready for OpenAI by the first quarter of next year, Meta has also made an order akin to this.

Nvidia is considering making the Blackwell chip available as a single GPU to compensate for the delay and fill some of its first orders.

It makes more sense for Nvidia to postpone the release of the Blackwell GPUs than to distribute potentially defective products. The industry has experience with such a move. Due to a simple error, AMD stated last month that it would be delaying the release of its Ryzen 9000 central processor units. Delays are preferable to disastrous mishaps, like the one that seems to have befallen Intel Corp.’s 13th and 14th-generation Core CPUs.

It is even significant that Nvidia makes the proper product because of how much its Blackwell GPUs are probably going to cost. The AI GB200 Grace Blackwell Superchips are reportedly going to cost as much as USD 70,000 each, while a whole server rack would set you back over USD three million. According to reports, Nvidia plans to sell between 60,000 and 70,000 entire servers; therefore, defects like the one discovered in Intel’s chips would be extremely costly and detrimental to the company’s standing.

Nvidia can most likely afford the delay in any event. The business is the clear industry leader for AI chips; according to several estimates, it controls up to 90% of the world’s market for these chips.

While competitors like AMD and Intel have created their own AI chips, the industry has yet to see any real adoption of them.

“For cloud providers, this will cause headaches in terms of delayed revenue,” the analyst added.

Mueller stated that since AI businesses are the end users, they will bear the brunt of this rather than the cloud corporations who will purchase the GB200 processors. He observed that many of them had planned their product roadmaps and development plans based on the expectation that the Blackwell chips would be available later this year. The analyst thinks that several initiatives will suffer significantly because of Blackwell’s possible arrival in early 2025.

“Overall, the pace of generative AI development may hit a plateau for a while, but this can be good news for enterprises as it means they’ll get a breather and the opportunity to catch up and think about where the technology can really make a difference,” Mueller stated. “The real winner may actually be Google, which runs all of its native AI workloads on its own TPUs. It might give Gemini a chance to get ahead.”

A spokesperson for Nvidia said, “Hopper demand is very strong, broad Blackwell sampling has started, and production is on track to ramp in the second half.”

This delay is Nvidia’s second setback in a short period of time. It was disclosed last week that the chipmaker is the focus of two inquiries into its AI operations by the US Department of Justice. Nvidia’s USD 700 million acquisition of the Israeli AI startup Run:AI Inc. is reportedly the subject of two investigations. The first examines potential antitrust violations, while the second looks at allegations that Nvidia compelled cloud computing companies to purchase its chips.