Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Nvidia and its main supplier Taiwan Semiconductor Manufacturing Company are facing production challenges with the highly anticipated next generation of its most powerful artificial intelligence chips, threatening to delay shipments planned for this year.
Nvidia’s cutting-edge designs, which take advantage of a new TSMC manufacturing process, have caused complications with certain models in the forthcoming Blackwell family of data centre chips as it prepares for mass production, according to people familiar with the situation.
Amid a broader stock market sell-off on Monday, shares in Nvidia fell as much as 15 per cent in early trading in New York, while TSMC closed down 10 per cent on the Taiwan stock exchange, after The Information first reported the potential delays. By 12pm in New York, Nvidia’s fall had moderated to 6 per cent.
Customers including Microsoft, Google, Meta and Amazon, as well as AI start-ups such as OpenAI, have been lining up to buy Nvidia’s latest chips to build the next generation of their AI systems.
Big technology companies are investing tens of billions of dollars each quarter on AI infrastructure and have pledged to increase capital expenditures further in the coming months. Some analysts now project $1tn of spending on data centres to fuel AI over the next five years.
However, anxiety on Wall Street about the sustainability of the AI boom has been growing in recent weeks. Nvidia’s market value has fallen by about $750bn since it briefly became the world’s most valuable company in mid-June.
Hedge fund Elliott Management told investors in a recent letter, seen by the Financial Times, that it believed Nvidia and other large technology stocks were in “bubble land” and that AI was “overhyped with many applications not ready for prime time”.
When he unveiled the new chips in March, Nvidia chief executive Jensen Huang said Blackwell would be twice as powerful for training AI models as its predecessor, Hopper.
He told investors and analysts in May during a quarterly earnings call that Nvidia would see “a lot of Blackwell revenue this year” and said last week that it had begun to ship engineering samples.
However, one person familiar with the manufacturing process said there were “difficulties” with moving towards mass production of the Blackwell chip. “Those are related to the interposer,” the person said. The interposer is a layer connecting the different dies packaged together in the complex chips needed for AI applications.
Nvidia declined to comment but reiterated that “Blackwell sampling has started and production is on track to ramp” up to mass production, in the second half of 2024. Demand for its existing Hopper chips remains “very strong”, Nvidia added.
TSMC did not respond to a request for comment.
The problems highlight the immense engineering challenges of packing the power needed for the newest AI chips into a limited space, and are likely to further exacerbate the capacity bottleneck in advanced packaging, the final stage in producing the most sophisticated chips.
TSMC, which is the world’s largest chip manufacturer and Nvidia’s sole manufacturing partner, has struggled over the past year to increase capacity of its most advanced production technology fast enough to meet AI chip demand.
When the company announced its second-quarter results last month, chief executive CC Wei said the company would not be able to balance demand and supply by the end of this year as previously aimed. TSMC now hoped to achieve this “sometime in 2025 or 2026”, Wei said.
Mark Li, a semiconductor analyst at Bernstein, said Nvidia would likely have to make a minor design tweak to address the problem.
BNP Paribas analysts said that fixing problems of this kind would typically take two or three months. But they did not expect the delay to change “the mid- to long-term story for Nvidia or AI adoption”, although it may be positive news for Nvidia’s closest competitor, AMD.
Analysts at Citi said in a note to clients that the delay could reduce Nvidia’s data centre revenues in the quarter ending in January by as much as 15 per cent, although sales could be higher in the subsequent period as a result.
Read the full article here