Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Nvidia’s claim that its business can sidestep the effects of US controls on the export of semiconductors to China is set to be tested when it publishes quarterly earnings on Tuesday.
Investors have, for now, brushed off fears that Nvidia will suffer after the US tightened restrictions on sales to China of advanced processors that are suitable for developing large artificial intelligence systems. The chipmaker has said the controls would not have a “near-term meaningful impact” on its business.
Since the Biden administration’s announcement in mid-October, Nvidia’s stock market valuation has lost and then regained about $200bn — equivalent to the entire market capitalisation of Intel — hitting an all-time high last week that valued the company at $1.2tn.
Analysts at Morgan Stanley said in a research note last week that the impact of the China restrictions was “the most asked question into earnings” among Nvidia investors.
China accounts for up to 25 per cent of sales in Nvidia’s data centre business but the company confidently asserted last month that it did not yet expect significant fallout from the new US rules.
“Given the strength of demand for our products worldwide, we do not anticipate that the additional restrictions will have a near-term meaningful impact on our financial results,” it said.
The results to be announced on Tuesday cover the three months to the end of October, so there is likely to be little in the earnings themselves reflecting the change in the rules. But investors will look closely at the forward guidance.
Nvidia has attempted to skirt US restrictions on its chip sales in China with new processors that work around the performance restrictions.
Earlier this year it launched a modified version of its flagship H100 processor for Chinese customers, called H800, following the initial US restrictions announced in October 2022. Then, this month, the Financial Times reported that Nvidia had again adapted its products, with a new product for China called H20.
The company has not officially announced the H20 but circulated its specifications to prospective customers in China. The details have left observers conflicted about whether Nvidia has been able to accommodate US requirements while also satisfying Chinese customers’ demand for AI chips.
Analysts and senior managers at two Chinese cloud computing providers said these details suggested the H20 may not provide enough computing power to efficiently train AI systems equivalent to OpenAI’s latest model, GPT-4. The chips would not allow Chinese companies to remain competitive with their US counterparts, these people said — a crucial goal of the Biden administration.
Executives at Tencent and Alibaba suggested last week during their quarterly earnings calls that the Chinese tech groups may not be able to rely on new chips from Nvidia to train their AI models. Both said they planned to increase their focus on domestic alternatives to Nvidia, with Huawei seen as the most likely beneficiary.
Eddie Yongming Wu, Alibaba chief executive, said he expected to see “multiple different chips being used [from] multiple different providers, meeting demand for AI computing power in the China market”.
Charlie Chai, a Shanghai-based analyst at 86Research, said it was “essentially difficult” to use the H20 for training big AI models, restricted by a performance cap on a single chip.
Nvidia has long argued against US controls on the grounds that holding back American companies in China would only fuel the advance of local chipmakers.
Analysts and industry insiders say that it would take years for Chinese tech companies to adapt their systems to use domestic chips instead of Nvidia’s. Nonetheless, this process has been accelerated by the US government’s tightening of export controls.
“Now everyone is forced to adapt the Huawei chips and software,” Chai said, adding that Nvidia “may have underestimated Huawei’s capability”.
Yet not every analyst is convinced that Nvidia’s business will suffer, at least in the next year.
Morgan Stanley said it was “somewhat surprised to report that there is minimal impact near-term” from the new controls, in part because — as Nvidia itself indicated — demand for its AI processors was so far ahead of supply in other parts of the world.
Dylan Patel, chief analyst at chip consultancy SemiAnalysis, said that despite Chinese companies’ public comments, they had ordered a “very significant amount” of H20 chips, which despite its limitations were still more capable than the Nvidia A100 chips that OpenAI used to train previous generations of GPT. He estimated that Chinese customers would spend about $15bn on H20-based systems next year.
“Hundreds of thousands of H20s are going to be manufactured and sold,” Patel said. “It’s not optimal but it is the best chip [for AI development] that China can purchase.”
Nvidia declined to comment.
