Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Amazon’s cloud computing arm has halted orders of Nvidia’s most advanced “superchip” to wait for a more powerful new model, as investors fret about a dip in demand between the $2.3tn chipmaker’s product cycles.
The Silicon Valley-based chipmaker unveiled a new generation of processors dubbed Blackwell in March, barely a year after its predecessor Hopper began to be shipped to customers. Nvidia’s chief executive Jensen Huang said the new products would be twice as powerful for training large language models, the technology behind OpenAI’s ChatGPT.
Amazon Web Services, the world’s largest cloud computing provider, told the Financial Times that it had “fully transitioned” its previous orders for Nvidia’s Grace Hopper superchip, which was launched in August, and replaced them with its successor Grace Blackwell.
The company said the move “made sense” given “that the window between Grace Hopper and Grace Blackwell was small”.
Nvidia declined to comment ahead of its quarterly earnings report on Wednesday, citing quiet period rules.
Analysts expect the chipmaker to report that sales tripled in its most recent quarter, driven by a frenzy of spending by Big Tech companies on artificial intelligence technology. But some investors have begun to question how long it can maintain its extraordinary growth streak.
While Big Tech companies have pledged to continue to invest tens of billions of dollars into building out data centre infrastructure for AI this year, “there is anxiety [on Wall Street] about a pause in front of Blackwell”, analysts at Morgan Stanley wrote in a note to clients this week.
Production of the new Blackwell chips will ramp up over the course of this year. Analysts expect them to be delivered in the fourth quarter. In the interim, analysts at Citi pointed to a “potential air pocket” in demand for AI chips, after last year’s long wait times for Nvidia’s chips dissipated.
Chips based on Nvidia’s Hopper architecture, such as its sought-after H100 graphics processing units, went into full production in September 2022. The Grace Hopper superchip, also known as GH200, packages several H100 GPUs alongside high-speed memory, connectivity and a central processing unit.
Its successor, the GB200, is the first product to take advantage of Blackwell.
Neither Amazon nor Nvidia would confirm the value of the order. Analysts at HSBC have estimated a GB200 chip, which includes two B100 chips, will cost as much as $70,000, with the price of an entire server carrying the new technology running up to $3mn.
AWS continues to offer other Nvidia chips including H100s to its cloud customers. But, as one of Nvidia’s largest customers, the AWS move is likely to concern investors already worried that tech companies would delay purchases as they await Blackwell’s release.
For much of last year, demand for Nvidia’s H100 chips vastly outstripped supply as the launch of OpenAI’s breakthrough ChatGPT triggered a wave of investment in AI infrastructure by cloud and internet companies, start-ups and corporate buyers.
But since the start of 2024, the long waits for delivery of H100 chips have eased.
Although Nvidia’s stock has almost doubled in value since the beginning of the year on confidence among investors about demand for AI chips, the shares have struggled to show a consistent increase since the chipmaker’s GPU Technology Conference in March where the Blackwell was unveiled.
Nvidia has battled in the past to manage supply and demand smoothly between product upgrades. Recent surges in demand for chips suitable for video games and cryptocurrency mining during the Covid-19 pandemic were followed soon after by a glut of GPUs.
Analysts at Morgan Stanley said they were confident that, even if the supply of Nvidia’s last-generation chips was more readily available now, “in the interim we are seeing new cloud, enterprise and sovereign customers taking all of the available Hopper supply”.
