Chipmaker Qualcomm said it plans to enter the crowded field of artificial intelligence (AI) accelerator chips next year, expanding beyond its best-known businesses of chips for mobile wireless connectivity and smartphone CPUs.
An analyst questioned the 11 percent jump in Qualcomm’s stock price following the announcement, at a time when industry-watchers are increasingly concerned that valuations of AI-related companies have become overinflated.
Qualcomm said it would offer its forthcoming AI200 and AI250 chips both as discrete processors and complete rack-scale server systems using its own CPUs.
Rack-scale system
The AI200 is planned to arrive next year, followed by the AI250 in 2027 and another chip the year after that, matching companies such as Nvidia and AMD that offer AI chip upgrades every year.
Qualcomm said the chips build on its experience deploying its Hexagon neural processing unit (NPU) in Windows PCs.
It said the chips would offer better total cost of ownership than competitors, at a time when data centre operators are working to contain the massive energy costs of training and operating generative AI systems.
The upcoming chips are aimed at inference tasks, rather than training, Qualcomm said.
The AI200 supports 768GB of LPDDR memory per card, while the AI250 will use a more efficient memory architecture that it promised would deliver a more than tenfold improvement in memory bandwidth, along with decreased power consumption.
Qualcomm said the chips would be available as discrete units for companies looking to build their own rack systems.
Customers for data centre components, such as Qualcomm’s CPU, could include competitors Nvidia and AMD, the company speculated.
‘Press release’
In May Qualcomm announced a deal with Saudi Arabia’s Humain, which it said would deploy 200 megawatts’ worth of its AI chips.
Gene Munster, co-founder of Deepwater Asset Management, noted that Qualcomm faces intense competition from market leader Nvidia, AMD and others in the field of AI chips.
“The fact that they want to get into a crowded business I think speaks to how much supply constraint we have around different pieces of inference,” Munster told CNBC’s Squawk Box.
He questioned the substantial stock price jump that followed the announcement, which is similar to other large valuation increases that have accompanied forward-looking AI-related announcements this year – a notable example being a September announcement by Oracle around expected AI-driven cloud revenue that buoyed the value of its shares by more than one-third.
“What we saw today, there is no substance behind this. This is a press release that has unlocked $20 billion (£15bn) in market cap,” Munster said of Qualcomm’s announcement.
“Keep in mind, this is really hard to do what they’re talking about here. AMD talked about this two years ago. They’re on version two of their chips and they really haven’t had that much progress.”
Sky-high valuations
Other industry watchers have become increasingly concerned about the sky-high valuations of companies with exposure to the generative AI industry, particularly following a series of infrastructure-related deals announced by OpenAI in recent weeks.
The deals with Nvidia, AMD, Broadcom and Oracle are worth some $1.5tn, but few details have been provided on how OpenAI plans to pay for them.
The deals were largely put together by OpenAI chief executive Sam Altman and an inner circle of executives, with little input from external advisers, the Financial Times reported on Sunday.
