Qualcomm launches new AI chips to challenge Nvidia and AMD…stock soars

Qualcomm announced on the 27th (local time) that it will launch a new artificial intelligence (AI) accelerator that will become a new competitor to Nvidia and AMD. Because of this announcement, Qualcomm (QCOM) shares on the U.S. market were trading at $202 after a 20% surge around 10:30 a.m. Eastern Standard Time.

According to CNBC, Qualcomm said it plans to release AI200 in 2026 and AI250 in 2027, and that both chips could be released as systems that can fill a liquid-cooled server rack.

Qualcomm's AI chips have so far focused on semiconductors for wireless connectivity and mobile devices and have not emphasized chips for data centers, which Nvidia and AMD have focused on.

Nvidia and AMD provide graphics processing units (GPUs) as full-rack systems that can use up to 72 chips like a single computer.

Qualcomm's data-center chips are based on the AI component of Qualcomm smartphone chips called the Hexagon neural processing unit (NPU).

Qualcomm said its AI chips have advantages over other accelerators in terms of power consumption, total cost of ownership, and a new approach to memory handling. Qualcomm said its AI cards support 768 gigabytes (GB) of memory, higher than Nvidia and AMD products.

Durga Malladi, the company's head of data center and edge, said last week, "We first wanted to prove ourselves in other domains and after building our strengths there it was easy to step up to data center level."

With Qualcomm entering the data-center chip field, a new competition has begun in the fastest-growing market in technology: equipment for AI-centric new server farms.

According to McKinsey's estimates, about $6.7 trillion in capital expenditure will be made for data centers by 2030, most of which is expected to be used for AI chip-based systems (about 9,609 trillion won).

Currently, Nvidia dominates the data-center chip market. Nvidia's GPUs account for more than 90% of the market, with the rest held by AMD and others.

Nvidia surpassed a $4.5 trillion market capitalization through GPU sales. Nvidia's chips were used to train OpenAI's large language model for ChatGPT.

However, companies like OpenAI have been seeking alternatives, and earlier this month OpenAI announced plans to acquire chips from AMD, the second-largest GPU supplier, and to acquire a stake in that company. Cloud service companies like Google, Amazon, and Microsoft are also developing their own AI accelerators.

Qualcomm said its chips focus on inference rather than training, i.e., running AI models. This is the way places like OpenAI handle terabyte-scale data to create new AI features.

Qualcomm said its rack-scale system will ultimately reduce operating costs for customers such as cloud service providers, and that it uses 160 kilowatts per rack. This is similar to the high power consumption of some Nvidia GPU racks.

Malladi said Qualcomm will sell AI chips and other components separately, particularly to customers like hyperscalers who prefer their own rack designs. He added that other AI chip companies like Nvidia or AMD could become customers for Qualcomm's data center components such as its central processing unit (CPU).

Qualcomm declined to comment on the price of chips, cards, or racks, and the number of NPUs that can be mounted in a single rack. Qualcomm announced in May that it had formed a partnership with Humane of Saudi Arabia to supply AI inference chips to data centers in the region. Humane became a Qualcomm customer and plans to build systems that can use up to 200 megawatts of power.

Kim Jeong-ah guest reporter kja@hankyung.com