Qualcomm announced on Monday that it will launch a new artificial intelligence acceleration chip. This move means that Nvidia, which currently dominates the artificial intelligence semiconductor market, will usher in new competitors. As of press time, Qualcomm's U.S. stocks were up nearly 20% intraday.
The launch of artificial intelligence chips marks a change in Qualcomm’s strategic direction. Previously, Qualcomm's business focus has been focused on semiconductor research and development in the fields of wireless communications and mobile devices, and has not been involved in the chip market related to large data centers.
Qualcomm said that the AI200 chip, which will be launched in 2026, and the AI250 chip planned to be launched in 2027, can be installed in a complete liquid-cooled server rack system.
Qualcomm's move is in line with Nvidia and AMD - the latter two companies have launched full-rack GPU (graphics processing unit) systems that can integrate up to 72 chips to achieve "multi-core-in-one" computer functions. Artificial intelligence laboratories need such computing power support when running the most advanced AI models.
Qualcomm's data center chip is developed based on the core artificial intelligence technology in its smartphone chips, namely the "Hexagon Neural Processing Unit" (Hexagon NPU).
"We first hope to prove our strength in other areas. Once we gain a foothold in these areas, it will be much easier to move into the field of data center-level chips." Durga Malladi, general manager of Qualcomm's data center and edge computing business, said in a conference call with reporters last week.
Qualcomm's entry into the data center field means that the fastest-growing market in the technology industry - the new artificial intelligence dedicated server cluster equipment market will usher in a new competitive landscape.
McKinsey predicts that by 2030, global data center-related capital expenditures will be close to US$6.7 trillion, most of which will be invested in systems with artificial intelligence chips as the core.
Currently, the industry is still dominated by Nvidia: its GPU chips account for more than 90% of the market share, and strong chip sales have also pushed Nvidia's market value to exceed US$4.5 trillion. In addition, the chip used by OpenAI to train the GPT series large language model (the core technology behind ChatGPT) comes from NVIDIA.
However, companies represented by OpenAI have begun to seek alternatives. Earlier this month, the startup announced plans to purchase chips from AMD, the world's second-largest GPU maker, and potentially take a stake in AMD. At the same time, technology giants such as Google, Amazon, and Microsoft are also developing exclusive artificial intelligence acceleration chips for their cloud service businesses.
Qualcomm said that the core positioning of its chips is "inference" (that is, running AI models), rather than "training" (referring to the process of developing new AI functions by laboratories such as OpenAI by processing terabytes of data).
The chipmaker said its rack-mounted system will ultimately be cheaper to operate and maintain for customers such as cloud service providers, and the rack system's power consumption of 160 kilowatts is comparable to the high power consumption levels of some Nvidia GPU racks.
Maradi revealed that Qualcomm will also sell its artificial intelligence chips and other related components separately. This model is especially aimed at customers such as hyperscale data center operators - such customers are more inclined to design their own rack systems. He also said that other artificial intelligence chip companies such as Nvidia and AMD may even become customers of some components of Qualcomm's data center (such as central processing units/CPUs).
“We have been working hard to ensure that customers have flexible choices: they can choose the entire system, or they can make their own combinations and combinations,” Maradi said.
Qualcomm has not yet responded to the specific pricing of chips, boards and racks, as well as the number of NPUs that can be carried in a single rack. In May this year, Qualcomm announced a cooperation with Humain Company in Saudi Arabia to provide artificial intelligence inference chips for data centers in the region; Humain Company has confirmed to become a customer of Qualcomm and has promised to deploy enough systems to bring the total power consumption to 200 megawatts.
Qualcomm said that compared with other acceleration chips, its artificial intelligence chips have advantages in three aspects: power consumption, total cost of ownership (referring to the cost of the entire life cycle of the device) and new memory management solutions. In addition, Qualcomm claims that its AI board supports a memory capacity of 768 gigabytes, which is higher than similar products from Nvidia and AMD.