Qualcomm enters the AI data center with inference chips, stock jumps 11%

Qualcomm announced Monday that it will release two new data center AI accelerator chips, the AI200 and AI250, sending its stock up 11% on the news. The move puts the mobile-chip giant in direct competition with Nvidia, which holds over 90% of the AI chip market, and AMD, the distant second-place player.

The announcement is not just another entrant into a crowded market. It is a bet on a specific architectural thesis: that inference, not training, will be where the volume and economics of AI compute shift in the coming years. Qualcomm is not trying to beat Nvidia at the training game. It is trying to own the part of the market that runs models after they are built.

Qualcomm’s AI200 will go on sale in 2026. The AI250 is planned for 2027. Both chips come in a full-rack, liquid-cooled system that can link up to 72 chips as a single computer, matching the form factor and scale of Nvidia’s DGX and AMD’s Instinct racks. The chips are based on the Hexagon neural processing units (NPUs) that Qualcomm already ships in its smartphone SoCs. Durga Malladi, Qualcomm’s general manager for data center and edge, told reporters that the company wanted to “prove ourselves in other domains” before moving into the data center.

The timing is strategic. Nearly $6.7 trillion in capital expenditures will flow into data centers through 2030, according to a McKinsey estimate cited in the report. Most of that spending will go to AI infrastructure. But the nature of that infrastructure is changing. Training the largest frontier models requires clusters of tens of thousands of GPUs running for months. Inference, by contrast, is a continuous, distributed workload that runs every time a user queries a chatbot, generates an image, or invokes an agent.

Qualcomm is making a specific architectural claim. Its AI cards support 768 gigabytes of memory, which the company says is higher than comparable offerings from Nvidia and AMD. The rack system draws 160 kilowatts, comparable to Nvidia’s high-power GPU racks. Qualcomm says its chips will ultimately cost less to operate for cloud service providers, citing advantages in power consumption, cost of ownership, and a new approach to memory handling.

The company is also offering flexibility. Malladi said Qualcomm will sell its AI chips and other parts separately, especially for hyperscalers that prefer to design their own racks. He even suggested that Nvidia or AMD could become clients for Qualcomm’s data center parts, such as its CPU. “What we have tried to do is make sure that our customers are in a position to either take all of it or say, ‘I’m going to mix and match,’” Malladi said.

That is a different sales strategy from Nvidia’s. Nvidia sells the full stack: chips, boards, networking, software, and the CUDA ecosystem that locks customers in. Qualcomm is offering an alternative path: buy the NPU, buy the CPU, design your own system. It is a bet that the hyperscalers — Google, Amazon, Microsoft — who are already building their own AI accelerators (TPUs, Trainium, Maia) will want more modular options, not less.

The market is already moving in that direction. OpenAI announced earlier this month that it plans to buy chips from AMD and potentially take a stake in the company. Google, Amazon, and Microsoft are developing their own accelerators. The hyperscalers want alternatives to Nvidia, and they want leverage. Qualcomm is offering itself as another option.

But the real question is whether Qualcomm’s NPU architecture, designed for the power and thermal constraints of a smartphone, can scale to data center workloads. Nvidia’s GPUs are general-purpose parallel processors that excel at both training and inference. AMD’s Instinct line uses a similar GPU architecture. Qualcomm’s NPU is a more specialized design, optimized for the matrix operations that dominate neural network inference. That specialization could be an advantage in power efficiency, but it could also limit flexibility as model architectures evolve.

The company declined to comment on pricing, the number of NPUs per rack, or the exact performance specifications. It did announce a partnership with Saudi Arabia’s Humain in May to supply data centers in the region with AI inferencing chips. Humain committed to deploying up to as many systems as can use 200 megawatts of power.

For AI builders, the entry of Qualcomm into the data center market is a signal that inference is becoming the dominant compute workload. Training the next GPT-class model will remain a capital-intensive, Nvidia-dominated affair. But running that model billions of times a day is a different problem, with different economics. Qualcomm is betting that the inference market will be large enough, and competitive enough, that a mobile-chip company with a specialized NPU can carve out a profitable niche.

The stock market liked the story. The real test comes in 2026, when the AI200 ships and the industry gets to see whether Qualcomm’s smartphone heritage translates into data center performance.