Qualcomm, a powerhouse in smartphone chips, has moved to target the artificial intelligence (AI) data center market. It unveiled a new AI chip and a server central processing unit (CPU) that use general-purpose memory for mobile and PCs instead of high-bandwidth memory (HBM), and secured Meta and Microsoft as customers.
On the 24th (local time), according to Reuters and CNBC, Qualcomm held an investor presentation in New York and announced a semiconductor roadmap for AI data centers. Qualcomm unveiled its High Bandwidth Compute (HBC) chip to be deployed in Microsoft's AI data centers and the server CPU "Dragonfly GF C1000."
HBC uses general-purpose memory found in smartphones and laptops instead of HBM. HBM is a key component that boosts AI accelerator performance, but it is expensive and in limited supply. Qualcomm's strategy is to combine relatively inexpensive memory with low-power design technology to reduce AI infrastructure build expense and power burden.
Meta adopted Qualcomm's Dragonfly GF C1000. The product is scheduled for mass production in 2028, and Meta plans to use next-generation chips as well. Qualcomm said it also signed customized chip supply deals with two unnamed hyperscalers. Hyperscalers are cloud corporations that operate large-scale data centers.
Qualcomm set a revenue target of $15 billion for the data center institutional sector in fiscal year 2029. For fiscal year 2027, it expected $5 billion in revenue from this institutional sector alone. It raised the revenue target for non-smartphone businesses, including data centers, from the previous $22 billion to $40 billion.
Cristiano Amon, Qualcomm's chief executive officer (CEO), said the company has assembled a comprehensive portfolio to enter the next phase of the data center market. Investors took it as a signal that Qualcomm aims to reduce its dependence on smartphones and pivot into an AI infrastructure corporations. Qualcomm shares fell 3.3% during regular trading, but rose more than 12% in after-hours trading following the presentation.
Qualcomm also presented a data center strategy bundling AI accelerators, CPUs, memory, and software. It plans to roll out four product lines sequentially over the next 24 months to enter the AI inference market. Inference is the execution phase in which a trained AI model answers user questions or generates images.
It will also strengthen software competitiveness. Qualcomm said it will acquire the AI software startup Modular in a stock transaction worth about $3.9 billion. Modular is a corporations that developed software that helps run AI models across multiple semiconductors. The key is reducing the burden of having to rewrite code for a specific chip.
Modular developed the AI programming language "Mojo" and the AI execution engine "MAX." Qualcomm plans to use Modular's technology as the core of the data center software stack to counter Nvidia's AI development platform "CUDA." CUDA is a software ecosystem that helps develop and run AI models on Nvidia GPUs.
Qualcomm is also eyeing the custom semiconductor market. Reuters reported that Qualcomm is discussing application-specific integrated circuit (ASIC) designs for ByteDance Ltd., the parent company of TikTok. An ASIC is a custom semiconductor designed for a specific customer or use. However, it was reported that it is not yet certain whether the talks will lead to a final design and manufacturing contract.
The market views that, because Qualcomm entered the AI data center market late, which is led by Nvidia and AMD, it will need time to show results. Nvidia dominates the AI infrastructure market with its GPUs and the CUDA ecosystem, and AMD is preparing next-generation server platforms. Big Tech companies such as Amazon, Google, Microsoft, and Meta are also expanding their in-house AI chip development.
Still, as AI data center investment continues to grow, some say opportunities have opened for Qualcomm. With power consumption and token processing expense emerging as key constraints on data center expansion, the industry sees the crux as whether the low-power design and general-purpose memory strategy will translate into actual expense savings.