DigiTimes 分析称高通凭 Dragonfly 平台双线进军云端 AI 加速与定制硅，长期对标英伟达/博通及联发科/世芯/创意

Qualcomm has historically provided standard chips primarily through its Snapdragon product line, serving a broad range of end customers. After officially launching its Dragonfly data center product line in June 2026, the company expanded its business to include multiple hyperscalers.

Compared with the consumer market, hyperscalers place greater emphasis on workload optimization, multigenerational product planning, and system integration capabilities. Qualcomm is therefore entering the market as an AI accelerator, high-speed interconnect, and custom silicon provider, while gradually expanding into CPUs, memory, software stacks, and rack-scale systems. Its data center strategy is shifting from supplying individual chips toward developing a complete AI infrastructure platform.
DIGITIMES believes that as Qualcomm increases its involvement at the system level, its competitive scope will expand from computing chip vendors such as Nvidia and AMD to custom silicon providers including Broadcom, Marvell, and MediaTek, as well as cloud ASIC design service providers such as Alchip Technologies and Global Unichip. This indicates that Qualcomm is joining the competition in the cloud data center market with both AI computing platforms and custom silicon offerings.
AI200/250 lays the foundation
At an analyst roundtable held during the Snapdragon Summit 2026, Qualcomm said its data center AI accelerators have different objectives at different stages.
The AI100 is primarily intended to help the company enter the hyperscaler market and establish initial partnerships. The AI200/250 will be used to pursue larger-scale deployments and introduce high-bandwidth compute (HBC), while the AI300 represents the product generation that Qualcomm is truly targeting for gigawatt-scale AI infrastructure deployments.
This shows that Qualcomm's current focus is to use the AI200/250 to open up the data center market; establish a track record in software compatibility, system deployment, and mass production; and then secure co-design opportunities for customers' next-generation products.
For hyperscalers, collaboration experience with previous generations of products often influences next-generation architectures and supplier selection. As a result, Qualcomm's ability to remain part of customers' product roadmaps and participate in defining next-generation products will be more strategically significant than shipments from any single generation.
In fact, as AI-related chips continue to diversify, the boundary between standard products and custom silicon in AI computing platforms will gradually blur. Hyperscalers may not need to develop chips entirely from scratch. Instead, they may adopt existing computing, memory, and high-speed interconnect IP and then make extensive adjustments based on AI models, power consumption, and system requirements.
Qualcomm is pursuing both standard products and custom silicon precisely to use reusable platform IP as a foundation, thereby increasing customization flexibility and its level of participation in the architectures of large customers.
The recent rise of agentic AI has also renewed attention on the role of CPUs. Qualcomm has planned different products, including head-node CPUs and agentic CPUs, with its first-generation CPU product, the C1000, expected to enter mass production in 2028.
DIGITIMES believes that as agent workloads shift from one-time model inference toward long-running, multi-stage tasks, CPUs may take on a greater role in task orchestration, memory management, and heterogeneous computing coordination. Data center CPU products are also likely to evolve toward multiple SKUs differentiated by workload.
Qualcomm also emphasized that CPUs, XPUs, HBC, interconnects, optics, and modular software stacks will serve as common foundational technologies within the Dragonfly data center platform strategy. Once the C1000 enters mass production, Qualcomm will fill out its in-house CPU capabilities and combine them with its existing XPU and high-speed interconnect products to create a more complete rack-scale computing and interconnect architecture.
Notably, Qualcomm is also elevating high-speed interconnects beyond data transmission to make them part of the compute fabric, enabling computing, memory, and networking to be jointly optimized at the rack and cluster levels. As these core capabilities gradually come online, Qualcomm's system involvement will expand from individual chips to these levels.
However, the key to advancing its strategy will be whether these multiple platform technologies can form actual deployments within the same hyperscaler architecture and enter customers' long-term AI infrastructure plans across product generations. Only if it can continue expanding its adoption will Qualcomm have a chance to move beyond being a chip supplier and become an AI infrastructure partner.
Qualcomm's dual-track strategy
However, in addition to providing design implementation, tape-out, and mass production services, custom silicon providers are increasingly integrating their own IP in areas such as CPUs, memory, SerDes, chiplets, and software while participating deeply in the definition of customers' products and systems. If these reusable key IP components can be incorporated into customers' product architectures across generations, suppliers will have greater opportunities to increase their technological stickiness and share of value.
Qualcomm has clearly chosen a custom silicon supply model, which is also relatively similar to those of Broadcom and Marvell. At the same time, Qualcomm continues to expand its data center platform capabilities and extend its presence into complete AI infrastructure, bringing it into competition with complete AI infrastructure platform providers such as Nvidia.
In other words, Qualcomm is deepening its co-design relationships with hyperscalers through its custom silicon supply model while increasing the contribution of its proprietary IP and system integration capabilities to expand its share of the overall AI platform value.
As AI chips expand from individual dies to chiplets, advanced packaging, optoelectronic interconnects, and rack-scale systems, the capabilities required of chip suppliers are also expanding from chip design to IP portfolios, system architecture, supply chain integration, and capacity allocation.
For Taiwan-based companies such as MediaTek, Global Unichip, and Alchip, Qualcomm's strategy shows that future competition in custom silicon will depend not only on design execution capabilities, but also on the ability to integrate diverse IP and increase participation in defining customers' products and system architectures.