Qualcomm is innovating with HBC to solve the inference memory wall but placing the memory on top of the compute die. QUALCOMM
Qualcomm has reached a major milestone in the development of the memory technology which underpins their entire data center strategy revealed at the company’s Investor Day event in June: High Bandwidth Compute, or HBC.
It is unusual for a company to announce an internal development benchmark, but the industry has been skeptical of Qualcomm’s multi-year roadmap into the datacenter of the AI world, so this will be seen as positive news. The strategy is to focus on servers used in AI inference, power efficiency, high-bandwidth memory, and custom hyperscaler silicon—not on competing head-on in GPU training. The company is packaging these components under the new Dragonfly brand, with a stated goal of more than $15 billion in annual data-center revenue by fiscal 2029.
The AI landscape, once dominated by compute efficiency, is now in the era of being memory limited. Hence the term, the memory wall.
Discussing this goal with ex-Alphawave Semi CEO Tony Pialis (now EVP and GM of the data center segment at Qualcomm) at the Company’s recent Snapdragon Summit, Mr. Pialis asserted that he sees so much demand for Dragonfly at major AI players that he expects to meet or exceed that goal, even if he managed to “massively screw this up”. After all, the $15B goal is less that 10% of Nvidia’s 2026 data center revenue, and Qualcomm has already announced deals with major players. (Note that Qualcomm is a client of my company, Cambrian-AI Research.)
HBC is the Cornerstone of Qualcomm’s Data Center Strategy
The underlying thesis here is that inference increasingly becomes constrained by memory movement, latency, and data-center power—not simply raw compute. Solving this problem, ideally first and better than the competition, would launch Qualcomm Qualcomm’s High Bandwidth Compute (HBC) architecture is meant to put compute and memory closer together, while its connectivity portfolio extends from copper to optical links for rack- and cluster-scale systems. The HBC memory approach will be deployed on both AI accelerators (AI250/300) and CPUs (C1000 and Snapdragon).
Qualcomm believes HBC will increase capacity per watt by 200X over SRAM like that found in Groq. QUALCOMM
Target Customers Like What Qualcomm is Showing Them
While still about a year out, Qualcomm has already been able to secure some major wins with its roadmap:
- Meta: A multigenerational agreement to deploy the Dragonfly C1000 CPUs in Meta’s next-generation server fleet, with production expected to begin in 2028.
- Amazon Web Services: A multigenerational custom-silicon and optical-networking agreement; Qualcomm said associated commercial processors were already in production, with revenue expected to begin in fiscal Q1 2027.
- Two additional hyperscaler custom-silicon wins: Qualcomm disclosed two large, unnamed hyperscaler designs, with meaningful revenue expected to start in fiscal 2027.
- Microsoft/Azure signal: Microsoft leadership indicated Azure deployment interest around Qualcomm’s HBC accelerator, though Qualcomm’s most explicit named CPU customer at launch was Meta.
- Capability-building acquisitions: Qualcomm acquired Alphawave Semi for high-speed wired/optical connectivity and Modular to strengthen a hardware-neutral AI software layer—important pieces for selling complete data-center platforms rather than isolated chips.
As a reminder, Qualcomms roadmap includes connectivity, custom silicon, AI accelerators, and Dragonfly CPUs. QUALCOMM
Demonstrating and Measuring Performance
Qualcomm’s differentiation claim is tokens per watt and total cost of ownership for inference. The company markets Dragonfly as delivering up to 8x better tokens per watt than GPUs, though this is a Qualcomm performance claim and will need independent workload-by-workload validation. The current testing now being disclosed measured 18X better memory bandwidth (TB/S) vs. traditionally-attached LPDDR5 and 5X better efficiency (TB/S/W) than HBM.
The Practical Takeaway
Qualcomm now has a credible, multi-product path—CPU, inference accelerators, memory architecture, networking, custom ASICs, and software—with initial hyperscaler endorsements. The main execution test is whether those design wins convert into sustained volume vs. sampling, and whether Qualcomm’s efficiency advantage holds against Nvidia, AMD, and internal hyperscaler silicon.
Every AI accelerator company is working to solve the memory wall, so Qualcomm needs to double down and meet schedule milestones to secure an important advantagein iots sentry into the datacenter. Some say Qualcomm is too late to the data center party, but if the current industry demand projections hold true, there is remains massive demand for a solution that solves the memory wall that Qualcomm could be able to convert to become the largest segment in Qualcomm’s portfolio.
Disclosures: This article expresses the author’s opinions and should not be taken as advice to purchase from or invest in the companies mentioned. Cambrian-AI Research is fortunate to have had many, if not most, semiconductor firms as our clients, including Blaize, BrainChip, Cadence Design, Cerebras, D-Matrix, Eliyan, Esperanto, Flex, GML, Groq, IBM, Intel, Micron, NVIDIA, Qualcomm Technologies, Si-Five, SiMa.ai, Synopsys, Ventana Microsystems, Tenstorrent and scores of investment clients. We have no investment positions in any of the companies mentioned in this article except Nvidia, where we are long.