Qualcomm is making a bold play for the artificial intelligence data center market, unveiling a radically new chip architecture designed to bypass the industry’s reliance on expensive High Bandwidth Memory (HBM). The new design, dubbed High Bandwidth Compute (HBC), was showcased at the company’s Investor Day in Manhattan and represents a direct challenge to Nvidia’s dominance in AI hardware.
Crucially, Qualcomm executives confirmed that the new architecture, which will be marketed under the “Dragonfly” brand, is being developed with the Chinese market firmly in mind. By designing compliance with US export controls into the silicon from the ground up, Qualcomm aims to capture a significant share of China’s booming demand for AI compute.
The HBC Architecture: Ditching HBM for LPDDR
The core innovation of the HBC architecture is its near-memory compute design. Instead of relying on HBM, which is currently facing severe global supply shortages and exorbitant pricing, Qualcomm’s design places a stack of low-power double data rate (LPDDR) memory chips directly on top of the logic die.
This vertical integration drastically reduces the physical distance data must travel between the memory and the processor. According to Durga Malladi, Qualcomm’s vice president of data center, this approach yields massive efficiency gains. The HBC architecture reportedly provides six times the bandwidth per watt compared to traditional HBM setups, while simultaneously expanding overall memory capacity and lowering energy costs.
To achieve this, Qualcomm is leveraging its deep expertise in mobile chip design, specifically utilizing through-silicon via (TSV) packaging technology to connect the stacked LPDDR chips to the logic die. “No one has actually done this before,” Malladi stated during the presentation. “It’s not like a science experiment. It’s a commercial product for us.”
The first product to feature the HBC architecture will be the AI250 data center rack, which is slated to begin shipping to customers in fiscal year 2027. The announcement was met with enthusiasm on Wall Street, with Qualcomm’s stock jumping 12 percent following the Investor Day presentation, during which the company also announced a $15 billion revenue target for its AI data center business.
Designing for the Chinese Market
While HBC’s technological innovation is significant, Qualcomm’s strategic focus on China is equally noteworthy. CEO Cristiano Amon was explicit about the company’s intentions. “Everything that we’re building for the data center, you should assume that China is going to be a target market as well,” Amon told investors.
According to Nikkei Asia, Qualcomm is designing specific versions of all four of its new data center product lines tailored for the Chinese market. These chips are being engineered to comply strictly with current US Commerce Department export control regulations, ensuring they can be sold legally to Chinese tech giants and cloud providers.
Amon highlighted Qualcomm’s existing relationships with Chinese original equipment manufacturers (OEMs) in the mobile sector as a key advantage. “That is a strength that we’re going to bring on the data center side,” he noted, suggesting the company plans to leverage its deep ties in Shenzhen and Beijing to rapidly deploy its new AI hardware.
Furthermore, Qualcomm is actively exploring the use of the open-source RISC-V instruction set architecture for its Chinese data center products. Following its acquisition of RISC-V startup Ventana Microsystems last year, Qualcomm is well-positioned to capitalize on Beijing’s push for open-source silicon. “Clearly, in China, there’s a lot of RISC-V interest,” Malladi acknowledged. “We are placing our bets … we’re in the midst of putting together the right kind of solution for that.”
Packaging Challenges and Yield Risks
Despite the promise of the HBC architecture, industry analysts have raised concerns about the manufacturing complexities involved. Stacking multiple LPDDR chips directly onto a logic die using TSV technology is notoriously difficult and prone to packaging defects.
Analysts at Morningstar warned that lower final yields due to packaging errors could “artificially strangle supply” of the new chips, potentially offsetting the cost advantages of ditching HBM. If Qualcomm struggles to manufacture HBC chips at scale with acceptable yields, the product’s commercial viability could be severely compromised.
Interestingly, these potential manufacturing challenges could inadvertently benefit the very memory makers Qualcomm is trying to bypass. If HBC yields are low, the overall supply of AI compute will remain constrained, supporting the high pricing dynamics currently enjoyed by HBM manufacturers like SK Hynix and Samsung. Both South Korean giants, backed by an 800 trillion won ($520 billion) government investment plan, are aggressively expanding their own advanced packaging and memory capabilities.
A New Front in the AI Hardware War
Qualcomm’s entry into the AI data center market with a radically new architecture marks a significant escalation in the battle for compute dominance. By offering a high-bandwidth, energy-efficient alternative to Nvidia’s HBM-reliant GPUs, Qualcomm is providing cloud providers and AI labs with a much-needed second source for advanced hardware.
The company’s explicit focus on China is a calculated risk. While navigating US export controls is complex, the potential reward is massive. Chinese tech giants are desperate for advanced compute, and a compliant, high-performance alternative to Nvidia’s restricted offerings would likely find a highly receptive market. If Qualcomm can overcome the packaging challenges and deliver the AI250 rack on schedule, the Dragonfly architecture could fundamentally reshape the global AI hardware landscape.