Qualcomm's proposed solution to catch up in AI infra: Bury the compute under the DRAM

Qualcomm detailed its near-memory “high-bandwidth compute” (HBC) architecture for AI250-class Dragonfly datacenter racks, targeting up to 133 TB/s effective memory bandwidth per card and 768GB of memory capacity, but the article flags these as potentially overstated due to “effective” bandwidth assumptions. Qualcomm also claims 18x (AI250 vs AI200) and 54x (AI300 vs AI200) effective bandwidth improvements and argues the approach lowers power/heat by stacking DRAM closer to compute. Separately, Qualcomm announced its planned acquisition of Modular (Mojo/Max software) to reduce reliance on Nvidia CUDA and improve LLM-serving portability, though regulatory risk remains and peak compute metrics were not disclosed.

Analysis

The investable point is not that Qualcomm has a better AI chip; it is trying to carve out the lowest-cost watt-per-token segment of inference, where buyers care more about opex and utilization than peak FLOPS. If the stacked-memory architecture works in production, the first budget line it attacks is serving infrastructure, not training clusters, which makes the revenue opportunity smaller than Nvidia’s but potentially stickier and more recurring.

Relative pressure falls on the marginal serving configurations inside NVDA and AMD, not the core franchises. Nvidia can defend share with software and system-level integration; AMD is more exposed because any credible vendor-neutral runtime reduces the value of its already-fragile software gap. The bigger second-order effect may be that cheaper inference expands total token consumption, so the industry could see mix shift rather than a broad AI capex reset.

The risk is execution: 3D stacking, thermal headroom, yield, and software integration can easily push meaningful revenue contribution out 12-18 months. The clean catalysts are third-party benchmarks on real LLM serving, reference customer wins, and the Modular close; if those do not materialize, this remains a narrative trade. Contrarian view: the market may be underestimating how much cheaper inference increases demand, which would help Qualcomm’s wedge without necessarily hurting Nvidia’s overall earnings power.

Key Decisions for Investors

• Initiate a small starter long QCOM / short AMD pair on any pullback; 6-12 month horizon, targeting a 2:1 reward/risk if Qualcomm gets even one credible production design win. Falsifier: no reference customer, benchmark, or software traction by the next earnings cycle.
• Do not short NVDA outright on this headline. If expressing downside, use limited-risk NVDA or SMH put spreads only into a benchmark or customer-announcement window, because the CUDA moat can absorb isolated inference-share threats.
• Put TSM on watch, not trade: if Qualcomm’s design meaningfully depends on advanced packaging, TSM stays embedded in the value chain even if HBM intensity changes. Thesis breaks if Qualcomm can ship without meaningful advanced packaging or slips materially beyond 2027.
• Consider buying QCOM on confirmation of the Modular close and a third-party serving benchmark, not on the investor-day claims alone. The market is likely undervaluing the software angle more than the chip specs; upside is a platform re-rating if Mojo/Max lowers integration friction.