Qualcomm bets $15 billion that AI chips don't need Nvidia's memory trick

Qualcomm has laid out a plan to grow its data centre chip revenue from around $0.3 billion this financial year to more than $15 billion by fiscal 2029, one of the boldest revenue targets any chipmaker has set for the AI infrastructure race so far.

The plan centres on its AI200 inference accelerator, launching later in 2026, followed by an AI250 chip built around what Qualcomm calls High Bandwidth Compute, or HBC. It's a 3D-stacked, near-memory architecture designed to do the job of High Bandwidth Memory, the specialist chips that Nvidia and AMD's data centre GPUs depend on and that have been in short supply for much of 2026. Qualcomm claims HBC delivers 133 terabytes per second of bandwidth per card, an 18-fold increase over its prior generation.

Microsoft and Meta have already signed on as customers, and Qualcomm says it's building custom silicon for two further unnamed hyperscale cloud providers. It's a serious challenge to Nvidia's grip on AI data centre hardware, not a side project.

Should you care?

This isn't something you'll ever buy directly. But the HBM shortage Qualcomm is trying to design around is the same shortage pushing up prices on the handhelds, laptops and budget phones covered elsewhere in this newsletter. If Qualcomm's approach works at scale, it's one of the more plausible routes to that pressure easing over the next few years.