Amazon ropes Qualcomm into something, something AI, networking chips
- ID
- 22570
- Status
- summarized
- Published
- 09 Sep 2026, 1:29 AM
- Fetched
- 09 Sep 2026, 7:04 AM
- Provider
- The Register
- Category
- technology
- Original URL
- https://www.theregister.com/off-prem/2026/09/08/amazon-ropes-qualcomm-into-something-something-ai-networking-chips/5295071
- Source URL
- https://www.theregister.com/headlines.atom
Summary
- Score
- 4.0
- Created
- 09 Sep 2026, 7:05 AM
- Tags
- Audience
- developersai_ml_learnerssaas_founders
What happened
AWS and Qualcomm announced a multi-generation partnership to develop custom AI inference silicon and 1.6 Tbps optical networking gear, with Amazon's commitments potentially totaling up to $60 billion through September 2036. Qualcomm's AI250-series Dragonfly rack systems, launching next year, replace expensive HBM with cheaper LPDDR memory by shifting compute to the memory's base die, claiming higher effective bandwidth for inference workloads.
Why it matters
If Qualcomm's HBM-substitution approach delivers, AWS inference pricing could eventually drop for workloads running on this custom silicon—but the timeline stretches to 2036 and the article itself notes the deal is 'more buzzwords than compute' with almost no technical detail. Builders should not plan around this yet; watch whether AWS exposes these chips as distinct instance types when Dragonfly ships next year.
Discussion angle
The real story for SEA builders may be the sidebar mention that Thailand paused all datacenter builds and approvals—how does that shift regional cloud capacity and latency expectations for Malaysian startups, and does it make Singapore/Malaysia more or less attractive for hyperscaler investment?