Amazon will buy up to 60 billion dollars of Qualcomm AI chips
Key takeaways
- Amazon could buy up to 60 billion dollars of Qualcomm AI data centre hardware over time
- Amazon also gets warrants for as much as 4 billion dollars of Qualcomm stock
- The agreement covers inference silicon and 1.6T optical interconnects, not training chips
- Qualcomm has a public target of 15 billion dollars in data centre revenue by fiscal 2029
Qualcomm signed Amazon as both a customer and an investor on 8 September. The agreement is multi generational, and Amazon could buy up to 60 billion dollars of AI data centre hardware under it. Qualcomm shares jumped around 10 percent.
What the Qualcomm Amazon AI chip deal covers
Custom AI inference silicon and 1.6T optical interconnects. Amazon also receives warrants to acquire as much as 4 billion dollars of Qualcomm stock, which ties the buyer's interests to the supplier's share price in a way a normal supply contract does not.
The word doing the work is inference. This is not training silicon, and Qualcomm is not trying to compete with an H series GPU on raw training throughput.
Why inference is the opening
Inference is now around two thirds of AI compute spend. It is also the part of the workload where a purpose built chip can beat a general purpose GPU on cost per token by a wide margin, because the problem shape is narrow and predictable.
Qualcomm has spent two decades building chips that do a great deal of arithmetic inside a very small power budget. That is the same constraint a data centre operator faces, scaled up and measured in megawatts rather than milliwatts. The skill transfers more cleanly than the company's history in phones would suggest.
Qualcomm has a public target of 15 billion dollars in data centre revenue by fiscal 2029. This single agreement is most of the path there.
The wider pattern
Every hyperscaler is now building or buying custom silicon to get out from under Nvidia pricing. AWS already runs the largest ASIC fleet in production, and adding Qualcomm to that mix gives it a second source for the workload it runs most. We set out how the options compare in our breakdown of GPUs, TPUs, Trainium and custom ASICs.
That matters for everyone else in a specific way. Hyperscalers get to run inference on hardware they part own, at a cost nobody outside sees. Everyone else rents GPUs at list price, against Nvidia's 279 billion dollars in purchase commitments through 2027.
What to watch
Whether Qualcomm names a second hyperscaler. One anchor customer at this size is a dependency, not a business line, and the 2029 revenue target assumes the pattern repeats. Watch the fiscal reporting for the split between phone and data centre revenue, because that is where the transition either shows up or does not. The same question of who can afford to run inference at scale sits underneath OpenAI's projected cash burn through 2030.
