Qualcomm and Amazon Web Services announced a multi-generation partnership in data centre infrastructure on 9 September 2026. Qualcomm will design customised silicon for AWS.
The most striking part of the deal is that it runs both ways: Qualcomm uses Amazon Bedrock while designing those chips.
Inference, not training
The focus is squarely on inference. AWS's reasoning is that what decides inference economics is not raw speed but energy cost per token. Qualcomm's low-power heritage, carried over from mobile processors, fits that metric precisely.
The collaboration is not limited to the compute unit. Optical interconnects at 1.6 terabits of bandwidth are part of the package too. In large models the bottleneck has long been not the chip itself but the data path between chips.
The cost side of inference weighs more heavily on cloud providers every year. You train a model once but rerun it on every request; over time the bill comes from that repeated work, not from training. Efficiency per watt is therefore not a marketing line but a direct question of margin.
The AWS chip shelf gets wider
Qualcomm's designs join the families AWS already runs.
| Family | Job | Origin |
|---|---|---|
| Trainium | Model training | AWS in-house design |
| Graviton | General-purpose server work | AWS in-house design |
| Nitro | Virtualisation and security | AWS in-house design |
| Qualcomm silicon | Inference | External design partner |
The table sums up the AWS strategy. The company has designed its own chips for years, but it does not want to depend on a single supplier, not even its own design team.
That variety is not immediately visible to customers. An AWS customer mostly does not pick which chip runs underneath; they see the price and the latency of the service. The return on this competition shows up in that price.
A question of scale for Qualcomm
Qualcomm's data centre target is 15 billion dollars in revenue by 2029. Amazon is the company's third major data centre customer win since June.
- Qualcomm's revenue still comes mostly from smartphone processors, and that market is not growing.
- The data centre side sits at the centre of the company's attempt to escape that dependence.
- The name across the table is Nvidia, the dominant producer of AI processors.
Investors reacted well and Qualcomm shares rose on the day of the announcement. The company's data centre business is still a small share of total revenue, but it is now where the growth expectation has moved.
How much pressure on Nvidia
Nvidia's strength is not hardware alone but the software stack that settled in over years. Moving an inference workload to another architecture is easier than moving a training workload, but it is not free.
Against that, inference is the workload growing steadily in volume and under the heaviest cost pressure. Large cloud providers shopping for alternatives on that side does not mean the whole market changes hands; it does mean price is under serious pressure for the first time.