Qualcomm Lands AWS Deal to Build Custom AI Chips

Qualcomm announced on Tuesday a multi-generational product collaboration with Amazon Web Services to supply customized silicon at scale for large AI data centres, with the work aimed squarely at inference. The two companies will also co-develop optical connectivity solutions reaching up to 1.6T, built on Qualcomm's SerDes and optical DSP technology. Qualcomm shares rose roughly 10% on the news.

"As AI demand accelerates, data center infrastructure will require advances in both computing and connectivity to deliver greater performance with more efficiency," said Cristiano Amon, Qualcomm's president and chief executive. Prasad Kalyanaraman, the AWS vice president who runs its infrastructure services, framed it as an extension of an existing relationship: "By working together on customized silicon and advanced connectivity, we're delivering more performant, efficient, and cost-effective infrastructure for our customers."

The financial hook is the equity component. Qualcomm issued Amazon a warrant to purchase up to 25 million shares of QCOM common stock — a stake that would be worth several billion dollars at recent prices. The press release itself disclosed neither the exercise price nor any vesting conditions; outlets reading the accompanying filing report a strike near $161 with a ten-year term and vesting tied to the volume of business Amazon actually places, and a figure of up to $60 billion in potential business has circulated in secondary coverage without confirmation from either company. Treat the headline dollar figures as reported rather than announced until the filing language is public.

For Qualcomm this is the first named hyperscaler behind a data-centre push it has been assembling in public for a year. Its June 2026 roadmap covered the Dragonfly C1000 server CPU, built from custom Oryon chiplets with more than 250 cores, and the AI300 inference accelerator; the company also weighed a $10 billion acquisition of Tenstorrent in June to buy its way further up the stack. Winning AWS as a design partner converts a roadmap into volume, and it does so in inference rather than training — the workload where cost per token, not peak FLOPS, decides the purchase.

AWS is not short of silicon of its own. Its Annapurna Labs team designs the Trainium accelerators and Graviton CPUs that already underpin much of its fleet, and it has signed multi-billion-dollar Graviton commitments with customers including Meta. Qualcomm is being added as a merchant partner for customized inference parts and the optics that stitch racks together, not as a replacement for that in-house work. The relationship also runs the other way: Qualcomm says it will lean harder on AWS AI infrastructure, including Amazon Bedrock, for electronic design automation, with the goal of shortening its own chip design cycles.

Strip out the warrant arithmetic and what remains is a second supplier being cultivated for the part of the AI build-out that is about to get enormous. Training runs are concentrated in a handful of labs; inference is every query, every agent step, every one of them billed. Amazon has spent years proving it can design that silicon itself. Paying Qualcomm to design some of it too, and handing over stock to align the incentives, reads less like a vote of no confidence in Annapurna than an admission that no single design team is going to cover the demand curve alone.