Table of Contents
The Qualcomm deal covers custom inference silicon and 1.6T optical links
In sum – what we know:
- A multi-generation deal – Qualcomm will supply customized AI-inference silicon and 1.6T optical connectivity for AWS data centers, though architecture, process node, foundry, quantities and deployment timing are all undisclosed.
- A milestone-vested warrant – Amazon holds a warrant for up to 25 million Qualcomm shares at $161.26, with only 3.75 million vested on issuance and the remaining 21.25 million tied to qualifying purchases of up to $60 billion by 2036.
- A diversification bet – The deal is high-profile validation of Qualcomm’s push beyond mobile, but it arrives late against Nvidia, AMD, Intel, and AWS’s own Trainium and Inferentia programs.
Qualcomm and Amazon have announced a major partnership on September 8, 2026 that will see Qualcomm deploying multiple generations of custom AI-inference silicon and optical interconnects for AWS data centers. It’s a pretty major deal for Qualcomm, which is newer to large-scale AI inference in hyperscale data centers than competitors out there.
The two companies aren’t strangers though. AWS has offered EC2 instances powered by Qualcomm’s AI 100 accelerators since late 2023. But this arrangement is a different animal — made up of bespoke server silicon, optical connectivity, and a multi-generation commitment aimed squarely at AWS’s large-scale AI infrastructure.
Scope of the deal
According to the two companies, the deal includes working together on customized silicon “at scale” for large AI data centers, with the workload emphasis on AI inference. Neither side disclosed the architecture, process node, foundry, quantities, deployment schedule, or any performance and power benchmarks. That’s a lot of blanks. The deal is mostly focused on inference — Nvidia remains the dominant supplier of AI GPUs and full systems, and training in particular is still an overwhelmingly GPU-driven business.
The collaboration also covers high-performance optical connectivity reaching up to 1.6T, or 1.6 terabits per second, plus future-generation solutions. Qualcomm says the connectivity work will lean on its SerDes technology — the serializer/deserializer circuits that convert data between parallel and serial forms for high-speed links. That’s along with its optical DSPs, the digital signal processors that clean up and manage signals traveling over optical fiber. AI clusters live and die by how fast data moves among compute, networking, storage, and memory, so a connectivity play alongside the compute play makes sense. The release doesn’t say which AWS regions, products, or systems will actually use any of it.
There’s a reciprocal element too. Qualcomm plans to expand its use of AWS AI infrastructure, including Amazon Bedrock, for electronic design automation, with the goal of shortening its design cycles. No economics or contract duration were disclosed there either.
Notably absent from all of this is any suggestion that AWS is stepping back from its own silicon. The Trainium, Inferentia, and Graviton programs continue unchanged, and AWS announced no customer-facing EC2 product tied to the Qualcomm work.
How the warrant and financial terms work
Qualcomm issued a warrant to Amazon.com NV Investment Holdings LLC, an Amazon affiliate, for up to 25 million Qualcomm common shares at an exercise price of $161.26 per share. At that fixed strike, the maximum exercise value works out to about $4.03 billion. To be clear, that’s the value of shares Amazon could buy at a set price — not a cash equity investment in Qualcomm.
The warrant vests over time rather than upfront: 3.75 million shares vested on issuance based on initial purchase commitments, and the remaining 21.25 million vest in tranches tied to Amazon’s executed commercial arrangements and binding purchase orders — up to $60 billion in qualifying payments over the warrant’s 10-year term, which runs to September 3, 2036, per the Form 8-K Qualcomm filed September 3, 2026. That $60 billion is a ceiling on payments that can count toward vesting, not an order, backlog, or guaranteed revenue.
Both companies are framing this as an extension of an existing relationship. Qualcomm CEO Cristiano Amon said AI infrastructure needs advances in both computing and connectivity to deliver more performance efficiently, and that Qualcomm would bring its power-efficient compute experience to customized silicon and connectivity. AWS vice president Prasad Kalyanaraman said the collaboration builds on the existing partnership and aims to deliver more performant, efficient, and cost-effective infrastructure for customers.
A big deal for Qualcomm
For Qualcomm, this is high-profile validation of a diversification strategy the company really wants to work. Its mobile business faces long-term pressure, particularly as more companies bring their silicon efforts in-house. Data-center infrastructure is the obvious place to look for growth, and Qualcomm laid out its ambitions earlier this year with its Dragonfly roadmap — with the Qualcomm Dragonfly C1000 CPU expected commercially in 2028, High Bandwidth Compute with the AI250 sampling in mid-2027, AI300 sampling in 2028, plus inference accelerators, connectivity products, and custom silicon. The pitch is built around performance per watt, token economics, and lower total cost of ownership for inference and agentic-AI workloads. What the Amazon announcement doesn’t say is whether the AWS work uses any of those named products or a separate custom design entirely.
The competitive setting is crowded, of course. AMD and Intel are both chasing hyperscaler custom-silicon and AI-infrastructure business, the cloud providers all have in-house programs, and Nvidia sits atop the whole market. Qualcomm’s disclosed inference focus is arguably the right wedge — inference demand is growing fast and hyperscalers are hunting for capacity, cost, and energy-efficiency gains wherever they can find them. But it’s still a late entry into a market where everyone else has a head start.
The real open questions are whether these custom chips complement or overlap Trainium and Inferentia, whether they’re destined for AWS’s internal use or customer-facing instances, and how the design relates to Qualcomm’s named Dragonfly parts.
