Qualcomm filed something far bigger than a normal vendor announcement on Tuesday, and the tell is in the SEC filing, not the press release. Qualcomm (NASDAQ: QCOM) is building customized silicon for Amazon across multiple generations, and it let Amazon lock in the price of that future relationship by taking warrants worth roughly $4 billion. The full scope is on the Qualcomm press release, but the financial mechanics buried in the SEC disclosure tell the real story.
What Qualcomm and Amazon Actually Agreed To
The public framing is clean. Qualcomm said it is working with Amazon across "multiple generations of customized silicon" focused on AI inference, meaning the chips are built for running models in production rather than training them from scratch. The companies are also pairing that silicon with optical connectivity solutions extending up to 1.6T, plus Qualcomm's own SerDes and optical DSP technology to move data between chips at scale.
The part that actually matters is the money. Per the SEC filing, as reported by CNBC, Qualcomm issued warrants letting Amazon buy 25 million Qualcomm shares at $161.26 apiece, for a total exposure of about $4 billion. Those warrants expire on Sept. 3, 2036, roughly a decade out. The shares vest in tranches tied to the "execution of certain commercial arrangements," including the purchase of up to $60 billion worth of Qualcomm's server chips and related technology over time.
Why the Money Structure Matters
Most custom silicon deals are non-disclosed NDAs until a product ships. This one is disclosed in full, and the warrant structure is a deliberate signal read by anyone who understands how these negotiations work.
From an engineering standpoint, Amazon is buying optionality at a fixed price. If Qualcomm's data center silicon works and AWS migrates workloads onto it, those 25 million shares are worth a lot. If it underperforms, Amazon's maximum cash out is the purchase price, while Qualcomm still carries the R&D risk. That is a risk profile tilted toward the customer. A data scientist watching the mechanics should note this is exactly how you price a multi-year R&D bet with no benchmark data to back it: you let the buyer cap downside and take upside.
From a hardware standpoint, the $60 billion purchase threshold is the real contract. The warrants are just the public banner; the actual obligation is that Amazon commits to buying $60 billion of Qualcomm server chips before the warrants fully vest. That number is larger than Qualcomm's entire current data center revenue, which means this deal is a floor, not a ceiling.
The Dragonfly Context
This is not Qualcomm entering the data center cold. In June, Qualcomm revealed its Dragonfly C1000 CPU, a chip built for agentic AI with an explicit focus on performance-per-watt rather than raw throughput. Meta agreed to use Dragonfly C1000 when it enters production in 2028. Qualcomm has publicly targeted $15 billion in data center sales by fiscal 2029 and laid out a roadmap with several products, including a standalone AI chip and a chip-tying interconnect.
The AWS deal is the validation of that roadmap. Meta was a single design win. Amazon is a multi-generational contract with billions committed upfront. The progression from "here is a chip" to "here is $4 billion in committed capital" is the difference between a prototype and a business.
What It Says About the Market
The data center silicon market has been a near-total Nvidia monopoly, built on GPUs that dominate both training and inference. But the economics are shifting. As CNBC reported, Bank of America predicts the CPU market could more than double, from $27 billion in 2025 to $60 billion by 2030, driven by inference workloads that favor fewer powerful cores over thousands of narrow ones. That is a structural tailwind for Qualcomm, AMD, and Intel.
Qualcomm's angle is power efficiency. A GPU is ideal for the massive parallel math of model training, but inference often runs workloads that are more sequential. Qualcomm's decades of mobile chip design, where power draw and thermal budget are existential constraints, translate directly into a data center pitch: do the inference cheaper per token. AWS, whose capital expenditure on AI infrastructure is already reaching into the hundreds of billions annually, benefits by not being locked into a single vendor. The parallel with the recent Nvidia move to protect its dominance is unmistakable: hyperscalers are actively building secondary silicon supply chains, and Nvidia is responding.
Our Read
Three things to watch from a technical and business standpoint.
First, the 1.6T optical interconnect is the hidden line item. Custom silicon is useless without a way to move data between chips fast enough, and Qualcomm is one of the few companies that designs both the compute and the connectivity. That vertical integration is the actual moat, not the chip alone.
Second, inference-first positioning matters for timing. Training a frontier model can take years and tens of thousands of GPUs. Inference infrastructure ships faster and generates revenue immediately, which is why Qualcomm scoped this deal to inference rather than training. It is the lower-risk path to the $60 billion target.
Third, the warrant structure is a template. If AWS migrates workloads onto Dragonfly silicon and hits its $60 billion purchase goal, the same disclosed-terms approach could appear in other hyperscaler deals. The market is moving from opaque NDAs to priced, time-boxed commitments, which makes these bets easier to value than the old hidden-deals model.
The bottom line: Qualcomm has converted a promising chip design into a multi-year contract backed by real capital from one of the world's largest hyperscalers. Whether the silicon performs at 1.6T speeds is still unproven, but the business bet is now on the record.