$60 billion. That is the number Qualcomm attached to its new agreement with Amazon, and the number every headline repeated. Look at the mechanism underneath and the story changes shape. The $60 billion is a ceiling on what Amazon Web Services may buy over time. The guarantee is smaller, and stranger: a warrant for 25 million Qualcomm shares struck at $161.26, vesting only as orders land.
Qualcomm and Amazon announced the multi-generation collaboration on September 8. The two companies will co-develop customized silicon for AI inference — running a trained model, not building one — plus optical links rated to 1.6 terabits per second inside Amazon's data centers. The chipmaker will also move more of its own chip-design work onto AWS, including Amazon Bedrock in parts of its electronic-design automation flow.
The structure says where the AI buildout has migrated. Training gets the attention. Serving pays the bills.
The pact makes Qualcomm a third hyperscaler supplier after Meta and Microsoft, and the first paid partly in its buyer's own upside.
The real risk is the calendar: no disclosed delivery dates, an initial scope narrowed to inference, and silicon that will be judged against 2028-era competition.
What Amazon actually bought
The word doing the work in the announcement is "multi-generation." Amazon is not buying a finished part off a shelf. It is buying a development relationship that spans several cycles of custom silicon, with the first phase aimed squarely at inference workloads.
That distinction matters for anyone pricing the deal. Training a frontier model is a one-time capital event. Inference is a running cost that repeats with every query, every agent action, every generated token. As usage grows, so does the bill — which is why the economics of serving, not building, now set the agenda for chipmakers. As we wrote in September, that shift was already tipping the AI compute build from training to serving; this deal moves the contest to who supplies the serving layer.
Deloitte forecasts that inference will account for roughly two-thirds of all compute in 2026, up from about a third in 2023, and that the market for inference-optimized chips will pass $50 billion this year. Those chips are cheaper to run per query than the superpowered accelerators used for training, and they need less of the expensive high-bandwidth memory that training clusters consume by the terabyte.
The optical piece is easy to overlook and hard to overstate. Moving data between accelerators has become a bottleneck as clusters grow past tens of thousands of chips. The 1.6-terabit links are aimed directly at that constraint, and they widen the deal beyond a single processor into the wiring of the data center itself.
Amazon is not the first hyperscaler to write this kind of check. Weeks earlier, Marvell Technology struck a custom AI chip agreement with Alphabet's Google that gave Google the right to buy a stake worth up to $12.2 billion. Reuters noted the pattern: financing in the AI boom is becoming intertwined, with customers taking equity in the suppliers they depend on.
The case that this time Qualcomm ships
Three facts support the bulls.
First, Amazon is paying in a currency that only makes sense if it intends to buy. A warrant that vests against purchases aligns both companies around volume. Amazon does not collect the shares by signing a memorandum of understanding, and the tranches are tied to commercial agreements and binding orders rather than goodwill.
Amazon does not take equity in a vendor it plans to test-drive.— Analyst quoted by Fierce Network
Second, the chipmaker arrives with a portfolio rather than a single part. Its Dragonfly data-center line, laid out in June, spans custom silicon, central processing units, inference accelerators and the high-speed SerDes links that move data between them. The company's stated edge is energy efficiency — a metric that matters more for inference than for training, because inference never stops.
Third, the customer list is forming. The chipmaker had already signed Meta for its C1000 processor and drawn Microsoft into its data-center plans. Amazon is the third hyperscaler on that list, and the one accepting equity upside instead of a discount. Each design win makes the next one easier to sell, and a hyperscale reference customer is the hardest reference to land.
Bob O'Donnell, chief analyst at TECHnalysis Research, read the announcement as the reassurance the chipmaker needed — proof that its data-center ambitions "could indeed be met." Its chief financial officer, Akash Palkhiwala, called it "a landmark deal" that "kicks us off" in the data-center business. Management is not treating this as a science project.
The case that a ceiling is not an order
The bears have the structure on their side.
$60 billion is not revenue. It is the maximum that determines how far the warrant vests, and the filing language drifts between "could" and "may." If Amazon orders a tenth of the ceiling, the headline shrinks and the warrant delivers a tenth of its shares. Nothing in the agreement obliges Amazon to buy anything.
The warrant is also a cost. Issuing 25 million shares struck at $161.26 hands Amazon upside that the existing shareholder base funds through dilution if the stock climbs. The shares rose about 3% on the announcement — a polite nod, not a market convinced it had just met the next Nvidia. By mid-September the stock traded near $188.
The chipmaker has tried this before. Its Centriq server processor and data-center unit were shut down in 2018, roughly one year after launch, after failing to loosen Intel's grip on the server market. The episode taught investors a lesson that still applies: a capable chip and a large customer are not the same thing as a durable franchise.
Then there is the calendar. Neither company disclosed when the first silicon arrives, how much Amazon expects to deploy, or whether AWS customers will be able to rent it directly. Jon Peddie Research told its readers to watch the space "between now and 2028." Two years is a long time in a market where Nvidia ships a new generation roughly every year.
Shares vest on purchases, so the chipmaker commits engineering capacity and grants an option before it books the revenue. If the program stalls, the cost stays with Qualcomm.
The terms that matter
Strip the framing and the deal is a set of numbers. The comparable is Marvell's agreement with Google, which set the template for hyperscaler equity in a custom-silicon supplier.
| Term | Qualcomm × Amazon | Marvell × Google |
|---|---|---|
| Focus | Custom AI inference silicon and 1.6T optics | Custom AI silicon |
| Headline value | Up to $60 billion in purchases | Stake right up to $12.2 billion |
| Equity component | Warrant for 25 million shares at $161.26 | Right to buy a stake |
| Announced | September 8, 2026 | August 2026 |
Terms as disclosed by the companies and reported by Reuters.
Amazon's maximum spend on Qualcomm chips
A ceiling that sets warrant vesting, not a binding order. · Qualcomm filing, September 2026
Strike price on Amazon's 25 million shares
Cashless exercise, expiring September 2036. · Qualcomm filing, September 2026
Signals to track before 2028
The debate will not be settled by press releases. It will be settled by a short list of observable events, most of which are now missing from the public record.
First purchase orders against the warrant, and how the vesting tranches are disclosed in the quarterly filings.
A named AWS inference instance type running Qualcomm silicon, rather than a general promise of future capacity.
Its data-center revenue line, which is still small enough to disappear inside the handset business.
Nvidia's next-generation launch cadence and pricing, which sets the bar any challenger has to clear.
If the chipmaker converts even part of the ceiling into shipped product, the $60 billion stops being a headline and starts being a franchise. If it does not, the agreement will look like a well-timed option — one that cost Qualcomm real engineering time and a slice of its equity to write.
The AI buildout has produced a new species of deal: customers paid in shares of the suppliers they depend on. Amazon now owns a small bet on Qualcomm, and the chipmaker owns a large bet on inference. Both sides will find out which one was the better trade long before 2036.