Amazon's Bet on Qualcomm Is About Inference, Not the $60 Billion

On September 8, Qualcomm said it would work with Amazon across multiple generations of customized silicon to build out Amazon Web Services' AI data centers, and the stock jumped about 10% that day. What turned a partnership announcement into a story was the scale attached to it: reports of a purchase opportunity as large as $60 billion, plus warrants giving Amazon the right to buy up to 25 million Qualcomm shares. The stock now sits around $182, roughly 3% above the previous close.

Look past the headline number, and the most telling word in the release is not "$60 billion." It is "inference."

An inference deal, not a training deal

There are two very different jobs inside an AI data center. Training is the one-time, months-long slog of building a model on tens of thousands of GPUs. Inference is the ongoing work of running that model for every request — every chatbot answer, every image, every agent action — and it grows with usage, not with model releases. Today the AI compute market is dominated by training chips, where one supplier holds an overwhelming share and its software lock-in is the moat. Inference has different economics: it cares far more about cost, power per request, and latency, and less about the proprietary developer software that training demands.

That is precisely the frame where a challenger can get a foot in the door, and it is where Qualcomm is aiming. The company's entire engineering identity is power-efficient processing — the same skill that runs a phone all day is the one required to serve millions of inference requests without melting a data center's power budget. Amazon, one of the largest buyers of AI compute on earth, choosing to design chips across multiple product generations with Qualcomm is the strongest endorsement yet of that argument. It is not a forecast from a sell-side analyst; it is a hyperscaler betting its own infrastructure roadmap on Qualcomm's silicon.

The $60 billion is an opportunity, not an order

This is also where the fine print matters, because the deal's structure is a commitment with conditions, not a guaranteed revenue stream.

The $60 billion figure is a reported purchase opportunity — an agreed ceiling on how much Amazon could buy over the life of the collaboration, not an order book and not revenue that has been booked. Actual purchases depend on Qualcomm shipping chips that work, at scale, on schedule, across generations. The warrants make the conditionality explicit: Amazon can acquire up to 25 million Qualcomm shares at $161.26, vesting in tranches tied to its purchases and expiring in 2036. Think of the warrants as a sweetener that aligns Amazon's interest with Qualcomm's success — Amazon only makes money on them if it actually buys the silicon and Qualcomm executes.

For Qualcomm's existing shareholders that structure cuts two ways. It is a genuine vote of confidence from a customer that will only realize its optional value if Qualcomm delivers. But it also means the payoff is diluted, back-loaded, and far down the timeline. The company has said it targets $15 billion in annual data-center revenue by fiscal 2029 — a goal that, even if met, is years of execution away and would represent a meaningful but not dominant slice of a business that does roughly $40 billion today.

What it does and doesn't change

The strategic context explains why this deal matters so much. Qualcomm's core handset and modem business is the slow part of the story — in the quarter it reported in late July, revenue was about $9.9 billion, and the company told investors that revenue from Apple products was declining faster than expected as Apple builds more of its own silicon. That Apple overhang is the reason Qualcomm has been pushed to diversify, and data center is now its most important new pillar. Two marquee customers in a few months — Meta, which will use Qualcomm's Dragonfly C1000 server CPU when it starts production in 2028, and now Amazon for custom inference silicon — are evidence the pivot is attracting real names, not just road-map promises.

Yet even after the pop, the market is not pricing Qualcomm as an AI compounder. It trades around 20 times trailing earnings and roughly 16 times forward earnings, with a dividend yield near 2%. Nvidia, by contrast, trades above 27 times trailing earnings, and AMD far higher. That gap is the market's way of saying it still sees Qualcomm as a phone company with an interesting data-center option attached.

On the merits, the Amazon deal is real progress: an inference leader-in-waiting just signed one of the largest possible customers to a multi-generation commitment. But the income statement will not show the effect this year or next, the $60 billion is conditional on execution, and the warrants would dilute existing holders only if Amazon exercises them. The honest read is that Amazon's bet de-risks — it does not complete — the Qualcomm AI story. It gives investors a reason to keep Qualcomm on the watch list and to re-examine the thesis; it does not turn a handset business into a delivered inference business in one announcement.