Qualcomm handed Amazon a 4 billion dollar warrant to win one customer

Qualcomm spent a decade being described as a phone company that kept failing to become a server company. On 8 September it became a server company, and the price of admission tells you a lot about who holds power in AI hardware right now.

The deal is a multi generation agreement to build customised silicon for large scale AWS AI data centres, aimed squarely at inference rather than training. Amazon can buy up to 60 billion dollars of Qualcomm data centre chips and related products across the term. Qualcomm shares rose roughly 9 percent.

Read the warrant, not the headline number

The detail that matters sits below the 60 billion dollar figure. Qualcomm granted Amazon a warrant to acquire up to 25 million Qualcomm common shares at a fixed 161.26 dollars each, worth about 4 billion dollars. A first tranche of 3.75 million shares vested on issue, tied to initial purchase commitments. The rest vests as Amazon hits agreed commercial milestones and keeps buying.

Strip the financial language away and the shape is simple: the supplier is paying the customer, in its own equity, for the privilege of being designed in. Amazon gets upside on the stock that its own purchase orders will move. Qualcomm gets a reference customer that makes every other hyperscaler conversation easier.

That is not a scandal. Warrant structures show up all over the AI buildout now, because the people with the power are the ones who own the data centres. It does tell you that nobody wins a hyperscaler socket on merit alone in 2026. You win it by sharing the upside.

Why inference is the opening

Training is where Nvidia is effectively unassailable, because training rewards raw throughput, interconnect and a mature software stack. Inference is a different problem. It rewards performance per watt, cost per token and predictable latency at enormous volume. It also runs forever, while a training run ends.

Amazon already builds its own inference silicon with Inferentia and Trainium, so this is not a company short of options. Adding Qualcomm suggests AWS wants a second credible supplier with real power efficiency engineering behind it, which is exactly what twenty years of fighting for milliwatts in phones produces.

The two companies also said they intend to work together on high speed optical connectivity, including solutions scaling to 1.6 terabits per second. That is the quieter half of the announcement and arguably the more strategic one. Moving data between accelerators is now the binding constraint in large clusters, and co packaged optics is where the next round of advantage gets decided.

What this changes

For Nvidia, one inference deal does not dent a business this dominant. What it does is normalise the idea that inference capacity gets bought from whoever is cheapest per watt, which is a harder position to defend than owning the training stack.

For everyone renting compute, more suppliers eventually means softer pricing, but not this year. Contracts of this shape run over multiple chip generations, and the capacity being negotiated now lands in data centres from 2027.

The honest caveat: up to 60 billion dollars is a ceiling, not a commitment. Milestone based deals get quietly rescoped all the time, and the warrant vesting schedule exists precisely because neither side trusts a press release. Watch the vesting disclosures over the next four quarters. They will tell you whether Amazon is really buying, long before any revenue number does.