Qualcomm–AWS AI chip deal: read the $60B warrant terms

Qualcomm announced a multi-generation collaboration with Amazon today to build custom silicon for AWS data centers, aimed squarely at AI inference. The stock jumped about 10%. The press release is the usual language about next-generation infrastructure. The interesting document is the 8-K Qualcomm filed the same day, because that is where the deal stops being a partnership and starts being a number.

What actually happened

Per Qualcomm's announcement, the two companies will work on "customized silicon at scale for large-scale AI data centers," plus optical connectivity "extending up to 1.6T" for high-bandwidth interconnects inside Amazon's networks. Inference, not training. Interconnect, not just accelerators.

Now the 8-K, Item 3.02. Qualcomm issued Amazon a warrant for up to 25,000,000 shares at $161.26 per share, cashless exercise, expiring September 3, 2036. The shares vest in tranches tied to commercial arrangements, binding purchase orders, and actual purchases of Qualcomm's server chip products and manufacturing services — up to a maximum of $60 billion in payments. 3,750,000 shares vested on issuance against initial purchase commitments.

Read that structure honestly. Amazon gets paid in Qualcomm equity for buying Qualcomm silicon, on a ten-year clock. Only 15% of the warrant is vested today. The other 85% is a commitment schedule dressed as an option.

Why this matters for your business

Your inference price is now set by a supply chain, not a pricing page. When your cloud vendor co-designs the chip, the endpoint you call, and the network between them, the marginal cost of a token becomes an internal transfer number. That can go your way — vertical integration usually does lower cost eventually. It also means the number is no longer disciplined by a competitor's price list.

Ten-year hardware deals are a lock-in signal, not a stability signal. A vendor with $60 billion of purchase incentives has a reason to steer workloads onto its own silicon. Expect the cheapest inference on AWS to increasingly be the inference that runs on chips AWS helped design. That is fine until your model of choice is not the one running there.

The portability test is the same as it always was. Keep the provider and model as config values. Keep an eval set of 30–50 real cases from your own business, in version control. Run them against a second provider once a quarter and record cost and latency. If you have never done that, you don't have a fallback — you have a hope. We wrote up the token-bill version of this argument earlier.

Watch the hardware side of your budget too. Qualcomm already raised prices during the memory crunch, and Broadcom is carrying debt against AI chip demand. The capital going into inference silicon has to be earned back from somebody's invoice. Sign short contracts and re-price annually.

Sources: Qualcomm, SEC Form 8-K, CNBC.