Amazon is paying Qualcomm in stock warrants, up to 60 billion dollars of chips

Twenty-five million Qualcomm shares at 161.26 dollars each. That is part of what Amazon is being given in exchange for buying chips, disclosed in a Qualcomm SEC Form 8-K on 8 September.

The rest of the agreement is more conventional. Qualcomm and AWS signed a multi-generation deal covering custom AI inference silicon and 1.6T optical connectivity for AWS data centres, with a ceiling of 60 billion dollars in purchases running to 2036.

The 60 billion dollar number is a ceiling, not a contract

Only 3.75 million of the warrant shares vested at signing. The rest vests against binding purchase orders that have not been placed yet, spread across the next decade. Qualcomm did not disclose the value of the initial commitment.

So the headline figure describes the most AWS could buy, not what it has agreed to buy. That distinction is worth holding onto, because 60 billion is the number that will travel.

Why the Qualcomm AWS AI chip deal matters more than its size

Two things make this more than a supply arrangement.

The first is the customer. Qualcomm has spent a decade being the phone company, and this is its first Western hyperscaler customer for data centre silicon. Landing AWS specifically is the credential that makes every following conversation easier.

The second is Alphawave Semi. Qualcomm bought the SerDes and optical DSP business in 2025, and this is its first commercial validation. Alphawave supplies the building blocks for the 1.6T optics half of the deal, which is the part that moves data between racks without the power cost eating the gains.

Inference is a memory problem

The chip fight has quietly moved from training to inference, and the two are not the same engineering problem. Training rewards raw compute throughput. Inference, and specifically the decode phase where a model produces tokens one at a time, runs into memory bandwidth and capacity long before it runs out of FLOPS. That distinction is worth understanding properly if you want this year's hardware announcements to make sense.

Qualcomm calls its approach High Bandwidth Compute and aims it at decode-heavy workloads. AMD made the same bet with Helios. Nvidia built its moat on training, which is now the smaller half of the market.

What to watch

The first real purchase orders, because those convert warrant shares into vested equity and a ceiling into revenue. Nothing here is proven until an order lands.

It also arrives in a year where national compute spending and chip supply politics already decide who can buy what. Qualcomm being asked to pay partly in its own shares is a reminder that hyperscalers have leverage nobody else has.