Qualcomm hands Amazon a $4B stake option to win AWS inference silicon

Share
Qualcomm hands Amazon a $4B stake option to win AWS inference silicon

Two big infrastructure moves and one surveillance story you should read in full. Here's the midday round.

Qualcomm and Amazon signed a multi-generational chip deal, and Amazon got an option to buy roughly $4 billion of Qualcomm stock as part of it. Per a regulatory filing, the warrant lets Amazon buy shares at $161.26 and is tied to as much as $60 billion of business under the agreement. The work itself is inference-first: custom silicon for AWS AI data centers, plus high-speed optical connectivity reaching 1.6 terabits per second. Qualcomm said it will also run more of its own chip-design workloads on AWS, and shares rose more than 7% in early trading.

This is the second hyperscaler to vouch for Qualcomm's data center push this year, after Meta signed on for the Dragonfly C1000 CPU in June — and it's a bigger one, because a warrant tied to future volume is a commitment, not a pilot. The tell is where the money is going: inference, not training, and optical interconnects, not just accelerators. That's the part of the AI stack where Nvidia's advantage is thinnest and where a company with Qualcomm's power-efficiency pedigree can plausibly win. Qualcomm has talked about $15 billion in data center revenue by fiscal 2029; this deal is the first outside evidence that target isn't just a slide.


Border Patrol is running secretive "Predictive Intelligence Targeting Teams" that mine Americans' financial activity and hand tips to local police, who then make traffic stops. The units' name is being reported for the first time by 404 Media, which obtained a DHS document in the case of Kyle William Olson: an agent in the Spokane Sector's PITT flagged "financial activity patterns commonly associated with illicit narcotics activity," passed it to the Montana Highway Patrol, and a trooper stopped Olson for an obstructed license plate. The stop produced a DUI charge and a possession-with-intent charge. The agent wrote that his intelligence was not the "sole basis" for the stop.

That phrase is the whole story. An officer still has to find a real reason to pull someone over — but the reason he was looking is a financial-pattern score nobody outside DHS can inspect. That's parallel construction with a machine-learning wrapper, and it launders a hunch into probable cause. CBP declined to say what financial data it monitors or whether it gets a warrant. Two sectors have PITT units that we know of; the honest assumption is more.


OpenAI says its first custom inference chip, Jalapeño, starts deploying by year-end. In a post today, CFO Sarah Friar said the chip delivered 1.5 to 1.9 times the peak token throughput per watt of the commercial systems it tested across three public models, with end-to-end latency 1.7 to 3.6 times lower. It will run alongside accelerators from Nvidia, AMD and others — not instead of them. Separately, Friar said the research organization now uses 3.1 agent-workdays for every workday of human labor.

Custom silicon is the difference between a lab that rents compute and a company that controls its unit economics. OpenAI is following the same path Amazon did with its own chips: keep the partners, but own the layer where the margin leaks. Jalapeño shipping is the moment that claim stops being a roadmap and becomes a procurement decision someone has to defend.

What to watch: whether other hyperscalers follow Qualcomm's equity-warrant structure — it turns a supplier into a partner in a way purchase orders never do.

Should a chip supplier's stock be part of a hyperscaler's comp package, or does that just lock out competitors? Tell us in the comments.

Sources: Qualcomm press release · Reuters · CNBC · 404 Media · OpenAI — The Work Now Within Reach · OpenAI — Jalapeño first results