Qualcomm issues Amazon warrants on 25 million shares to co-design AWS inference chips

An ML platform lead reserving inference capacity for next year can read this announcement twice and change nothing in the plan. Qualcomm and Amazon did not publish a part name, an availability date, or a price ^[14].

The SEC filing cited by TechRadar covers 25 million Qualcomm shares priced at $161.26, which the publication puts at around $4.03bn of stock ^[9]. Multiply the two and you get $4.0315bn ^[16]. Qualcomm traded at $174.09 when TechRadar published, up nearly 5% over five days ^[10]. The strike sits $12.83 under that price, so the warrant carried roughly $320.75m of value above what Amazon would pay to exercise it ^[17].

Qualcomm's side of the deal is a route into a market where, as TechRadar notes, Nvidia has had major success, with the custom AWS silicon focused on inference ^[7]. Amazon already has extensive experience designing its own data center silicon through AWS ^[13]. The buying runs both ways. Qualcomm plans to deepen its use of AWS AI infrastructure, including Amazon Bedrock, for electronic design automation, targeting a reduction in chip design cycles ^[8].

AWS vice president Prasad Kalyanaraman said, "By working together on customized silicon and advanced connectivity, we're delivering more performant, efficient, and cost-effective infrastructure for our customers" ^[6]. The verb there is present tense, and Qualcomm says the collaboration will span multiple generations of customized silicon ^[5].

The only measured efficiency figures in either account describe buildings. Amazon's 2025 sustainability report declared a global average PUE of 1.14, nearly 9% better than the public cloud industry average of 1.25 ^[11], and Google says it managed 1.09 fleet-wide ^[12]. The difference is 0.05 ^[18]. TechRadar writes that Amazon's work with Qualcomm "could be key to it overtaking Google" on that measure ^[20]. That link is the publication's inference. Neither company claimed it.

The connectivity half of the deal gets less attention and may land sooner. Interesting Engineering's account describes the movement of data between processors as a significant bottleneck as AI workloads scale ^[15], and the stated target is optical connectivity extending up to 1.6T ^[2], built on Qualcomm's SerDes and optical DSP technologies ^[4].

Before a custom accelerator goes into a 2027 capacity plan, I want four things: a part number, an instance type that launches in a real account, a published hourly rate, and a benchmark on the model you actually serve. At renewal, buyers price the part they can measure. This announcement supplies multi-generational product language ^[19] and a warrant struck at $161.26 ^[9].