Amazon's $60B Qualcomm deal expands AI inference hardware
A $60 billion investment deal between Amazon and chipmaker Qualcomm reflects a growing emphasis on optimizing AI inference hardware and containing its costs.
Amazon and Qualcomm will develop custom silicon and high-speed optical connectivity for AI inference, potentially giving AWS users and data center operators a new set of hardware options as data center power and infrastructure costs rise.
Amazon's Qualcomm deal, revealed by the chipmaker on Sept. 8, opens up another hardware path for AI inference as power constraints push data center architects toward more specialized infrastructure. Qualcomm is targeting processors and connectivity designed around those constraints.
The companies will work across multiple generations of customized AI inference silicon and optical connectivity designs, reaching up to 1.6 terabits per second for large-scale AI data centers.
Stephen Sopko, practice lead, semiconductor and deep tech at HyperFrame Research, said the emphasis on inference is important because the requirements differ from AI training.
"Power is the constraint that decides which architectures get built," Sopko said in an email.
AI model training workloads tend to emphasize peak floating-point performance and high-bandwidth memory, while inference puts more weight on metrics such as tokens per watt, tokens per rack and available power.
For IT buyers weighing cloud providers for AI workloads, the deal could broaden the hardware options under AWS inference services. For data center operators considering their own AI infrastructure, Qualcomm is adding another processor and connectivity supplier to the market, although neither company disclosed what products Amazon plans or when production begins.
Commercial commitment runs through 2036
Qualcomm issued Amazon.com NV Investment Holdings LLC a warrant to acquire up to 25 million Qualcomm shares at $161.26 per share, according to a Qualcomm SEC filing. The warrant is tied to commercial arrangements, binding purchase orders and actual purchases of Qualcomm server chip products, technology, systems and manufacturing services.
It's definitely significant … because of the multiple-generational aspect of the partnership. Hyperscalers don't typically commit to multiple generations for commodity parts.
Matt Kimball, Vice president, Moor Insights & Strategy
Those purchases can total up to $60 billion in payments over the warrant's term, which expires in September 2036. Qualcomm said 3.75 million shares vested when the warrant was issued based on initial purchase commitments.
Matt Kimball, vice president and principal analyst for Data Center Technologies at Moor Insights & Strategy, said in an email that the multi-generational structure is what makes the deal notable.
"It's definitely significant … because of the multiple-generational aspect of the partnership," Kimball said. "Hyperscalers don't typically commit to multiple generations for commodity parts."
Kimball said the structure suggests Amazon is planning around a longer-term hardware architecture rather than evaluating a single component.
Amazon has spent years developing its own data center silicon through its Trainium AI accelerators and Graviton server CPUs.
Qualcomm is working with Amazon on customized silicon for a specific part of the AI infrastructure stack while Amazon continues to develop its own chips.
Qualcomm is building a broader data center business
The work between Qualcomm and Amazon also covers multi-generational development of high-bandwidth network connectivity.
"Connectivity is not an accessory to the accelerator story; it is half the system," Sopko said.
Qualcomm will use its SerDes and optical digital signal processor technologies. The company gained additional high-speed connectivity capabilities through its acquisition of Alphawave Semi.
Kimball said AWS already has multiple connectivity suppliers, including Marvell and Broadcom, so the networking work should not be automatically interpreted as Amazon replacing those vendors.
The more significant shift is Qualcomm's combination of custom compute and connectivity capabilities, he said.
Qualcomm is also expanding into software and AI infrastructure. The company plans to deepen its use of AWS infrastructure, including Amazon Bedrock, for electronic design automation workloads to reduce chip design cycles.
Qualcomm's current data center portfolio includes the Dragonfly C1000 server CPU, AI200, AI250 and AI300 inference accelerators, high-bandwidth compute technology and 800G and 1.6T connectivity.
The company announced a separate multi-generational data center CPU agreement with Meta in June. Meta plans to use Qualcomm's C1000 in its next-generation server fleet, with production expected in the second half of 2028.
Qualcomm completed its acquisition of Modular in July, adding technology aimed at the software and AI infrastructure stack. The company's broader data center strategy includes AI accelerators, server CPUs, high-bandwidth compute technology and custom silicon. It has targeted more than $15 billion in data center revenue by fiscal 2029.
Shane Snider is a Senior News Writer at TechTarget, covering AI infrastructure, hyperscale data centers, cloud platforms, and the power and energy systems driving modern compute expansion. You can reach Shane at [email protected] or on LinkedIn.
