Qualcomm is preparing to provide custom AI chips to Amazon, Bloomberg reports, a move that would land the San Diego chipmaker its most significant datacenter design win to date and pit it directly against the merchant AI silicon business that Nvidia has built across the hyperscalers.

The report, attributed to people familiar with the matter, indicates Qualcomm would join the roster of suppliers developing application-specific accelerators for Amazon's cloud infrastructure, complementing — or in some cases displacing — the in-house Trainium and Inferentia product families that Amazon Web Services has developed for training and running large AI models. Bloomberg did not disclose financial terms, volumes, or a timeline for the engagement, and neither Qualcomm nor Amazon has publicly confirmed the arrangement.

A deal of this kind would mark a strategic escalation for Qualcomm, whose datacenter ambitions have waxed and waned since it sold its Centriq server processor business in 2018 after that Arm-based line failed to gain commercial traction against Intel's Xeon dynasty. Since then, the company's revenue has rested heavily on handset silicon — Snapdragon application processors for Android flagships, RF front-end components, and automotive platforms — with chairman and CEO Cristiano Amon repeatedly stating his intent to diversify the company's revenue base beyond mobile.

Custom silicon for hyperscalers is one of the few segments where that diversification thesis has already produced revenue. Qualcomm's design win with Meta, disclosed in 2023, involves custom server processors built on Arm cores, and the company has signaled that additional hyperscaler engagements are central to its growth story. Amazon, the largest cloud operator by revenue and the operator of one of the industry's most aggressive internal chip programs, would be a marquee addition to that customer list.

The competitive backdrop explains the timing. AWS continues to expand its Trainium line — the second generation, Trainium2, entered large-scale deployment across its facilities in late 2024, and Amazon has discussed clusters linking tens of thousands of the accelerators for frontier-model training. Yet demand for AI compute continues to outrun what any single internal program can supply, and cloud operators have grown more willing to split orders across multiple silicon suppliers to secure capacity and pricing leverage. Broadcom and Marvell have built multibillion-dollar businesses on exactly this dynamic, designing custom accelerators for hyperscalers that also buy Nvidia's GPUs. Qualcomm, which licenses Arm architectures and has deep experience in low-power design and systems-on-chip integration, is chasing the same opportunity.

For Amazon, adding Qualcomm as a custom-silicon partner would extend a sourcing strategy already spanning internally developed accelerators, Nvidia GPUs, and AMD's Instinct MI300-series parts. Multiple suppliers give AWS pricing leverage at a moment when AI infrastructure capital expenditure is climbing across the industry, and design competition among vendors tends to accelerate the performance-per-dollar gains that cloud operators use to defend their margins.

Investors treated the report as material. Qualcomm shares moved on the Bloomberg headline, reflecting the market's long-running valuation of the company's diversification effort: handset-connected businesses still dominate its revenue mix, and each datacenter or automotive win is read as evidence that the growth profile can broaden. An Amazon engagement, if confirmed, would give Qualcomm a second named hyperscaler customer alongside Meta and a credible reference for pursuing further custom accelerator business.

Caution is warranted on specifics. The report does not establish whether the Qualcomm-designed parts would serve training workloads, inference, or auxiliary functions such as storage and networking offload; it does not name a process node, a foundry, or target deployment dates; and talks of this kind can change scope or stall before resulting in shipping silicon. Custom chip programs typically run 18 to 30 months from design start to volume deployment, so any revenue effect would likely arrive in later fiscal years even if the partnership proceeds on schedule.

Neither company commented publicly at the time of the report. If the arrangement advances to confirmed volumes, it would tighten the three-way contest among merchant GPU suppliers, established custom-silicon houses like Broadcom and Marvell, and Qualcomm's newer entrant effort — with Amazon's datacenter procurement becoming one of the clearest signals of how the AI accelerator supply balance shifts over the next several quarters.