← Back to feed

Amazon Web Services and Qualcomm announced a multi-generation collaboration on September 8 covering custom silicon development and optical connectivity, with milestone-linked warrants tied to potential hardware purchases of up to $60 billion over ten years. Qualcomm shares surged as much as 11% intraday before paring gains to close up 3.2% at $174.09.

AWS joins Microsoft and Meta as named partners in Qualcomm’s data center silicon program, with a fourth unnamed global hyperscaler confirmed in progress by Qualcomm CFO Akash Palkhiwala at the Goldman Sachs Communacopia + Technology Conference the same day.

Qualcomm’s Re-Entry Into Data Center

Qualcomm shut down its Centriq server CPU program in 2018 after limited commercial traction. The 2026 Dragonfly portfolio is its structured re-entry. Unveiled at Qualcomm’s Investor Day on June 24, the portfolio covers two categories:

• Dragonfly C1000: An Arm-compatible server CPU built on Qualcomm’s Oryon cores, designed for general-purpose data center workloads and AI host-node orchestration
• Dragonfly AI200/AI250/AI300: Dedicated inference accelerators for AI workloads at scale

At the June launch, Qualcomm named Meta as a customer for Dragonfly C1000 server CPUs and separately disclosed engagements with two unnamed hyperscalers on custom silicon. AWS, revealed September 8, is the first of those two custom silicon partners to be named. The second remains undisclosed.

Three Partners, Each With a Different Deal Structure

The deals are not interchangeable. Meta’s agreement covers Dragonfly C1000 server CPUs for its internal infrastructure. Microsoft’s partnership is on Qualcomm’s High Bandwidth Compute architecture for Azure. AWS’s agreement spans custom silicon co-development and optical interconnects.

That variety is deliberate. Qualcomm is not selling a commodity chip to hyperscalers — it is entering co-development arrangements where each partner gets tailored hardware targeting its specific workload mix. The $60B AWS figure represents a ceiling on potential purchases across the collaboration’s duration, tied to milestone completion rather than a committed order.

Why Hyperscalers Are Signing

Inference workload costs are the primary driver. As frontier model deployment shifts from training to serving, hyperscalers are under pressure to reduce per-inference costs at the infrastructure level. Custom silicon — whether Google’s TPUs, Amazon’s Trainium and Inferentia, Microsoft’s Maia, or now Qualcomm-partnered designs — lets cloud providers optimize the cost structure of that compute rather than paying Nvidia margin on every inference operation.

Qualcomm’s re-entry matters because it adds a new supplier option. Broadcom, Marvell, and internal labs have been the main alternative-silicon paths for hyperscalers. A company with Qualcomm’s chip design and Arm expertise at this scale changes the competitive dynamic for procurement teams evaluating the next generation of inference infrastructure.

With three public partnerships and a fourth confirmed in progress, Qualcomm has rebuilt credibility in data center silicon faster than most expected when it exited the market eight years ago.