Summary: Qualcomm and Amazon are pairing custom inference silicon with 1.6T optical connectivity, but the filing behind the deal makes clear that the headline $60 billion figure is a long-dated, conditional ceiling—not a contracted revenue windfall or proof that a new chip has beaten Nvidia.
Qualcomm’s newest Amazon deal has an unusually sharp edge: the company says it will help build custom silicon for AWS’s AI infrastructure, while a regulatory filing says Amazon can earn stock warrants against up to $60 billion in payments. The number is large enough to redraw a few investor slides. It is not, on the evidence available, $60 billion of Qualcomm revenue.
The collaboration, announced September 8, is aimed at large-scale AI inference, the expensive business of serving model responses after training. Qualcomm says the two companies will work across multiple generations of customized silicon and optical connections reaching 1.6 terabits per second. Amazon, for its part, described the goal as more performant, efficient and cost-effective infrastructure.
The $60 billion figure comes with conditions
Qualcomm’s Form 8-K, filed September 8 and reporting an agreement dated September 3, provides the essential caveat. An Amazon affiliate received a warrant to buy up to 25 million Qualcomm shares at $161.26 each, expiring September 3, 2036. The warrant shares vest in tranches tied to commercial arrangements, binding purchase orders and actual purchases of Qualcomm server-chip products, technology, systems and manufacturing services. The filing puts the maximum payment amount at $60 billion; 3.75 million shares vested on issuance based on initial purchase commitments.
That structure matters. A warrant linked to purchases can align a supplier and a hyperscaler for a decade, but it does not turn the ceiling into booked sales. Actual volumes, product mix, deployment timing and the value that ultimately vests all remain contingent. It also creates a trade-off for Qualcomm shareholders: the commercial upside comes with potential dilution if the warrant is exercised.
Inference is more than an accelerator problem
Qualcomm is positioning the partnership inside a broader data-center plan unveiled in June. Its Dragonfly roadmap combines the AI300 inference accelerator, C1000 CPU, High Bandwidth Compute memory technology and 800G/1.6T connectivity. Qualcomm projects better energy efficiency and token economics, but those are company estimates based on product specifications, not independently comparable cloud benchmarks.
That qualification is more than legal fine print. Inference cost is determined by the accelerator, but also by memory bandwidth, packaging, networking, cooling, utilization and software. A chip can look efficient in isolation and still lose its advantage if a model’s memory behavior, networking pattern or serving stack becomes the constraint.
The optical work is therefore not a decorative add-on. Qualcomm says the companies will use its SerDes and optical DSP technology for high-bandwidth links. If model serving shifts toward larger clusters and more disaggregated racks, data movement can become as important to the bill as raw compute. But 1.6T links also require a real ecosystem of optics, switches, cables, packaging and operational reliability. The press release does not disclose which workloads, topologies or deployment dates the partnership has committed to.
AWS already has an in-house answer
Amazon is not entering this partnership because it lacks AI chips. It sells Trainium and Inferentia alongside Nvidia-based instances, and has repeatedly argued that custom silicon lowers its costs and reduces dependence on a single supplier. In its latest annual-report materials, Amazon said Trainium2 was fully subscribed with 1.4 million chips landed, that it powers most inference on Bedrock, and that Trainium3 was already running production workloads. Those are Amazon’s own disclosures, not a neutral comparison of every workload.
This makes Qualcomm more likely to be a complement than a clean replacement. AWS can use a customized external design where it has a workload, supply-chain or system-level reason to do so, while keeping Trainium as a core vertical-integration bet and Nvidia GPUs for the broad software ecosystem. Google’s TPUs and Meta’s internal accelerators point to the same market shift: cloud giants are seeking more options, not necessarily a single universal winner.
Who should care, and who can wait
Cloud operators and AI services with sustained, predictable inference demand have the most to gain if the partnership produces lower tokens-per-watt and better capacity availability. They can justify the engineering work to tune models, schedulers and networking around a new platform. Optical and advanced-packaging suppliers may benefit if such rack-scale designs move into volume.
Smaller teams do not need to make a buying decision from this announcement. For them, the relevant comparison is likely to remain the delivered price, model availability, latency and operational friction of AWS instances—not a future custom chip’s claimed efficiency. Enterprises should ask whether a provider can show workload-specific cost, throughput and power data, and whether switching requires model-porting or networking commitments that erase the savings.
The next evidence to watch is practical: named deployments, sampling dates, software support, independent benchmarks using the same model and serving setup, and any indication that lower infrastructure cost reaches customer pricing. Until then, the deal is best read as a serious option on inference infrastructure—not a completed upset of the GPU market.