Amazon has given Qualcomm something every would-be challenger in artificial-intelligence infrastructure wants: a path to tens of billions of dollars in real purchases. A new long-term agreement ties Amazon's acquisition of Qualcomm server chips, technology, systems and manufacturing services to warrant vesting milestones that extend all the way to $60 billion in payments.

The financial structure is unusually concrete. In an SEC filing, Qualcomm disclosed that it issued an Amazon affiliate a warrant for as many as 25 million Qualcomm shares at an exercise price of $161.26. The warrant expires in 2036, and 3.75 million shares vested immediately based on initial purchase commitments. The rest vest in tranches as commercial arrangements, binding orders and actual purchases reach specified thresholds, up to the $60 billion ceiling.

At current scale, the warrant is worth roughly $4 billion, according to Reuters. But the more important number is the purchasing framework behind it. Qualcomm is trying to turn decades of mobile-chip expertise into a durable data-center business; Amazon is trying to expand AI capacity while avoiding dependence on any single architecture or supplier. The agreement gives both companies a reason to push that transition faster.

Custom silicon, not a simple off-the-shelf chip sale

Qualcomm described the relationship as a multi-generation collaboration on customized silicon for large-scale AI data centers, focused initially on inference—the phase in which trained models answer queries and perform tasks. The companies are also working on optical connectivity reaching 1.6 terabits per second and future generations of interconnect technology, according to Qualcomm's announcement.

That scope is important because modern AI systems are increasingly constrained by more than raw arithmetic. Data must move among accelerators, CPUs, memory and storage quickly enough to keep expensive computing hardware busy. Qualcomm's push into optical digital-signal processing and high-speed SerDes technology—strengthened by its acquisition of Alphawave—means the company is selling a combination of compute and connectivity rather than merely another accelerator card.

The deal also runs in the opposite direction. Qualcomm said it plans to deepen its use of AWS infrastructure and Amazon Bedrock for electronic-design-automation workloads, with the goal of shortening chip-design cycles. In other words, Amazon becomes both a large prospective buyer of Qualcomm silicon and a cloud supplier supporting Qualcomm's own development process.

Qualcomm is betting on inference as its next large market

Qualcomm has already set a public target that makes the Amazon relationship central to its diversification. At its June investor day, the company said it wants more than $15 billion in annual data-center revenue by fiscal 2029 and $40 billion in non-handset revenue overall. Its investor plan positions AI inference accelerators, CPUs, custom silicon and networking as one integrated business.

The broader product roadmap includes Dragonfly AI200, AI250 and AI300 inference accelerators, a C1000 data-center CPU and a high-bandwidth-compute architecture aimed at lowering energy use per generated token. Qualcomm's June roadmap promised an annual accelerator cadence and emphasized performance per watt, a metric that has become more consequential as power availability limits where and how quickly AI clusters can be built.

The Amazon warrant does not guarantee that all $60 billion will be spent. Vesting depends on future orders and purchases, and the agreement stretches across years. It does, however, make the commercial opportunity far less hypothetical than a conventional design-win announcement. Amazon has an economic incentive to reach the purchase thresholds, while Qualcomm has a clear incentive to deliver products that can win those orders.

Amazon is adding suppliers, not abandoning its own chips

The agreement should not be read as Amazon stepping away from its internally designed Trainium processors. Amazon CEO Andy Jassy said earlier this year that Trainium and the company's other custom silicon had become one of the world's largest data-center chip businesses, with more than $225 billion in Trainium revenue commitments. His Q1 update said Trainium3 was nearly fully subscribed and much of the still-future Trainium4 capacity had already been reserved.

Amazon has also committed to enormous volumes of third-party accelerators. In August, AWS and Nvidia announced plans to deploy 2 million additional Nvidia GPUs in 2027 and 2028, alongside deeper integration of networking, memory and CPU technologies. That expansion makes clear that AWS's strategy is heterogeneous: custom Amazon silicon, Nvidia GPUs and specialized partner hardware can coexist when customers need different price, performance or software characteristics.

The same pattern is visible in AWS's AI-inference partnerships. Amazon has paired Trainium with Cerebras systems for disaggregated inference, while its Neuron software stack is designed to let developers run PyTorch, JAX and common model frameworks on Trainium and Inferentia. AWS describes Neuron as the developer layer connecting those purpose-built processors to mainstream machine-learning software.

Why a $60 billion ceiling changes the competitive picture

Nvidia still dominates the accelerator market, and the Qualcomm-Amazon agreement does not establish that Qualcomm can match Nvidia's software ecosystem or displace AWS's own silicon. What it does establish is a credible route to scale. In semiconductors, large customers can justify manufacturing investment, accelerate software optimization and help suppliers spread development costs across more units. A purchase framework measured in tens of billions of dollars is therefore strategically meaningful even before every tranche is earned.

Amazon's own AI commitments explain why it wants more options. In February, Amazon and OpenAI expanded their relationship by $100 billion over eight years, with OpenAI committing to consume about 2 gigawatts of Trainium capacity through AWS. The companies' agreement spans current and future Trainium generations while Amazon continues to buy GPUs and now works with Qualcomm on custom inference silicon.

For Qualcomm, the message is equally direct. The company has spent years trying to reduce its dependence on smartphones as Apple brings more modem work in-house and handset growth matures. The Amazon deal gives its data-center plan an anchor customer, a measurable purchase ladder and a decade-long warrant structure tied to execution. The next test is no longer whether Qualcomm can announce an AI-server strategy. It is whether it can ship enough competitive silicon and networking technology for Amazon to climb toward that $60 billion ceiling.