The most important part of the deal between Amazon and Qualcomm is not the potential $60 billion figure. It is the way that figure is structured. Amazon is not merely buying a new batch of processors: it is tying a portion of its future infrastructure to multiple generations of custom silicon, and Qualcomm is using the contract to prove that its ambition to become a serious supplier for AI data centers is no longer just an investor pitch.

On September 8, Qualcomm announced a multi-year, multi-generational collaboration with Amazon to develop custom chips for large AI data centers, with an initial focus on inference—the stage where a trained model generates responses, images, code, or decisions. The two companies will also work on high-speed optical connectivity, with solutions reaching up to 1.6 terabits per second and subsequent generations.

Filings submitted to the SEC make the deal even more compelling. Qualcomm has issued Amazon a warrant to purchase up to 25 million Qualcomm shares at $161.26 per share, expiring in September 2036. The shares vest in tranches tied to orders, commercial agreements, and actual purchases of Qualcomm chips, technologies, systems, and manufacturing services, up to a maximum of $60 billion in payments. At signing, 3.75 million shares vested based on initial commitments.

This detail reframes the news. The $60 billion figure is neither a signed check nor a guarantee that Amazon will spend that amount. It represents the financial ceiling around which the incentive structure is built. The more Amazon buys and advances the program, the more the warrant vests. It is a framework that aligns customer and vendor, clearly signaling how critical Qualcomm views turning AWS from a partner into a customer of scale.

This is not a deal to “beat Nvidia” tomorrow

The temptation is to frame every new AI chip as a direct strike against Nvidia. That would be an oversimplification. Amazon is not walking away from Nvidia, and the numbers demonstrate this far better than any corporate statement.

On August 26, less than two weeks before the Qualcomm announcement, AWS and Nvidia announced an expansion of their partnership, scheduling the deployment of two million additional Nvidia GPUs in 2027 and 2028, alongside new integrations across CPUs, networking, memory, and infrastructure for agentic AI and robotics. Furthermore, AWS continues to develop its in-house silicon, spearheaded by the Trainium family.

The objective, therefore, is not to swap one vendor for another. It is to build a heterogeneous infrastructure where general-purpose GPUs, proprietary accelerators, CPUs, custom chips, and networking systems operate side-by-side, assigned to workloads based on cost, performance, power draw, and availability.

It marks a significant transition. During the opening phase of the generative AI boom, GPU scarcity made the ability to procure accelerators almost more urgent than the ability to optimize them. Now that operational volumes are surging and inference has become an ongoing cost, hyperscalers have an ever-stronger incentive to design specialized hardware tailored to their specific workloads.

Why inference is the right battleground for Qualcomm

Training frontier models demands massive clusters, ultra-fast interconnects, and deeply mature software. It is a market in which Nvidia has established an overwhelming lead, not just through GPUs, but through an end-to-end ecosystem of libraries, interconnect fabrics, and developer tools.

Inference presents very different economics. A model may be queried millions or billions of times a day, frequently under strict constraints regarding latency and cost per token. In such an environment, even a marginal gain in energy efficiency or memory utilization produces a substantial financial impact when scaled across entire data centers.

This is precisely the arena where Qualcomm aims to leverage the expertise it forged in mobile: delivering high compute throughput under tight power envelopes. In June, the company introduced the Dragonfly family for data centers, featuring CPUs, AI accelerators, and memory and connectivity technologies specifically engineered for inference. The announced roadmap includes Dragonfly AI200, AI250, and AI300 on an annual cadence.

In its joint announcement with Amazon, Qualcomm has not yet disclosed detailed chip specifications, mass-deployment timetables, or production volumes. It is therefore premature to evaluate these products against Nvidia GPUs, Trainium, or competing accelerators on real-world performance metrics. However, the existence of a multi-generational commitment shows that Amazon is not simply sourcing a standalone component: it is helping architect a platform designed to evolve over time.

The real bottleneck is no longer just the processor

One element of the agreement warrants nearly as much attention as the silicon: optics. Qualcomm and Amazon will collaborate on interconnects operating at up to 1.6 Tbps, leveraging Qualcomm’s SerDes and optical DSP technologies.

Inside hyperscale AI clusters, moving data has become almost as critical a challenge as running calculations. As models grow, memory must be distributed across multiple accelerators, requiring thousands of components to communicate at blistering speeds. When a processor sits idle waiting for lagged data delivery, compute capacity is wasted. When networking consumes too much power, overall infrastructure energy costs soar.

This is why semiconductor companies are seeking to control a larger portion of the stack: compute, memory, interconnect, networking, and software. Value is no longer concentrated in a single chip. It lies in the ability to make thousands of chips work together as if they were a single machine.

The fact that Amazon included optical connectivity in the same partnership shows just how central this systems-level logic has become. And it offers Qualcomm a second point of entry: not just selling accelerators, but becoming a supplier of the infrastructure that connects them.

For Qualcomm, it is the most concrete proof of diversification

Qualcomm has been trying for years to reduce its dependence on smartphones. At its Investor Day in June, it set a target of over $15 billion in data center revenue in fiscal year 2029 and raised its overall non-handset revenue target to $40 billion.

These are ambitious goals. Until a few months ago, the risk was that the data center market would remain mostly a promise. The deal with Amazon gives that strategy a customer with one of the largest cloud infrastructures in the world, along with an economic incentive tied to the growth of purchases.

This does not mean Qualcomm has already achieved its targets. The warrant itself highlights the phased nature of the relationship: a significant portion of the value vests only as orders and payments materialize. But that is precisely why the agreement matters. It does not just gauge the enthusiasm of the two companies; it ties Amazon's equity stake to concrete commercial results.

For AWS, on the other hand, the logic is almost mirror-image. The more AI demands grow, the riskier it becomes to depend on a single architecture or vendor. Amazon can continue buying massive quantities of Nvidia GPUs, develop Trainium in-house, and simultaneously commission custom silicon from Qualcomm. The three strategies are not mutually exclusive: they reinforce each other by increasing bargaining power and cloud flexibility.

Custom silicon is becoming an industrial policy for the cloud

The phenomenon goes beyond Amazon. Major cloud operators are trying to bring an increasing share of hardware design in-house or under their control. The reason is simple: when spending tens of billions on infrastructure, even a few percentage points of efficiency are worth massive sums.

Custom silicon makes it possible to eliminate unneeded features, optimize memory and interconnects for specific workloads, and integrate hardware and software more deeply. In exchange, it increases complexity: designing chips is expensive, takes years, and carries manufacturing, software, and compatibility risks.

This is where a company like Qualcomm comes in. Amazon can define the requirements and maintain control over the service architecture without necessarily having to build every component in-house. Qualcomm brings expertise in SoC design, energy efficiency, connectivity, and system integration; AWS brings scale, real-world workloads, and the ability to deploy the outcome across a global infrastructure.

The agreement also includes an intriguing move in the opposite direction: Qualcomm plans to increase its use of AWS's AI infrastructure, including Amazon Bedrock, for electronic design automation workloads. In practice, while Qualcomm helps Amazon design cloud hardware, Amazon helps Qualcomm use the cloud to design new hardware faster. It is a cycle in which chip design and AI infrastructure become increasingly interdependent.

The 60 billion caveat

The headline figure is also the one that calls for the most caution. In the SEC filing, the 60 billion represents the maximum payments tied to the warrant's vesting conditions. There is no indication that Amazon has already entered into an irrevocable commitment to spend the entire amount.

This means running a headline simply saying “Amazon buys 60 billion in Qualcomm chips” would be misleading. The agreement creates a commercial runway that could reach that scale, but the gap between potential and actual revenue will depend on chip quality, delivery timelines, cost-effectiveness, and AI demand.

Then there are other elements that remain unknown: who will physically manufacture the chips, which process nodes will be used, how much memory will be integrated, what software stack will accompany them, and how open they will be to AWS customers versus internal use. These are details that will determine the real technological impact of the agreement.

The AI chip war enters its second phase

The first phase of the AI race was dominated by one question: who can get enough GPUs? The second will be more sophisticated: which combination of GPUs, custom accelerators, networking, memory, and software produces a token at the lowest cost while maintaining sufficient performance and reliability?

It is in this second phase that the Amazon-Qualcomm agreement becomes significant. Not because it crowns a winner, but because it shows that hyperscalers do not want a single hardware answer. They want a portfolio of solutions and the ability to shape silicon around their workloads.

For Qualcomm, it is an opportunity to transform from a mobile chip champion into an AI infrastructure provider. For Amazon, it is another lever to reduce costs and dependencies. For Nvidia, it is a sign that market growth does not eliminate competition: it enriches it, because every major customer has enough scale to finance alternative architectures without stopping their GPU purchases.

The key metric to watch in the coming months will therefore not be the theoretical value of the warrant. It will be the initial orders, the products actually deployed, and the share of AWS inference workloads running on silicon co-designed with Qualcomm. That is where it will become clear whether this deal is merely a major strategic commitment or the beginning of a new balance of power in the AI chip market.

Sources