Qualcomm Pays Amazon in Stock for a Place in the AI Data Centre

On September 8, 2026, Qualcomm said it had won Amazon as a data-centre customer in a "multi-generational" collaboration to design custom silicon for artificial-intelligence inference and high-speed optical connectivity inside AWS. The headline number attached to the deal was $60 billion. The number that will actually show up in Qualcomm's accounts is smaller, and the way it gets there says a lot about how the AI build-out is being financed.

Related reading: TSMC Posts a Sixth Straight Record Quarter as AI Demand Outruns Supply · Taiwan Plans to Make Unauthorised AI-Chip Exports to China a Crime.

The Qualcomm press release described work across "multiple generations of customized silicon," focused initially on inference (running trained models rather than training them) plus optical links scaling to 1.6 terabits per second, drawing on the SerDes and optical DSP technology Qualcomm acquired with Alphawave. Qualcomm also said it will deepen its own use of AWS, running electronic-design-automation workloads on Amazon Bedrock to shorten chip-design cycles.

The mechanism is in the regulatory filing. As Reuters and CNBC reported, Qualcomm issued Amazon's investment arm a warrant for up to 25 million of its own shares at $161.26 each (a notional $4.03 billion) expiring in September 2036. The shares vest in tranches tied to commercial arrangements, binding purchase orders and actual purchases, up to a maximum of $60 billion in payments. About 3.75 million shares vested on issuance against initial commitments. Qualcomm's stock rose roughly 4 percent on the news.

The $60 billion is a ceiling, not an order

The most common misreading of the announcement is that Amazon committed to buy $60 billion of Qualcomm chips. It did not. The 8-K filing language is specific: the warrant keeps vesting as Amazon places purchase orders, "up to a maximum amount of $60 billion in payments." That figure caps the incentive schedule. It is the point at which the warrant stops rewarding further purchases, not a floor under Amazon's spending. No purchase obligation is disclosed.

Jon Peddie Research made the same point in plainer terms: the $60 billion is the top of the vesting ladder, not contract value or guaranteed backlog. The distinction matters because a genuine $60 billion order would be roughly 1.4 times Qualcomm's annualised revenue arriving from a single customer. A $60 billion vesting ceiling is a statement about how far the equity incentive reaches before it is exhausted.

Read correctly, the structure is a familiar one. Qualcomm is paying Amazon in future equity to choose its silicon. If the partnership works, Amazon buys chips and Qualcomm's dilution tracks the resulting revenue. If it does not, Amazon walks away holding an unvested instrument, and Qualcomm has spent nothing but potential dilution. The seller, not the buyer, carries the risk, which is why the structure has become standard in AI infrastructure. AMD granted OpenAI a warrant over up to 160 million shares at a nominal strike in October 2025 on the same template, and Marvell issued Google a warrant worth up to about $12.2 billion in August 2026.

What each side gets

For Qualcomm, the prize is a flagship reference customer. The company is the dominant supplier of smartphone processors, but it has been trying to build a data-centre business as its handset market matures and Apple moves to in-house modems. Its current portfolio includes the Dragonfly C1000 server CPU and the AI200, AI250 and AI300 inference accelerators, along with high-bandwidth compute and 800-gigabit and 1.6-terabit connectivity. Qualcomm has said it is targeting more than $15 billion in data-centre revenue by fiscal 2029 and completed the acquisition of Modular in July to strengthen its software stack. A multi-year AWS relationship is the credibility (and the production scale) that a challenger needs.

For Amazon, the deal widens an already broad menu of compute architectures. AWS runs Nvidia and AMD accelerators, its own Trainium and Inferentia AI processors, Graviton CPUs and Nitro infrastructure silicon. Amazon's custom-chip business passed a $25 billion annualised run rate at the end of the June quarter, and the company has been explicit that it wants to match silicon to workload rather than standardise on one vendor. Adding Qualcomm reinforces both technical flexibility and negotiating leverage. Neither company has said how the Qualcomm products relate to the Trainium and Inferentia roadmaps, and nothing ships until 2028 at the earliest.

The optical angle deserves as much attention as the compute. Linking thousands of accelerators into one training or inference cluster is increasingly a data-movement problem, and Qualcomm's Alphawave acquisition now looks like a deliberate bet that interconnect (not raw compute) is where the next bottleneck sits. Amazon validating that thesis with purchase-linked warrants gives it weight.

Why inference is the prize

The deal's emphasis on inference rather than training is a deliberate reading of where AI spending is heading. Training a frontier model is a concentrated, episodic expense borne by a handful of labs; inference (running models in response to users) is a recurring cost that grows with every product and every query. As models proliferate across search, code assistants, customer service and devices, the economics of serving them start to dominate the total cost of ownership, and serving costs are largely a function of performance per watt and memory bandwidth.

That is the battle Qualcomm is entering. It claims roughly twice the performance per watt for its Dragonfly C1000 server CPU and four to eight times the memory bandwidth per watt for its AI300 accelerators, figures that will have to be demonstrated on production workloads before they mean much. The company's pitch to hyperscalers is that inference can be done more cheaply on specialised, power-efficient silicon than on the general-purpose accelerators that dominate training. Nvidia still controls the overwhelming majority of the AI accelerator market, and its own inference products are improving quickly, so Qualcomm is not claiming to displace it. It is claiming a place beside it, another option in an infrastructure stack where AWS already runs Nvidia and AMD parts alongside its own Trainium and Inferentia chips.

For Amazon, suppliers are leverage as much as technology. Every credible alternative accelerator gives AWS more room to negotiate on price and to match silicon to workload, and its own custom-chip programme gives it a fallback if a supplier underperforms. The risk for Qualcomm is timing: nothing from this collaboration is expected to ship before 2028, by which point Nvidia, AMD and Amazon's in-house teams will each have shipped at least two more generations. Qualcomm's equity is riding on a roadmap that has to stay competitive for years before it earns a single dollar.

The accounting twist

There is a detail that will surface in Qualcomm's financials before any chip does. Under US generally accepted accounting principles, equity granted to a customer is generally treated not as a marketing expense but as a reduction of the transaction price. Accounting rules codified in 2019 direct companies to measure share-based consideration payable to a customer under share-based-payment rules and then reflect it against revenue. AMD states the treatment plainly in its own filings: the grant-date fair value of the warrant is recognised as a reduction to revenue as revenue is recognised.

The practical effect is that reported revenue arrives net of the incentive. If Qualcomm recognises a dollar of sales to Amazon under this arrangement, the top line it reports is that dollar minus the amortised fair value attributed to it, so both revenue and gross margin come in below what the gross purchase figures imply. Qualcomm has not yet disclosed how it will classify this particular warrant, and its fiscal year ends September 27, 2026, so its annual report is the first document likely to carry the number. Investors who track the partnership through purchase announcements alone will be looking at the gross side of a net picture.

What to watch

Three things will test the arrangement. First, whether Qualcomm's silicon reaches AWS production on schedule, a 2028 horizon gives Nvidia, AMD and Amazon's own Trainium team two more product generations to respond. Second, whether the optical and inference silicon is genuinely custom and competitive, or a second source that AWS uses mainly as negotiating leverage. And third, how much of the $60 billion vesting ladder actually gets climbed: the warrant structure gives both sides an incentive to grow volumes, but no obligation to do so, and the disclosed initial commitment covers only 15 percent of the shares.

The larger read is that AI infrastructure is being financed with a widening array of instruments, from debt and prepayments to equity-linked customer warrants. That lowers the cash cost of building for the vendor, but it also binds buyers and sellers more tightly and pushes more of the risk into share counts and future revenue. The Qualcomm–Amazon deal is not an outlier; it is the latest sign that the AI boom is increasingly funded through the balance sheet and the share register, as much as through the purchase order.