On a Tuesday morning, QualcommQCOM-- and AmazonAMZN-- announced a multi-generation collaboration to build custom silicon for AWS's AI data centers, and traders bid Qualcomm shares up roughly 10% before the open. The stock closed the day up about 3% at $174. The headline carries a number so large it reads as a done deal: Amazon has committed to order chips worth "up to $60 billion." Before anyone treats that as a line item on Qualcomm's income statement, it is worth reading what the agreement actually says, because the structure tells you how much of that $60 billion is guaranteed and how much is a reward for Qualcomm actually delivering.
Here is the mechanism hidden inside the press release. As part of the deal, Amazon received a warrant to buy as many as 25 million Qualcomm shares at $161.26 each — giving Amazon the right to acquire up to roughly $4 billion in shares — expiring in September 2036. That warrant does not vest all at once. It vests in tranches tied directly to Amazon's chip orders, contingent on the orders actually reaching that $60 billion figure. In plain terms, the warrant is not a signing check Amazon can flip for a profit tomorrow; it is an incentive contract that pays Qualcomm's customer only as Qualcomm ships product, year after year, across multiple chip generations for up to a decade.
"Up to $60 billion" is a ceiling that only becomes real as Amazon actually buys — and the warrant is the proof-it structure that makes that explicit.
Quick Backtesting Tool
Now weigh what Qualcomm must do to earn even a fraction of it, because the timeline is the second thing the headline hides. Qualcomm's data-center roadmap, which it calls Dragonfly, came out of a June investor day, where it announced inference accelerators called the AI200 and AI250, expected mid-2027, plus a server CPU (the Dragonfly C1000) that launches with Meta as the first customer and is not in production until 2028. The company's stated ambition is more than $15 billion in data center revenue by fiscal 2029. Qualcomm's current revenue is roughly $45 billion a year, so even hitting that target means data center becomes about one-third of the business — a transformation that depends on chips that do not exist at deployment scale yet. Amazon's orders under this deal, and the warrant tranches tied to them, begin flowing only after those parts ship and prove out in the field.
That brings up the third, more interesting question: why would Amazon — which already designs its own custom AI accelerators in Trainium and Inferentia, with about $20 billion in annualized chip revenue across its Trainium, Graviton, and Nitro lines — hand a chunk of that work to Qualcomm? This is where the deal is easiest to misread as a share-gain win over Nvidia. It is not that. Amazon is not rejecting Nvidia or abandoning Trainium; it buys all of the above. Qualcomm is being added as complementary capacity in the two areas where its engineering bends the right way: low-power inference, and the high-speed optical interconnects (up to 1.6Tbps) that keep memory traffic from bottlenecking when inference clusters scale across racks.
Qualcomm's pitch for that inference slice is total-cost-based, exactly the frame a hyperscaler's custom-silicon strategy lives on: it claims 4x to 8x better performance-per-watt than GPUs on memory bandwidth, using cheaper LPDDR memory instead of HBM to cut cost, at up to 8x more tokens per watt. Note the word "claims." None of those figures have been verified by independent teardown or benchmark yet — they are Qualcomm's own marketing numbers on paper, and inference economics in a real serving stack (utilization, batch size, software maturity) rarely match the spec sheet. Qualcomm knows the software is the weak link, which is why it bought Modular in July to build an inference software stack independent of Nvidia's CUDA.
The market is pricing the ceiling — a $60 billion, decade-long relationship — while the evidence only supports the floor: a credible, complementary inference and networking customer, whose mountable revenue is still years and multiple proofs away.
None of this is to dismiss the deal. It is genuinely the strongest customer signal Qualcomm's data-center effort has produced, and it gives the revenue target a real path rather than a slide-deck ambition. Amazon's decision to insert a warrant that pays only on delivered orders is, if anything, a vote of interest rather than a firm commitment to spend $60 billion. And the numbers that would change the story are observable and specific: whether AI200 and AI250 ship on schedule next year, whether independent benchmarks confirm even a meaningful fraction of the efficiency claims, and whether the warrant tranches vest as orders actually convert into revenue. Today's ~3% close prices the promise of the biggest prize in the chip trade; the evidence so far supports a real but slower, smaller, and harder-won opportunity than the headline number suggests.
