Amazon Backed Qualcomm's Custom Chips and Nvidia's GPUs at Once. Is the GPU's Grip Slipping?

Amazon spent the last two weeks placing two very different bets on how it will power artificial intelligence. On August 26, Amazon Web Services and Nvidia said AWS would deploy two million additional Nvidia GPUs across its data centers in 2027 and 2028, an expansion of a commitment the two companies had already sized in the millions. On Tuesday, Qualcomm announced that the same customer had signed a multi-year deal for custom silicon of its own — chips Qualcomm will design specifically for AWS to run AI inference, the everyday work of answering user queries rather than training models from scratch. Qualcomm shares jumped about 10% on the news, according to Bloomberg and CNBC reporting.

The two announcements sit awkwardly together. That is the point. One says the world’s largest cloud provider still cannot get enough of Nvidia’s hardware. The other says it is paying to design around it. The market has spent two years asking whether custom, purpose-built chips will erode Nvidia’s commanding hold on the AI data center, or whether that hold is too deep to dislodge. Amazon’s two announcements land on opposite sides of that question.

Every hyperscaler is now designing its own accelerator

The strongest evidence for the challengers is that Amazon is not alone, and the trend predates this deal. Google has run its own tensor processing units for years. Amazon builds Trainium and Inferentia chips in-house, and Microsoft and Meta have each committed to custom accelerators tuned to a narrow band of AI workloads. Industry analyses of the hyperscaler chip programs describe a consistent logic: for the inference jobs that dominate a live data center, a chip built for that one task can deliver a large cost and power advantage over a general-purpose GPU. The Qualcomm deal extends the pattern to a merchant supplier, giving a hyperscaler a second source without building everything itself.

The structure of the agreement shows how serious Amazon is. Qualcomm granted Amazon a warrant to buy up to 25 million Qualcomm shares at $161.26 apiece, worth roughly $4 billion at current prices, according to reporting from 24/7 Wall St. and Startup Fortune. The shares vest only as Amazon actually buys hardware, and unlocking the full stake would require Amazon to spend as much as $60 billion on Qualcomm chips and services through 2036. An initial tranche vested immediately. The vesting schedule ties Qualcomm’s payout directly to how much Amazon ends up buying.

Two weeks earlier, the same customer ordered two million more GPUs

Then there is the size of the Nvidia order sitting next to it. The two million GPUs AWS committed to on August 26 are Nvidia’s newest Blackwell Ultra and Rubin parts, and both companies framed the expansion as a response to demand that had outrun AWS’s earlier plans. Nvidia still holds an estimated 80% to 90% of the AI accelerator market, a position built as much on its CUDA software as on the silicon. A decade of tools, libraries, and trained engineers runs on CUDA, and rebuilding that ecosystem for a new chip is a multi-year undertaking most buyers have not attempted.

The Qualcomm deal is also narrower than the headline suggests. It targets inference, the cheaper and more repetitive half of the AI workload. Training the largest models remains dominated by Nvidia GPUs, where the flexibility of CUDA and the multi-chip scaling of Nvidia’s interconnects are hardest to match. Qualcomm’s own history is a caution. Its earlier push into data-center server chips, the Centriq line, was wound down years ago, and the company has described the Amazon agreement as its largest move into cloud infrastructure to date. Building the silicon is one thing. Producing it at the volume and reliability a hyperscaler demands is a separate test, one Qualcomm has yet to pass at scale.

What the data shows

The tape has rewarded the incumbent. Before Tuesday’s jump, Qualcomm had traded down about 6% over three months and roughly flat over the past month, and at $168.74 it sat well below its 52-week high near $260, with a market value of $185 billion. Nvidia has gained more than 15% over the same three months and closed at $230.36, within a few percent of its own 52-week high, carrying a market capitalization of roughly $5.5 trillion. Amazon, the buyer in both deals, is down about 3% over three months at a $2.8 trillion value.

The broad market shrugged. The S&P 500 sat at 7,699, little changed, and technology has been the strongest sector group of late. The gap in scale between a $185 billion supplier and a $5.5 trillion one is the backdrop to every headline in this contest.

What would settle it

The proof is in shipments, and the dates are on the calendar. Qualcomm has said its data-center chips will sample and reach customers over the next two years, so the first hard read is whether working silicon arrives on schedule and AWS deploys it in volume. Amazon’s re:Invent conference, held each year in early December, is where AWS details its chip roadmap and typically discloses how much of its own and partners’ silicon it is running. Nvidia’s next quarterly report will show whether data-center revenue keeps climbing even as its largest customers diversify. And the warrant is a running scoreboard: each tranche that vests marks another block of Qualcomm hardware Amazon has actually bought.

None of those milestones will crown a winner. They will show whether the custom-silicon wave is taking real share or filling a niche at the edges of a market Nvidia still defines. Amazon is buying both stories at once. For now, so is everyone else.

This is a general-market summary for information only — not investment advice, and not a recommendation regarding any security.