What Happened
Markets & Semiconductors · Compute · Vertical Integration · Deal Briefannounced 26 Aug 2026 · record to 27 Aug 2026
Nvidia's biggest customer is also building the chips to need it less
AWS Adds 2 Million More Nvidia GPUs
Nvidia and Amazon Web Services have jointly announced a further 2 million Nvidia GPUs for AWS's data centres in 2027 and 2028, on top of the more-than-1-million agreed five months earlier at Nvidia's GTC 2026 conference — plus Nvidia CPUs, memory and networking technology, open models, data-processing software and a robotics stack, all moving deeper into AWS. Neither company has disclosed a price. The announcement landed on Nvidia's own quarterly earnings call, weeks after Amazon told its own investors its chip business had crossed a $25 billion run rate and held more than $225 billion in commitments for Trainium — the accelerator Amazon has built, by its own account, to compete with Nvidia's. What was actually announced, what each company's own numbers say, and what it means that both are true of the same relationship at once.
By the Numbers the GPUs committed, and what Amazon's own silicon has reached
- 2,000,000additional Nvidia GPUs heading to AWS in 2027–2028, per the joint announcement
- 3,000,000GPUs now committed to AWS in total, once March 2026's more-than-1-million is added in
- $225bnAmazon's own disclosed Trainium revenue commitments, from its 29 Apr 2026 earnings call
- $25bnthe annualised run rate of Amazon's own chip business, from a separate, later call on 30 Jul 2026
- 5 monthsbetween AWS's first 2026 Nvidia GPU commitment and this one
What's Confirmed, Graded each figure traced to its own call and its own date
| Standing | What | Detail |
|---|---|---|
| Confirmed | That Nvidia and AWS jointly announced the expansion | Nvidia's own Newsroom and AWS's own Newsroom published the same press release on 26 Aug 2026, naming the GPU counts, Vera CPU infrastructure, memory and networking work, open models, data-processing tools and the robotics stack; no financial terms appear anywhere in it. |
| Confirmed | That neither company has put a price on either GPU tranche | No dollar figure for either the March or the August GPU commitment appears in the joint press release or on Nvidia's own earnings call. A widely-relayed “tens of billions of dollars” estimate is TechCrunch's own, from GPU unit costs it does not show — not a figure either company has stated. |
| Confirmed | That Amazon is separately talking about selling Trainium outside AWS | Amazon's AI chief Peter DeSantis told Bloomberg on 18 Jun 2026 that AWS is in talks to sell Trainium chips to companies outside AWS; an AWS spokesperson confirmed the same to TechCrunch. As of the sources checked, this remains talks, not a completed sale. |
| Confirmed | That the $225bn and $25bn figures are three months apart, not one disclosure | Amazon disclosed “over $225 billion” in Trainium commitments on its 29 Apr 2026 call and “over $25 billion” in chip-business run rate on a separate call on 30 Jul 2026. The $225bn figure was not repeated on the later call. |
Timeline
March to August 2026, in Seven Moves the deal's own history, in order
- 16 Mar 2026AWS announces, at Nvidia's GTC 2026 conference, plans to add more than 1 million Nvidia GPUs starting in 2026.
- 29 Apr 2026On Amazon's Q1 2026 earnings call, Andy Jassy discloses more than $225bn in Trainium revenue commitments and says AWS will keep both a deep Nvidia partnership and a large chips business of its own.
- 20 May 2026On Nvidia's Q1 FY2027 call, Jensen Huang says the new Vera CPU has opened a “brand new $200 billion” market Nvidia has never addressed before.
- 18 Jun 2026Amazon's AI chief Peter DeSantis tells Bloomberg that AWS is in talks to sell Trainium chips to companies outside AWS.
- 30 Jul 2026On Amazon's Q2 2026 call, Jassy discloses the chip business has crossed a $25bn run rate, still growing at triple-digit rates, and again affirms the Nvidia partnership.
- 26 Aug 2026Nvidia and AWS jointly announce 2 million additional GPUs for 2027–2028, plus Vera CPUs, memory and networking work, and Nvidia's data-processing software and robotics stack moving further into AWS.
- 26 Aug 2026, same callNvidia reports record quarterly results on the same earnings call the partnership is announced on.
The Argument
Customer and Competitor, at Once why the same relationship holds both
Amazon is, at once, Nvidia's largest-scale customer and one of its most serious would-be competitors — and the 26 August announcement makes both truer at the same time, not one at the expense of the other. On the buying side, AWS has now committed to roughly 3 million Nvidia GPUs across two 2026 deals, and is bringing in Nvidia's Vera CPUs, its NVLink Fusion interconnect, its Nemotron open models and its robotics platform besides. On the building side, Amazon's own Trainium accelerator and Graviton server CPU exist expressly to reduce that same dependence: Amazon CEO Andy Jassy told investors on 29 April that AWS holds more than $225 billion in Trainium revenue commitments, and on 30 July that the chip business had grown its run rate to $25 billion — and on both calls said, in nearly the same words, that AWS would keep a deep Nvidia partnership and a large chips business of its own. Amazon's AI chief Peter DeSantis went a step further on 18 June, telling Bloomberg that AWS is in talks to sell Trainium chips to companies outside AWS — a move that would put Amazon's silicon directly into Nvidia's own customer base, confirmed to TechCrunch by an AWS spokesperson. Nvidia, for its part, is not merely selling more of the same GPUs; its own CPUs, networking and software are now moving into the infrastructure of the company building a rival to all three. AWS CEO Matt Garman frames the arrangement, in the companies' own joint statement, as being about customer choice: “Customers want the freedom to choose the best tools for their AI workloads, and they want confidence that everything works seamlessly together.” Nvidia CEO Jensen Huang, in the same statement, frames it as sixteen years of partnership deepening across “the full stack — GPUs, CPUs, networking, open models and software.” Both descriptions are true of the same relationship Amazon itself describes as also, separately, a competitive one.
Two Ways to Read It what favours each account, and what complicates it
| Reading | What favours it | What complicates it |
|---|---|---|
| Coexistence, not retreat | Both companies' own joint statement frames the deal as deepening co-engineering across the full stack; Converge Digest's own-day analysis reads the new GPU order as evidence Amazon is running Nvidia GPUs, Trainium, Graviton and Nitro together rather than choosing between them. | Amazon's own $225bn Trainium commitment figure and Peter DeSantis's confirmed talks to sell Trainium outside AWS both sit oddly beside “coexistence” if that word is read as meaning Amazon's own silicon is a secondary effort. |
| A serious, emerging rival | 24/7 Wall St's Alex Sirois reads Trainium's own scale — 1.4 million chips already deployed, $225bn in commitments, Trainium3 nearly fully subscribed — as having “crossed a threshold” into Nvidia's own direct-sales territory; Stocktwits frames Amazon, alongside Google and Meta, as a hyperscaler customer becoming a rival in the same week OpenAI unveiled its own chip. | Financial commentator Jim Cramer, quoted the same day, dismisses the pattern outright: “Every day I read about some chip that is superior to Nvidia. And every year I see no real competitors.” And the single largest data point in this record points the other way — Amazon's own 26 August order is for 2 million more Nvidia GPUs. |
What Others Add
What The Deal Actually Moves beyond the GPU count
Vera CPU
Nvidia's new CPU, built for agentic AI
- Jensen Huang called it a “brand new $200 billion” market on Nvidia's 20 May 2026 earnings call — Nvidia's own forecast of a market it has not yet captured, not a disclosed figure.
- On the 26 Aug call, Nvidia's own CFO Colette Kress said Vera shipments to AWS are already under way — “some integrated with Rubin, others standalone” — alongside earlier shipments to Oracle Cloud Infrastructure and SpaceX AI.
- Nvidia expects Vera to more than double its CPU revenue in fiscal 2028, positioning the company, in its own words, as “one of the world's leading server CPU suppliers.”
Beyond GPUs and CPUs
Networking, models and robotics move in too
- Nvidia's NVLink Fusion interconnect is extending to Nvidia's new custom high-bandwidth memory, NVHBM — giving Amazon's own Trainium chips, not only Nvidia's, access to it.
- Nvidia's Nemotron open models continue to run on Amazon Bedrock and SageMaker, alongside Amazon's own and third-party models.
- Amazon Robotics is adopting Nvidia's full physical-AI stack — Omniverse, Cosmos, Isaac and Jetson — for its warehouse robots.
- The two companies plan 100,000 GPUs on secure AWS infrastructure for U.S. government AI factories, cleared to the highest federal security level named in the announcement.
Named, on the Record who is saying what, and in what capacity
| Who | Says |
|---|---|
| Matt Garman | AWS's CEO, on the joint announcement: “That's why we've invested deeply with NVIDIA to make AWS the best place to run NVIDIA AI technologies…” |
| Jensen Huang | Nvidia's CEO, on the same announcement: “For 16 years, we have scaled NVIDIA computing in the cloud together. Now, we are expanding our partnership across the full stack.” |
| Andy Jassy | Amazon's CEO, on Amazon's own 29 Apr 2026 call: “We will always have customers who want to run NVIDIA on AWS. And we will also have a very large chips business ourselves.” |
| Jim Carroll | Editor and publisher, Converge Digest, the day of the announcement: the new order “does not represent a retreat from AWS custom silicon.” |
| Alex Sirois | Financial writer, 24/7 Wall St: Trainium's scale has “crossed a threshold” that “puts them in NVIDIA's direct sales lane for the first time.” |
Conclusion
Where This Stands what's settled, and what's still open
What the record supports
That Nvidia and AWS jointly committed to 2 million more GPUs for 2027–2028, on top of the more-than-1-million agreed five months earlier — a combined footprint of roughly 3 million — plus Vera CPUs, memory and networking work, and Nvidia software and robotics moving further into AWS. That Amazon's own chip business grew its disclosed figures from $20bn to $25bn run rate, and over $225bn in Trainium commitments, across two separate 2026 calls. That Amazon is separately, confirmedly exploring selling Trainium to companies outside AWS — at the same time it placed the largest GPU order of its own history with the company that arrangement would compete against.
What it doesn't yet
Whether Amazon's own silicon meaningfully dents Nvidia's business, or whether — as named commentators disagree, on the record, in the same week — rising total AI demand simply makes room for both to grow at once. No dollar figure for either GPU tranche has been disclosed by either company. And whether AWS's Trainium talks with outside buyers produce an actual, named sale remains, as of the sources read, exactly that: talks.
What Would Settle It the concrete tells
- A disclosed price for either GPU tranche. Neither the March nor the August commitment carries a dollar figure from either company; TechCrunch's own “tens of billions” estimate is not one.
- A completed, named sale of Trainium chips outside AWS. As of the sources read, this remains talks Peter DeSantis and an AWS spokesperson have confirmed, not a transaction.
- Amazon's next earnings call repeating, or not, the $225bn Trainium figure. It went unmentioned on the 30 July call; a fresh, larger figure or its continued absence would each say something the current record cannot.
- Nvidia's own future disclosures on how concentrated its hyperscaler revenue remains. A widening spread among named buyers would support the “plenty of demand for both” reading; continued concentration in a handful of giants, Amazon among them, would support the rivalry reading.