Amazon Built Its Own Chips to Escape Nvidia — Then Just Tripled Its Nvidia Order

Amazon Built Its Own Chips to Escape Nvidia — Then Just Tripled Its Nvidia Order

I've been running hosting infrastructure for over a decade, and I've learned to read hyperscaler purchase orders the way other people read tea leaves. So when Amazon — the company that spent billions building its own AI chips specifically so it wouldn't have to depend on Nvidia — announced Wednesday that it's adding another 2 million Nvidia GPUs to its data centers, I had to sit down. This isn't a footnote in the AI buildout. It's the most honest demand signal we've gotten all year, and it's coming from the one company you'd least expect to write it.

Let me tell you something straight: when the world's biggest cloud provider — the one with its own silicon division, its own custom chips, and a CEO who keeps talking about customer choice — walks onto Nvidia's earnings call and orders 2 million more of your competitor's chips, the story isn't about Amazon and Nvidia. It's about every independent hosting provider, every colo operator, and every startup waiting on GPU allocation. Because that order just reshaped the supply curve for the next two years.

The Announcement — Two Million More GPUs, Two Years Out

Here's what happened. During Nvidia's Q2 fiscal 2027 earnings call on Wednesday, AWS and Nvidia announced a major expansion of their strategic collaboration. AWS will deploy 2 million additional Nvidia GPUs across its global infrastructure in 2027 and 2028. The chips span the next three generations of Nvidia silicon — Blackwell Ultra, Rubin, and Rubin Ultra. Neither company shared financial terms, but at current unit costs, we're talking tens of billions of dollars. TechCrunch did the math and called it what it is: Amazon just tripled its order.

Context matters here. Five months ago, at GTC 2026, AWS committed to deploying more than 1 million Nvidia GPUs starting this year. Nvidia's statement on Wednesday was blunt: "demand has exceeded those expectations." That 1 million already got eaten, and they're coming back for two more. The expanded deal also includes 100,000 GPUs for AI factories serving the U.S. government at Impact Level 6 classification — one of the highest security tiers in the federal stack — plus new collaboration across CPUs, networking, open models, data processing, and robotics.

And this lands inside an earnings report that was already absurd. Nvidia posted $96.2 billion in quarterly revenue, up 106 percent year over year. The data center segment alone did $89 billion, up 117 percent. Guidance for next quarter is $108 billion, above every consensus estimate. Gross margins held at 75 percent. The stock still slipped after hours — that's now five straight post-earnings declines, because apparently a company growing revenue 106 percent isn't exciting enough anymore. But the AWS order is the number that actually matters, and it wasn't in the revenue line. It was in the future.

The Part That Stings — Amazon Can't Quit Nvidia Either

Now let me give you the part that should make every AI company think twice about its chip strategy. Amazon has been building its own silicon for years precisely to reduce its dependence on Nvidia. Trainium, its custom AI chip, has crossed a $25 billion annualized revenue run rate. Amazon says it has $225 billion in total commitments from AI labs — including Anthropic and OpenAI — tied to its custom chip business. Its Arm-based Graviton CPU is a legitimate challenger to Intel and AMD in the server market. Its AI chief, Peter DeSantis, has said AWS is in talks to sell Trainium chips to other companies for their own data centers. This is the most aggressive custom-silicon program of any hyperscaler on earth.

And despite all of that, Amazon just tripled its Nvidia order. Read that again. The company that built the most credible alternative to Nvidia in the industry just committed tens of billions more to Nvidia. That tells you two things. First, demand is so far ahead of supply that even a $25 billion custom-chip program can't fill the gap — Amazon's own silicon isn't enough for what its customers are asking for. Second, the integration is deeper than the marketing lets on. AWS is now working with Nvidia's NVLink Fusion interconnect technology in next-generation Trainium chips, and Annapurna Labs — Amazon's chip division — is tapping Nvidia's custom high-bandwidth memory technology. Amazon isn't just buying Nvidia chips; it's wiring its own chips into Nvidia's rails. If you can't beat them, plug into their scale-up fabric.

This is the part that stings for anyone who thought custom silicon would break the Nvidia grip. It hasn't. It's made the grip stronger, because now the two companies are co-dependent at the silicon level.

The Two Readings — Demand Running Ahead of Every Forecast, or the Most Concentrated Order Book in History

There are two ways to read this deal, and both of them are true. That's what makes it so uncomfortable.

Reading one: the demand is real, and it's accelerating. AWS is the most price-disciplined buyer in the cloud industry — it does not spend tens of billions on speculation. Its customers' reservations now extend into 2028. Frontier labs, enterprises, and governments are all in the queue. When the world's largest cloud provider says it needs 3 million Nvidia GPUs in a two-year window and its own chips can't cover the overflow, the "is AI demand real?" debate is officially over. It's real enough that the most conservative buyer in the business just tripled down.

Reading two: this is the most concentrated order book in the history of computing, and that's a systemic risk wearing a growth story. One customer — one — just locked millions of GPUs across three product generations. Add the government's 100,000 and the earlier commitment, and a meaningful slice of Nvidia's entire forward allocation for 2027-2028 is spoken for by a single company. That's great for Amazon's competitive position. It's less great for everyone downstream. And let's not pretend hyperscalers don't over-order. They've done it before. Announced is not delivered, and a commitment signed on an earnings call has to survive fab yields, HBM supply, power constraints, and the next 24 months of a market that has already surprised everyone.

So, ent? Both readings. The demand is real, and the concentration is real. The problem isn't that Amazon is buying. The problem is that Amazon buying 2 million means the rest of us are looking at what's left.

The Secondary Bottleneck Nobody's Talking About — The Stack, Not the Chip

Here's the part that keeps me up at night, and it's not the GPUs. Watch what else moved in this deal, because Nvidia just turned a chip purchase into a platform takeover.

Nvidia is bringing its Vera CPUs to AWS — some integrated with Rubin, others standalone. Nvidia's CFO, Colette Kress, said on the call that Vera will be deployed by "every major hyperscaler, neocloud, AI lab, and system OEM," with shipments already underway to lead partners including Oracle and SpaceXAI. Jensen Huang has been claiming a "brand-new $200 billion TAM" for Vera since May. That's not a chip announcement. That's a direct assault on Intel and AMD in the one market they still owned — the general-purpose CPU — and it's riding into every data center on the back of a GPU order.

Then there's the networking layer. AWS and Nvidia are deepening collaboration on high-speed interconnect — Elastic Fabric Adapter, NVLink Fusion, Nvidia's scale-up architecture — so that Trainium and Nvidia GPUs can share a common rack-scale fabric. Nvidia's memory technology, NVHBM, is going into Amazon's own chips. And on the robotics side, Amazon is adopting Nvidia's full physical AI stack — Omniverse, Cosmos, Isaac, Jetson — to power its warehouse robot fleet.

Do you see what's happening? The bottleneck used to be "can I get a GPU?" Now it's "can I get a GPU, the CPU that feeds it, the fabric that connects it, the memory that fills it, and the software that runs it — all from the same vendor?" Nvidia is packaging the entire AI infrastructure stack into one dependency. The GPU was the hook. The platform is the lock. For every independent operator, that's the real constraint to plan around: not a chip shortage, a stack shortage.

What This Means for Independent Hosting Providers

First: read this as an allocation signal, not a demand story. GPU supply for independents does not loosen in 2027-2028. The biggest buyer on earth just locked two more years of forward allocation. If you're planning capacity on the assumption that the market will open up, you're planning against the wrong curve. Lock your supply relationships now.

Second: watch the CPU layer. If Vera is going to every major hyperscaler, the x86 market is about to get squeezed from the top. That could mean aggressive pricing from Intel and AMD as they defend share — which is an opportunity for you — or it could mean they prioritize their own hyperscaler accounts, which is a risk. Either way, don't sleep on the CPU supply picture for the next 18 months.

Third: calibrate announced against realized. A 2 million GPU commitment is a promise, not a shipment. Track actual delivery milestones — quarterly volumes, memory availability, power contracts. The gap between announced and delivered is where the real market information lives, and it's the gap that has burned more than one operator this cycle.

Fourth: position as the independent alternative. When hyperscalers lock supply into their own clouds, the overflow demand — the customer who wants portability, the workload that needs a human being on the phone, the pricing that doesn't come with a migration to someone's ecosystem — still has to go somewhere. That somewhere is you. Build the story now: independent, flexible, not an extension of one vendor's platform.

Fifth: watch AWS cloud pricing. Two million more GPUs and a $25 billion Trainium run rate give AWS enormous capacity to push pricing in the market. But that capacity came at enormous capex cost, and hyperscalers don't eat capex quietly. Expect the pricing pressure to arrive in both directions: aggressive on the front end to fill the machines, quietly higher on the back end once the dependency is built.

The Bottom Line

Amazon built its own chips to escape Nvidia, and then handed Nvidia the biggest order in its history. That's not a contradiction. That's the reality of a market where demand is running ahead of every forecast and the integration is too deep to unwind. Nvidia isn't just selling chips anymore. It's selling the rails, and even the most independent-minded company on earth just decided to ride them.

For the rest of us, the lesson is simple: the AI buildout just told us who owns the supply curve, and it's not us. Plan for tight allocation, watch the CPU layer, calibrate the promises against the deliveries, and be the alternative the giants can't be. The stack is getting locked. Your job is to stay outside the lock — and stay useful to the people who want out.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: About Amazon (AWS/NVIDIA press release), NVIDIA Newsroom, TechCrunch, TechTimes, AI Business, Cloud Computing News, 24/7 Wall St., Yahoo Finance.