Andy Jassy dropped a number in his April shareholder letter that should have every chip analyst paying attention: Amazon’s custom AI chip business, if it were a standalone company, would have an annual run rate of $50 billion. That’s roughly the size of Intel’s entire revenue. And now, for the first time, Amazon is actually talking about selling those chips outside its own cloud.
AWS AI chief Peter DeSantis told Bloomberg this week that the company is in early-stage talks to sell its Trainium chips to other companies for use in their own data centers. No specific buyers were named, but the message is clear: Amazon wants a piece of the market that Nvidia has dominated for years.
What’s Actually Happening
Here’s the short version. Amazon designed its own AI chip called Trainium — a custom ARM-based accelerator built specifically for training and running machine learning models. Until now, you could only use Trainium by renting compute time on AWS. Amazon never sold the chips directly.
That’s about to change. DeSantis confirmed that demand for Trainium is so strong that Amazon is exploring external sales. The company says selling chips to third parties won’t cannibalize AWS revenue — they believe the market is big enough for both.
This isn’t a casual pivot. AWS has resisted selling chips externally for years because of what industry folks call the “waterfall effect.” When a company runs Trainium on AWS, Amazon doesn’t just charge for the compute — it also makes money on storage, security, networking, and monitoring. Selling the chip outright means giving up that bundled revenue stream. The fact that they’re reconsidering tells you two things: the demand is overwhelming, and the competitive pressure from Nvidia’s pricing is real.
The GPU Tax Is Getting Hard to Stomach
Here’s the thing about Nvidia’s dominance that doesn’t get talked about enough: it comes with a premium. When every AI lab, every cloud provider, and every startup needs the same GPUs, Nvidia can name its price. And they have. The company is currently on a $326 billion revenue run rate — most of it from AI chips that cost more than some companies’ entire IT budgets.
For companies training large models, the math is brutal. A single training run for a frontier model can cost hundreds of millions of dollars, and a huge chunk of that goes straight to Nvidia’s margins. That’s why every major tech company — Google with its TPUs, Microsoft with Maia, and now Amazon with Trainium (similar to what Google is doing with its Coralboard for edge AI) — is building custom silicon. Not because they want to be chip companies, but because they can’t afford to be entirely dependent on one supplier.
OpenAI, Anthropic, and Apple have all reportedly standardized portions of their workloads on Amazon’s Trainium 3 chips. The reason is straightforward: 40% lower total cost of ownership compared to Nvidia’s general-purpose GPUs for transformer-based models. When you’re spending billions on compute, that 40% adds up fast.
Why This Matters Beyond Amazon and Nvidia
The real story here isn’t just about two companies competing. It’s about the end of what analysts are calling the “GPU monoculture.”
For years, if you wanted to do serious AI work, you bought Nvidia GPUs. Period. There was no real alternative. That’s changing fast. Google’s TPUs have been handling inference for years internally. Microsoft designed Maia specifically for Azure workloads. And now Amazon is saying its chips are good enough to sell to anyone.
This shift has three big implications for developers and enterprises:
First, pricing pressure. When Amazon starts selling Trainium externally, Nvidia can’t charge whatever it wants anymore. Competition drives prices down — that’s basic economics. For Filipino developers and startups using cloud services, this could eventually mean lower inference costs and more affordable AI compute.
Second, chip diversity. Not every AI workload needs the same hardware. Training a massive language model is different from running inference on a production API. Custom chips optimized for specific tasks can be more efficient than general-purpose GPUs. More chip options means better performance-to-cost ratios for different use cases.
Third, supply chain resilience. Relying on a single chip manufacturer is a risk that enterprise IT managers understand well. Remember when chip shortages during the pandemic shut down everything from car production to gaming console launches? Diversifying the AI chip supply across multiple manufacturers — Nvidia, Amazon, Google, Microsoft — reduces that systemic risk.
The TSMC Bottleneck
There’s a catch, though. Amazon can design the world’s best AI chips, but someone still has to manufacture them. And that someone is almost always TSMC — the Taiwanese foundry that makes chips for most of the semiconductor industry.
Here’s where it gets interesting: Nvidia recently supplanted Apple as TSMC’s largest customer. That means Nvidia has more bargaining power with the manufacturer than anyone else. Amazon wants to scale Trainium production for external sales, but it’s competing for fab capacity against a company that’s ordering more chips than anyone on the planet.
This is the same dynamic that has played out in every chip generation. Design is only half the battle — manufacturing capacity is the other half, and right now, Nvidia has the upper hand there too.
The Bigger Picture: Silicon Sovereignty
What Amazon is doing fits into a larger trend that goes beyond just competing with Nvidia. Nations and corporations are increasingly looking at AI chips as a matter of strategic independence. India, the UAE, and several European countries are demanding that AI infrastructure be built on their soil, using hardware they can control.
When a company owns the chip design, the compiler, and the model stack, it gains a level of control that no software-only company can match. Amazon’s willingness to sell Trainium externally could eventually enable what some call “regional silicon sovereignty” — allowing countries and companies to run AI workloads on custom hardware without depending entirely on a single American chipmaker.
Like I explored in my analysis of Bezos’s Prometheus raising $12B for physical AI, the race to control AI hardware is heating up across the board. Amazon selling chips externally is the latest move in a game where the stakes keep rising.
What Happens Next
These talks are still in early stages. No deals have been announced, and Amazon hasn’t named any buyers. But the direction is unmistakable. When your CEO publicly says your chip business could be worth $50 billion, and your AI chief starts talking to Bloomberg about selling chips externally, you’re past the “exploring” phase.
The question isn’t whether Amazon will sell Trainium chips outside AWS. It’s how fast they can scale production to meet demand — and whether Nvidia’s TSMC advantage will slow them down.
For developers, this is good news. More competition means more choices, lower costs, and less dependence on a single vendor. The GPU monoculture that defined the AI era is cracking, and Amazon just pushed a crowbar into the gap.
As I covered in my piece on AI’s energy consumption problem, custom chips designed for specific AI tasks are inherently more energy-efficient than general-purpose GPUs. Amazon’s push into external chip sales could accelerate that efficiency gains across the industry.
And for anyone running infrastructure on AWS — which includes a lot of Filipino developers and startups — keep an eye on Trainium pricing. If Amazon starts offering competitive external rates, expect AWS to adjust its cloud pricing too. That’s the kind of ripple effect that benefits everyone downstream.
Also worth watching: Tesco just ditched VMware after Broadcom’s price hike, proving that enterprise customers will walk when vendor pricing gets too aggressive. Nvidia should take note.
The AI chip war just got a new combatant. And this one has $50 billion worth of motivation.