Amazon Just Made Leaving Nvidia a One-Line Change. Here's Who Should Care.

Amazon released a free tool that lets AI developers move their code from Nvidia chips to Amazon's own Trainium chips by changing one setting. Amazon stock jumped 5%. The headline is about Nvidia's moat, but the bigger number is buried in Amazon's last report: its chip business already runs at more than $25 billion a year.

HeyTheo Research · Tuesday, October 20, 2026

Quick Read

• Amazon (AMZN) launched TorchNeuron on Oct 19, a tool that lets code written in PyTorch, the most common AI framework, run on its Trainium chips with a one-line setting change. It goes open source on Oct 20.

• That targets Nvidia (NVDA)'s strongest defense: CUDA, the software that makes switching away from Nvidia chips slow and costly.

• Less noticed: Amazon's chip business passed a $25 billion yearly revenue run rate last quarter, growing at triple-digit rates, with multi-gigawatt Trainium commitments from OpenAI and Anthropic.

• Amazon reports third-quarter results on Oct 29. AWS growth has sped up for four straight quarters, to 37%.

What Amazon launched

Most AI models are written in PyTorch, a free, open-source framework. Amazon says about 85% of AI applications use it. But PyTorch has always run best on Nvidia chips, because Nvidia's CUDA software sits underneath it. Moving to another chip meant rewriting and retuning code, which costs time and engineers.

TorchNeuron removes most of that work for Amazon's chips. A developer changes one setting to tell the code which chip to use. It supports Amazon's second- and third-generation Trainium chips for both training models and running them.

Source: Amazon, TechCrunch, SiliconANGLE, Motley Fool

One early customer, the voice AI company Deepgram, says it processes 70% more audio per dollar after moving to Trainium. That is the pitch: not the cheapest chip, but more work per dollar.

Research cue · try it in HeyTheo

Ask Theo for the bull case, bear case and what to watch on Nvidia, and add: "how much of Nvidia's valuation depends on customers being locked into CUDA?" It frames how much a working alternative could matter to the stock, even if orders hold up. Open HeyTheo

Start with: search NVDA, then ask the question above.

The number most coverage skipped

The launch got the headlines. Amazon's last quarterly report holds the context.

AWS is speeding up. Growth rose from 17% a year ago to 37% in the second quarter of 2026, on $42.2 billion of revenue. That is four straight quarters of faster growth at a business of that size.

The chip business is already large. Amazon said its chips business and its AI business on AWS each passed a $25 billion yearly revenue run rate, both growing at triple-digit rates. OpenAI and Anthropic have each signed multi-year Trainium commitments measured in gigawatts.

The order book is huge. AWS had about $496 billion of signed commitments not yet recognized as revenue at the end of June, roughly double the end of 2025.

TorchNeuron is how Amazon fills that capacity with its own chips instead of Nvidia's. Every workload that runs on Trainium keeps more of the margin inside Amazon.

Research cue · try it in HeyTheo

Build two baskets: the companies building their own AI chips (Amazon, Alphabet, Meta) and the companies that mostly rent out Nvidia capacity. Track them against each other through earnings season. If the first group holds up better, the market is starting to price in-house chips as an advantage. Open HeyTheo

Start with: search AMZN for the first basket and NVDA for the second.

Rather read first? Our free daily brief lands in your inbox before the open. Send it to me

What it means for Nvidia

Not much in the next few quarters. Demand for AI chips still outruns supply, and TSMC (TSM) said last week that it cannot make enough. Large customers will keep buying Nvidia's best chips.

The risk is slower and shows up in price, not volume. Nvidia earns premium margins partly because leaving is hard. If even part of the work can move with one setting, its biggest customers gain leverage in every negotiation. Alphabet (GOOGL) is building a similar tool for its own chips, so the pressure comes from more than one side.

It also matters for the companies that rent out Nvidia chips. AWS pitches TorchNeuron to startups partly as a reason to avoid clouds that are tied to Nvidia hardware, like CoreWeave. As we covered in our note on the AI trade's two ledgers, those operators are already under pressure from their borrowing costs. And as we laid out in our TSMC note, the chipmaker gets paid whichever design wins.

The bull and bear case

Amazon, bull case. AWS growth is accelerating, the order book is near $500 billion, and a fast-growing chip business means more of each AI dollar stays in-house. TorchNeuron lowers the biggest barrier to using it.

Amazon, bear case. It plans to spend about $200 billion this year, and returns have to show up. A beta tool does not move deeply customized code, and customers may keep paying for Nvidia's best chips regardless.

Nvidia, bull case. Demand still exceeds supply, its newest chips lead on performance, and much of the industry's most advanced code is written directly for CUDA.

Nvidia, bear case. Its largest customers are all building their own chips and now the software to move work onto them. Even a modest shift weakens its pricing power over time.

Research cue · try it in HeyTheo

Set an earnings trigger on Amazon for Oct 29: "tell me what Amazon says about Trainium and its chips business." Then use the backtester to see how the stock moved after its last four reports. It shows how much a strong AWS quarter has moved the stock before. Open HeyTheo

Start with: search AMZN, then set the earnings trigger.

Research it yourself: rules HeyTheo tracks

• In-house chips vs renters. Two baskets of AI names, split by whether they design their own chips, tracked against each other.

• Growth acceleration trigger. Flags any holding whose main segment grows faster for three quarters in a row.

• Earnings keyword trigger. Flags when a company's results mention a product or business you follow.

• Ask Theo. Try: "Which of my holdings depend most on Nvidia's pricing power?" or "Give me the bull and bear case on Amazon before its Oct 29 report."

Check the rule behind any trigger before acting on it. You trade through your own broker; HeyTheo helps you decide.

The Switching Cost

Nvidia's lead was never only about chips. It was about how costly it is to leave. Amazon just made leaving cheaper for most AI code, and it did so with a chip business that already earns more than $25 billion a year.

Nothing changes for Nvidia's orders this quarter. What changes is the conversation in every pricing negotiation. Watch what Amazon says about Trainium on Oct 29, how fast developers adopt TorchNeuron, and whether Nvidia's margins start to give.

Start your own first pass. Download HeyTheo, add the AI names you hold, and run the three cues above. Get the app

Reading on a laptop, or not ready to install? Get the free daily brief by email, and open the app on your phone when you are ready. Get the daily brief

Sources

Amazon, TechCrunch, SiliconANGLE, Motley Fool, Drillr, Capp Notes, MarketScreener

Disclaimer

Disclaimer: HeyTheo is a research and education platform, not an investment adviser or broker-dealer. Nothing here is advice to buy, sell, or hold any security. You trade through your own broker; HeyTheo helps you decide. Backtested results are hypothetical and do not guarantee future returns. References to governments, officials, or policies are for market context only and are not political endorsements. All investing involves risk, including loss of principal. Data is as of the dates noted.