For years, Nvidia’s moat has been software, not silicon. On Monday, Amazon took its most direct swing at that moat yet—and investors noticed.
On October 19, 2026, Amazon Web Services released TorchNeuron, a free, open-source tool that lets developers run code written in PyTorch—the framework used by an estimated 85% of AI applications—on Amazon’s Trainium AI chips with a single configuration change and no code rewriting. Until now, developers who wanted to use Trainium had to modify their code to call programming interfaces specific to Amazon’s chips, a process that could take months of engineering work.
The announcement sent Amazon shares up 5% in morning trading, with Morgan Stanley naming the company a “top pick”. But the significance goes far beyond a single stock move. TorchNeuron is Amazon’s most serious attempt to dismantle the software advantage that has kept Nvidia on top of the AI computing market—and it arrives at a moment when the entire hyperscaler industry is converging on the same strategy.
The CUDA Moat: Why Nvidia Has Been So Hard to Dislodge
Nvidia’s dominance in AI chips—estimated at 80-85% of the data center AI accelerator market by revenue—has never been purely about hardware. It’s about CUDA, the software platform Nvidia has spent more than a decade building. CUDA makes PyTorch code run seamlessly on Nvidia GPUs, and because so much of the AI industry’s software was built around it, adapting that code to other hardware has historically been painful, slow, and expensive.
That lock-in has made many companies reluctant to switch, even as Amazon, Google, and Microsoft have developed their own chips. As Colin Brace, vice president of technology software at Annapurna Labs, Amazon’s chip-design unit, put it: “If you know PyTorch, then you know Trainium”. That’s the promise TorchNeuron is making—and it’s a direct challenge to the assumption that Nvidia’s software ecosystem is an unbreakable barrier.
What TorchNeuron Actually Does
The mechanics are elegantly simple. TorchNeuron is adapted from PyTorch itself, so developers who already know PyTorch already know how to use it. It runs existing PyTorch code on Trainium chips with minimal changes, and Amazon says more than 30 early customers—spanning robotics, world models, and voice AI—are already using it.
But Amazon isn’t stopping at making its own hardware easier to use. The company says it is contributing changes upstream to PyTorch itself, pushing the framework to be less automatically centered on Nvidia. That’s a longer-term play: if PyTorch becomes more hardware-agnostic at its core, the switching cost to Trainium—and to any other alternative—falls for everyone.
The value proposition Amazon is pitching isn’t just ease of migration. It’s price-performance plus platform integration. Brace argues that Trainium’s advantage is that it handles both training and inference on the same chip, tightly integrated with the rest of AWS, which improves the economics across a full workload. AWS points to Deepgram, an AI voice company that moved to Trainium using native PyTorch and is now processing 70% more audio per dollar at production-grade accuracy.
The Decart Talks: $7 Billion for a Software Bridge
TorchNeuron is only part of the story. Amazon is reportedly in talks to acquire Decart, a startup that makes software to help developers move AI models built for Nvidia chips to alternatives like Trainium, in a deal that could value the company at around **$7 billion**. Decart was valued at nearly $4 billion in May, according to PitchBook data.
Decart’s technology helps optimize AI models for different processors, allowing developers to move workloads between chips more easily—a capability that becomes increasingly valuable as companies seek alternatives to Nvidia. Notably, Nvidia is an investor in Decart, and Decart already runs its own AI models on Trainium. The potential acquisition signals that Amazon isn’t just building software tools; it’s willing to spend billions to buy the bridge that makes switching away from Nvidia frictionless.
The Broader Pattern: Everyone Is Attacking the Moat
Amazon is not alone in this fight. The TorchNeuron release comes amid a coordinated push across the hyperscaler industry to loosen Nvidia’s grip.
Google is working on TorchTPU, a similar project to run PyTorch natively on its TPU chips with minimal changes, and has started selling TPUs directly to some outside customers. Microsoft said last week that it plans to run most of its AI services on its homegrown Maia chips rather than Nvidia hardware. And Amazon’s custom chip business is already growing at a triple-digit percentage, helping boost AWS’s overall growth rate to 28%—its best quarter in nearly four years. Trainium3 chips, which became available at the start of 2026, are nearly sold out.
The irony is that Amazon is simultaneously one of Nvidia’s largest customers. The company has said it will deploy more than one million Nvidia GPUs starting in 2026, even as it pushes Trainium. That dual strategy—buy Nvidia’s chips for today’s workloads while building alternatives for tomorrow’s—is the defining posture of the hyperscalers in 2026. They are not trying to replace Nvidia overnight. They are trying to reduce dependency, improve margins, and build leverage.
Why This Move Matters
TorchNeuron matters for three reasons.
First, it attacks the actual barrier. For years, the conversation about Nvidia’s dominance focused on chip performance. But the real lock-in has always been software. TorchNeuron doesn’t claim to beat Nvidia on raw specs; it claims to make switching away from Nvidia easy. If that claim holds up in the public beta, it changes the calculus for every AI startup and enterprise that has been stuck on Nvidia by default.
Second, it’s open-source and free. Amazon is releasing TorchNeuron as open-source software and contributing changes upstream to PyTorch. That’s a strategic choice: by making the tool free and embedding it in the PyTorch ecosystem, Amazon lowers the barrier to trying Trainium for developers who have no relationship with AWS. It’s a customer acquisition strategy disguised as a developer tool.
Third, it arrives at a moment of maximum pressure on Nvidia’s valuation. Nvidia’s earnings may face little immediate pressure—the company still holds the overwhelming share of the market, and demand for its chips remains insatiable. But credible alternatives give Nvidia’s customers more leverage and could pressure its valuation over time. As The Motley Fool put it, “Nvidia’s earnings may face little immediate pressure, but its valuation could be another story”.
Amazon’s TorchNeuron is
not a knockout blow. Nvidia’s CUDA ecosystem is deeply entrenched, and the vast majority of AI workloads will continue to run on Nvidia hardware for years to come. Amazon itself will keep buying Nvidia GPUs by the million.
But TorchNeuron is the most credible shot yet at the software advantage that has kept Nvidia untouchable. It makes Trainium easier to try, cheaper to adopt, and harder to ignore. Combined with the potential $7 billion Decart acquisition and the parallel efforts at Google and Microsoft, it signals that the hyperscalers are no longer content to be Nvidia’s customers. They want to be its competitors.
The AI chip war is no longer just about who builds the fastest silicon. It’s about who controls the software layer that determines which silicon developers choose. Amazon just made its boldest move on that layer—and Nvidia should be paying attention.
