AWS Fires Its Anti-CUDA Weapon: One Config Line Moves PyTorch Apps onto Trainium
AWS released TorchNeuron, running PyTorch AI apps on its Trainium chips with no code rewrite — a direct hit on CUDA.
AWS on Monday released TorchNeuron: AI applications can run PyTorch models directly on Amazon's in-house Trainium chips by changing a single line of configuration — no code rewrite required. The move strikes at Nvidia's software moat: CUDA.
Attacking the software lock-in, not the chip
For years, Nvidia has ensured CUDA works seamlessly with PyTorch, making its hardware "the default you never have to choose." Amazon's number: roughly 85% of AI applications are written in PyTorch. Levering that layer means levering Nvidia's software base.
Colin Brace, VP of Technology Software at AWS Annapurna Labs (the unit behind Trainium), told TechCrunch that Trainium's value is price-performance plus the fact that one chip handles both training and inference, tightly integrated with the cloud: "the infrastructure to go from prototype to production is already there."
His example is speech company Deepgram: after moving to Trainium's native PyTorch support, it processes 70% more audio per dollar at production-grade accuracy.
Amazon also took a swipe at Nvidia-tied "neoclouds" like CoreWeave and Nebius: rather than betting on a single-purpose compute provider, move into AWS.
Why this matters now
The CUDA moat was never really the chip — it was the migration cost. Making the move a one-line configuration change puts Nvidia's pricing power on a comparison sheet. Deepgram's 70%-more-per-dollar number is the kind of concrete figure CFOs notice.
This lands alongside other fractures in Nvidia's position: Tenstorrent just raised $1.5 billion for open AI silicon, and open-weight models keep improving. The compute market is fragmenting from a single default into a menu of options.
What this means for your business
If you're running AI workloads on AWS — or planning to — this lowers the cost of experimentation. The practical move: when you next negotiate cloud compute or sign credits deals, benchmark Trainium pricing against GPU instances for your actual inference workloads, using the Deepgram 70% figure as the starting question.
More broadly, design for portability. Workloads written in standard PyTorch without CUDA-specific dependencies can now move between Nvidia GPUs, AWS Trainium, and emerging alternatives with minimal friction. That portability is negotiating power — against cloud providers and against hardware vendors.
For startups and SMBs, the message is even simpler: inference costs are dropping through competition, not just through better models. If you've been postponing an AI feature because GPU costs looked prohibitive, re-run the numbers this quarter. The price of intelligence is falling on two fronts at once — and vendors competing for your workload is the best time to be a buyer.