On two consecutive days, Google and Amazon announced tools that directly challenge Nvidia's dominance over the AI development ecosystem. On Tuesday, October 20, Google made TorchTPU public, a version of PyTorch optimized for its own TPU chips. On Monday, the 19th, AWS had already released TorchNeuron, which does the same for the Trainium chip.
Both tools address the same problem: PyTorch, used by 85% of AI developers according to Google, targets Nvidia GPUs by default. This creates a network effect that makes it hard for any other chipmaker to compete, even with superior hardware. TorchTPU and TorchNeuron break that coupling by allowing PyTorch models to run natively on TPU and Trainium, respectively.
Nvidia had been challenged on hardware by startups like Tenstorrent, which raised $1.5 billion on Monday, and by hyperscalers developing their own chips. But the true moat was always software: Nvidia's CUDA and native PyTorch compatibility. Google and Amazon now target exactly that point, offering portability without code rewrites.
For those operating AI systems in production, the news changes the cost and vendor calculation. Until today, migrating workloads between Nvidia GPUs and alternative chips required retraining models or manually adapting code. With TorchTPU and TorchNeuron, the switch can become a decision purely based on cloud capacity and price. The paper believes software portability is the most relevant move to break the AI hardware monopoly since PyTorch itself emerged.
The software competition now opening puts pressure on Nvidia to justify its prices through performance, rather than ecosystem lock-in. For engineering teams, the recommendation is to test compatibility of models in production with both tools before the cloud offerings that support them consolidate.