The two biggest cloud players in the world have just made the most audacious move against the monopolistic stranglehold of AI hardware.
In less than 24 hours, Amazon and Google announced radically new software tools that enable any developer to migrate their AI applications from Nvidia GPUs to proprietary custom chips with a simple configuration change — without rewriting a single line of code. It is the most coordinated attack ever seen against the advantage that has kept Nvidia as the industry's default choice.
Amazon was first, launching TorchNeuron on Monday (October 19), a tool adapted from PyTorch, the open-source framework used to build approximately 85% of current AI applications. With just a single line of configuration changed, developers can run their PyTorch models on Amazon's Trainium chips instead of Nvidia GPUs, and the company says the same Trainium chip is capable of both training and inference, improving economics across the entire workload journey.
According to Colin Brace, Vice President of Software at Annapurna Labs (Amazon's chip unit), more than 30 customers — including companies in robotics, world models, and voice AI — are already using TorchNeuron in production. Voice AI company Deepgram, for example, claims to have processed 70% more audio per dollar spent on Trainium while maintaining production-grade accuracy. Amazon is also contributing code directly to the main PyTorch project, with the explicit goal of decoupling framework features from specific chips, making PyTorch inherently hardware-agnostic.
The following day (October 20), Google responded with TorchTPU, an essentially equivalent tool enabling the same migration to the company's Tensor Processing Units (TPUs), through Google Cloud. But Google made an additional move: it worked directly with Meta, Hugging Face, Inferact (creator of vLLM), and RadixArk (creator of SGLang) to ensure TorchTPU integrates seamlessly with these popular AI development tools. The company also launched the TPU Agent Suite, a dedicated set of tools to help developers move their applications to Google's chips and cloud, fixing errors along the way. It further updated LiteRT, its tool for running AI models on mobile and embedded devices, to work better with PyTorch.
What is happening here is more than a battle between two companies for market share. It is a genuine attempt to dismantle what the industry calls Nvidia's "lock-in." For more than a decade, Nvidia invested heavily in CUDA — its software infrastructure that allows applications to run efficiently on Nvidia GPUs. CUDA became so deeply integrated into the development ecosystem that it became the path of least resistance: if you were building an AI model and wanted performance, the default was Nvidia. It was a perfect trap, built not through brute-force marketing but through incremental engineering that made developers willing hostages to a single hardware platform.
Now Amazon and Google are removing the bars. The reasoning is elegant: instead of competing directly with CUDA, they render it irrelevant. If PyTorch — the de facto language of modern AI — can run on any chip without rewriting, Nvidia's advantage disappears.
The landscape is complex because the AI industry will never be a zero-sum game. Startups like OpenAI have signed contracts with Nvidia, AMD, Broadcom, and other companies, maintaining a multi-supplier portfolio by design. The idea of everyone migrating to a single provider never worked in practice. But the point isn't that everyone abandons Nvidia tomorrow — it's that the narrative of "there is no alternative" is dying. And that fundamentally shifts the bargaining power of the only supplier that has controlled the market for years.
For Amazon, TorchNeuron is an opportunity to break dependence on neoclouds like CoreWeave and Nebius, which bet everything on Nvidia hardware. Amazon offers something different: the entire cloud, with the same chip training and inferring, improving economics across the entire workload journey. For Google, it's a chance to sell more TPUs to external cloud providers — The Information reported that Google already approached neocloud companies about using its TPUs, and companies including Anthropic, Meta, and Apple already use TPUs.
What remains to be seen is whether this chip war actually translates into real choice for developers or merely creates more configuration complexity. PyTorch exists precisely to hide hardware difficulties — and decoupling it from specific chips may, ironically, expose those very problems in ways that no framework ever wanted to face. The question isn't who will win, but whether the AI industry is prepared for an era where hardware-free suppliers will mean everyone will have to take more responsibility for what's running below the metal.
Sources: TechCrunch, Reuters, Business Insider
✓ Independent sources cross-checked and verified before publishing
