Key Highlights

  • Nvidia (NASDAQ:NVDA) faces a direct software challenge after Google and Amazon released native PyTorch backends for their own AI chips within about two days of each other in October 2026.
  • Amazon.com (NASDAQ:AMZN) open-sourced TorchNeuron, a native PyTorch backend for the Trainium chips used in its cloud data centres, under the Apache 2.0 licence on Tuesday, October 20, 2026.
  • Alphabet (NASDAQ:GOOGL) released TorchTPU, which lets PyTorch code run natively on Google's tensor processing units, the custom chips behind its own models and its cloud AI offering.
  • Nvidia shares still rose about 1.5% to 2.5% on Wednesday, October 21, 2026, which suggests the market did not treat the releases as an immediate threat to demand.

What Google and Amazon Released

Nvidia (NASDAQ:NVDA) has long relied on CUDA, its programming platform, as the default route from AI software to its chips. In October 2026 two of its largest customers moved to loosen that link. Amazon.com open-sourced TorchNeuron, a native PyTorch backend for its Trainium accelerators, under the Apache 2.0 licence on Tuesday, October 20, 2026. Alphabet released TorchTPU for its tensor processing units within about two days. PyTorch is the most widely used framework for building and training AI models. A native backend registers the chip as a standard PyTorch device, so a model written for Nvidia hardware can, in principle, target Trainium or TPUs by changing the device setting rather than rewriting code through a separate compiler stack.

How the CUDA Moat Works

The moat around Nvidia (NASDAQ:NVDA) is less the chip than the software stack around it. CUDA includes optimised kernels, which are the low-level routines that run each mathematical operation, plus libraries for networking, debugging and profiling built over almost two decades. Most PyTorch code reaches Nvidia hardware through those kernels, so switching chips has meant engineering time and performance risk. Native backends lower the cost of trying an alternative, especially for inference, where models are fixed and cost per query matters most. They do not remove the gap in tuned kernels, tooling and developer familiarity. Hyperscalers also gain bargaining power: a credible in-house option can lower the price they pay for Nvidia systems even if most workloads stay on Nvidia.

Risks and What to Watch

The first risk for Nvidia (NASDAQ:NVDA) is gradual share loss inside hyperscaler data centres, where Google and Amazon control both the software and the hardware choice. The second is pricing, since easier portability strengthens buyers in negotiations over next-generation systems. The third is AI developers such as Anthropic, which already trains on Trainium and TPUs and is reported to be preparing an initial public offering. Offsetting this, performance parity takes time, and PyTorch support alone does not match CUDA's tooling. Investors should watch Alphabet capex guidance and Microsoft results on Wednesday, October 28, 2026, Amazon results later in the season and Nvidia's fiscal third-quarter report in November for any comment on custom chips.

Where Things Stand

The answer is that Nvidia (NASDAQ:NVDA) faces a slow erosion risk rather than a sudden one, because native PyTorch backends make Trainium and TPUs easier to adopt without matching CUDA's depth. The releases mainly strengthen Google and Amazon in their own data centres and in price talks. The next catalysts are Alphabet and Microsoft results on Wednesday, October 28, 2026, Amazon results and Nvidia's November report. Adoption of new software is uncertain, and this is not investment advice.