1 change
Amazon and Google open PyTorch to their own AI chips
Video coming soonAmazon says moving PyTorch code onto its own AI chips now takes one configuration change. Amazon and Google both open-sourced tools this week that run PyTorch on their own chips.
What happened
Amazon open-sourced TorchNeuron at the PyTorch Conference on October 20. It runs existing PyTorch code on Amazon's Trainium2 and Trainium3 chips. Amazon says that takes a single configuration change and no rewrite. It says it tests TorchNeuron against hundreds of thousands of AI models on every build. Amazon is also contributing changes to PyTorch itself to reduce chip-specific code. On the same day, Google open-sourced TorchTPU, a backend that runs PyTorch on its TPUs. Its first public release is version 0.2.0.
Read the transcript
One configuration change. That's what Amazon says it takes to move PyTorch code to its own chips. This week, Amazon open-sourced a tool called TorchNeuron. It runs PyTorch on Amazon's Trainium chips, with no rewrite. Amazon says it tests it against hundreds of thousands of models on every build. It is also sending changes to PyTorch itself to cut chip-specific code. On Tuesday, Google open-sourced its own version, TorchTPU, for its TPU chips. Google's first public release is version 0.2.0, so it is early. We think the real target is Nvidia's software lock-in. If code moves with one setting, chips compete on price and supply. The bigger question is whether these tools work as well in practice.
That's your byte. Sources at sixtybytes.com.
Sources
Related
Morning brief · Wednesday, October 21, 2026
SpaceX now makes more from AI than from Starlink
SpaceX's AI unit brought in $7.6 billion last quarter, more than Starlink, out of $14.6 billion in total revenue.
Deep-cut · Tuesday, October 20, 2026
Nscale's IPO filing: $103.4 billion contracted, $2.6 billion switched on
Only about 2.5% of Nscale's contracted business was active in August, its IPO filing shows. Some investors now want a lower valuation.
Morning brief · Tuesday, October 20, 2026
Samsung will reportedly make part of Nvidia's next memory chips on TSMC's 3 nm process
A Seoul newspaper says Samsung will use TSMC's 3 nm process for the base die of Nvidia's custom memory. No company has confirmed it.
Narration is AI-voiced. Every story is human-checked.