pytorch

Google is opening a new front in the AI infrastructure race, this time through software—not just hardware. ⚡

Google announced that TorchTPU is now open source at the 2026 PyTorch Conference, in collaboration with Meta’s PyTorch team. The project lets AI models run on TPU devices within PyTorch, alongside GPUs and CPUs, with only minimal changes to existing code.

The biggest development? 👇

🔹 Frameworks such as vLLM, SGLang, TorchTitan, VeRL, and Ray have integrated TPU support, expanding developers’ options for running models.

🔹 And with Hugging Face Transformers v5, deploying models on TPU is now possible by setting `device_map="tpu"`.

In my view, the real competition with NVIDIA won’t be decided by chip power alone, but by how easily developers can switch to other alternatives without rebuilding their systems from scratch. 🚀

Can open-source software support give Google a bigger share of the AI computing market?
$GOOGLB
$NVDAB
$METAB

#Google #Aİ #PyTorch #NVIDIA #Crypto

Is this move by Cambricon even more clever than NVIDIA?

I just saw the news: Cambricon has jumped straight into PyTorch’s highest tier, sitting down at the same table as Huang.

Don’t rush to cheer as a “homegrown hero” just yet—ask one question first: how hard is the gold in this seat? In the PyTorch ecosystem, those who can get invited to the table are basically heavyweights like NVIDIA, which already comes with built-in traffic. This time, Cambricon has truly fought its way in, no shortcuts.

In plain terms, competition in AI chips isn’t only about raw compute power—it’s also about ecological discourse power. For a while, we’ve always said domestic chips lack software adaptation. This time, it’s basically like installing a chair inside someone else’s house.

That said, we should also be realistic: it’s easy to get a seat, but whether developers will willingly choose your framework is the real test. After all, NVIDIA’s CUDA moat can’t be filled by getting one invitation.

In my view, this is a smart play—at least it shows Cambricon has learned to “enter the community first, then set the rules.” But don’t hype it too early; the road is still long. The potholes in the software ecosystem are ten times harder to fill than the hardware gaps.

So tell me: is this seat just a door-opener, or another “epic good-news” roller coaster?

#寒武纪 #PyTorch #AI芯片 #英伟达 #财经热度