Google 與 Amazon 令轉離 NVIDIA 只需修改一行設定

Forty-Eight Hours That Broke the Moat: Google and Amazon Made Leaving Nvidia a One-Line Change

At PyTorch Conference 2026, Google announced that TorchTPU is now an open-source project: a native PyTorch backend for Google's TPUs, co-developed with Meta's PyTorch team.
The critical design point is that existing PyTorch code targets a TPU through a standard torch.device("tpu") call, with no rewrite required. Migration stops being a project and becomes a configuration change.
The previous day, Amazon released the open-source framework TorchNeuron in public beta, under which PyTorch code runs on AWS Trainium chips without modification.
Observers described the sequence as the second cloud provider in as many days driving the cost of "leaving Nvidia" to the floor — two rivals digging in the same trench, turning "switch chips" from a rewrite project into a one-line config change.
The strategic backdrop: NVIDIA's dominance has rested far more on software than on silicon. CUDA is roughly two decades of accumulated libraries, kernels and developer muscle memory, and that accumulation — not the chip — is the moat.
In the same week, Meta detailed the Prometheus gigawatt-scale supercluster in Ohio: phase one is already built and has been used to train the Muse model, using GB200 NVL72 racks, with an ultimate target of roughly one million GPUs, effective throughput above 85 percent, and mean time between interruptions rising from 47,000 to more than 54,000 GPU-hours.
The same week also saw DeepSeek open-source its dual-chip inference engine DeepInfer, and reports that DeepSeek is co-designing future AI chips with Huawei, planning to pretrain on Ascend 950DT as early as the end of 2026.
Why This Is the Most Exciting Story of the Day
Here is the thing that should make every AI builder sit up: the hardest part of leaving NVIDIA was never the hardware. You can buy TPUs. You can buy Trainiums. What you cannot easily buy is twenty years of CUDA — the kernel libraries, the operator coverage, the Stack Overflow answers, the institutional reflex of a thousand engineering teams who reach for the same idioms without thinking. That accumulated software, not the chip, is the moat. And this week two of the largest cloud providers on earth started paving over it.

The elegance of the attack is in its modesty. torch.device("tpu"). That is the whole pitch. Nobody is asking a team to rewrite their training stack, re-validate their kernels, or re-hire for a new hardware dialect. They are asking them to edit a string. When the cost of switching collapses to a one-line config change, the strategic question stops being "can we afford to migrate" and becomes "why would we not benchmark the alternative." Once you benchmark the alternative, you have pricing leverage forever. That is a structural shift, not a feature release.

The genuinely delightful twist is who is doing it. Google and Meta — co-developing a PyTorch backend whose entire purpose is to reduce dependence on a third party. Meta, meanwhile, is simultaneously one of NVIDIA's largest customers, building the gigawatt-scale Prometheus cluster on GB200 NVL72 racks. Everyone is a frenemy with everyone. Google wants TPU utilization. Meta wants framework portability so it never gets cornered. Amazon wants Trainium volume. And PyTorch itself, the neutral ground, becomes the Switzerland through which all three interests flow. It is hard to imagine a more entertaining alignment of incentives.

The second-order effect is the one that matters most for ordinary consumers of AI. Portability is a bargaining chip. The moment switching costs approach zero, pricing power migrates from whoever owns the lock-in to whoever can offer the best dollar-per-token. That is good for every startup, every research lab, and every enterprise that has been quietly resenting its compute invoice. It is also good for the industry's resilience: a world where models can move between silicon vendors is a world where a single supply-chain shock cannot strand an entire generation of training runs.

Now the honest caveats, because "public beta" and "no rewrite required" are marketing-adjacent phrases. A demo that runs clean is not the same thing as sustained parity. The real test is kernel coverage and long-tail operator support — the obscure custom ops, the mixed-precision edge cases, the distributed-training corner cases that only surface at scale. Google and Amazon have not made NVIDIA irrelevant this week. They have made it contestable, which is a different and arguably more important achievement. The next six months of independent benchmarks will decide whether this was a crack or a demolition.

Source
PyTorch Conference 2026 / Google TorchTPU / AWS TorchNeuron / TechCrunch / The Verge / AI News Brief / AI Frontier Daily / threerouter