Google DeepMind’s new training stack matters because datacenter boundaries are turning into frontier bottlenecks. Decoupled DiLoCo trained a 12B Gemma model across four U.S. regions on 2-5 Gbps links, more than 20x faster than conventional synchronization while holding 64.1% average accuracy versus a 64.4% baseline.
HN did not read Google’s TorchTPU post as another cloud pitch. The real question in the thread was whether a PyTorch user can really switch to `tpu` without falling back into the old PyTorch/XLA pain cave.
HN treated TPU 8t and 8i as more than giant datacenter numbers. The thread focused on the bigger shift: agent-era infrastructure is splitting training and inference into separate hardware bets.
Anthropic said it has signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity that will begin coming online in 2027. The company framed it as its largest compute commitment so far, tied to surging Claude demand and a rapid jump in large enterprise customers.
Anthropic said on April 7, 2026 that it has signed a deal with Google and Broadcom for multiple gigawatts of next-generation TPU capacity coming online from 2027. The company also said run-rate revenue has surpassed 30 billion dollars and more than 1,000 business customers are now spending over 1 million dollars annually.
A Hacker News thread with about 240 points focused attention on Anthropic’s April 6 announcement that it signed for multiple gigawatts of next-generation TPU capacity with Google and Broadcom starting in 2027, alongside claims of more than $30 billion in run-rate revenue and over 1,000 seven-figure business customers.
HN picked up Nanocode, an open JAX project that packages tokenizer training, pretraining, synthetic data generation, agentic SFT, and DPO into an end-to-end recipe for building a coding model on TPU infrastructure.