🏗️ Meta (~1M GPUs) — the supercluster already training Muse Wonder
Meta publishes the details of Prometheus: hundreds of thousands of GPUs already running in Ohio and the plan to reach ~1 million, with more than 85% goodput in pre-training.
Tenstorrent closes a $1.5B Series E at a $6.2B valuation —more than double 2024— to scale manufacturing of its Galaxy RISC-V chip servers.
Anthropic is negotiating a second ~$45B deal with Nscale for another 460 MW at the Monarch campus in West Virginia, using Vera Rubin chips.
The debate is no longer whether AI creates value. It is who controls the chips and the megawatts.
🏗️ Meta (~1M GPUs) — the supercluster already training Muse Wonder
On October 19 Meta’s engineering team published the deep-dive on Prometheus, its gigawatt-scale supercluster in New Albany, Ohio. Phase one is complete: hundreds of thousands of NVIDIA GPUs (GB200 Catalina and GB300 Clemente NVL72) interconnected across multiple buildings that act as a single machine. They have used it to train the Muse models, including Muse Wonder. The full-scale target is ~1 million distributed GPUs.
They report more than 85% goodput in pre-training and more than 90% in reinforcement learning, a 99% reduction in interruptions (from ~1,000 events per 8K GPUs per day to very low targets) and MTBI above 54,000 GPU-hours. The Backend Aggregation network spans more than 100 km. Cooling is closed-loop liquid. In short: Meta is no longer talking about “we are going to build”; it is measuring goodput in real frontier-model production.
🚀 Tenstorrent ($6.2B) — Jim Keller doubles down against Nvidia
On October 19 Tenstorrent announced the close of a $1.5B Series E at a $6.2B post-money valuation, more than double the ~$2.6B of late 2024. Trace Capital and Innovation X led; ARK, Fidelity, T. Rowe Price, Morgan Stanley, Sandisk and others joined. Jim Keller —ex Apple, AMD, Tesla— told Bloomberg they chose to raise capital instead of selling after demonstrations of their Galaxy servers running DeepSeek models.
The money goes to scaling manufacturing and deployment of Galaxy (32 Blackhole RISC-V chips with Tensix cores), to the CPU and AI IP they license, and to hiring. The bet is clear: open hardware, cheaper to customize and usable both in the cloud and on-prem for open-weight model inference. After years of “Nvidia or nothing”, a second player with $6.2B and Keller’s track record starts to be credible for certain workloads.
💰 Anthropic ($45B) — the second check to Nscale for Vera Rubin
Bloomberg and Reuters report on October 19-20 that Anthropic is in advanced talks with Nscale for a second agreement of approximately $45B. It would be a second building at the Monarch campus in West Virginia, another ~460 MW, powered by NVIDIA Vera Rubin chips that start arriving in late 2027. The first deal (August) was already ~$45B / 460 MW / six years. Anthropic would have a right of first refusal on a third building.
Nscale (backed by Nvidia) has an enormous IPO pipeline backlog concentrated on Microsoft and Anthropic. A second commitment of this magnitude confirms that frontier labs are signing multi-year contracts of tens of billions just to secure capacity. The debt and energy-delay risk is real —Microsoft already walked away from one site over connection problems—, but Anthropic seems willing to double down before the IPO.
The connecting thread
Meta shows the cluster it already has measuring goodput on frontier models. Tenstorrent raises $1.5B to offer an open alternative to Nvidia GPUs. Anthropic is about to sign another $45B just so it does not run out of megawatts. The model race has become a physical infrastructure race. Whoever does not have chips or energy contracts signed through 2027-2028 simply does not train at the same pace.