Meta published an engineering account of Prometheus in New Albany, Ohio on October 19, 2026. The company says its first phase is complete and was used to train Muse. It describes the current installation as hundreds of thousands of GPUs, while roughly one million is the target at full expansion. Those are different project states.
Useful work is more than installed hardware
Meta reports over 85% goodput for Muse Wonder pre-training and over 90% for reinforcement learning. These are company measurements, not independently reproduced SIExpo results. The account does not disclose an exact current GPU total, overall site consumption or final project cost. A description of gigawatt scale should not be converted into an invented measurement of instantaneous power demand.
Our explanation of the engineering issue is straightforward: an accelerator provides value when it contributes to useful work. Waiting, interruption and repeating lost steps can reduce output without a device being permanently broken. A comparison based on GPU count therefore needs measurements of the actual workload and recovery process alongside the hardware inventory.
Buildings need to function as one system
Meta’s earlier February technical account describes Backend Aggregation as an Ethernet layer connecting separate network fabrics. It discusses alternative topologies, path redundancy and constraints involving distance, buffering and latency. This is project background, not a newly launched October networking product. Keeping the dates separate avoids making an existing architectural design sound like a fresh commercial release.
The practical connection is coordination in distributed training. When one portion of the computation waits for data from another, an individual chip’s peak capability does not describe the speed of the complete job. This is our explanation of the dependency, not an additional benchmark. For an organization assessing its own cluster, we would examine communication delays, downtime and recovered work alongside processor utilization.
Closed-loop cooling does not establish zero consumption
Meta’s separate August explainer describes a water-and-glycol mixture circulating through equipment and heat exchangers before returning to the servers. It also discusses Air-Assisted Liquid Cooling for buildings without the corresponding built-in infrastructure. The material explains a principle; it is not an independent water audit of Prometheus or a complete measurement of every facility involved.
Our conclusion is to preserve measurement boundaries. Reusing coolant in one circuit does not by itself establish a campus’s total water or energy impact. Such a conclusion requires a reporting period, an explicit system boundary and published measurements. We do not transfer results from a separate pilot to every building or assume that a cooling description resolves all environmental questions.
For an infrastructure comparison, a useful record would identify installed capacity, the work completed during the measured period and the interruptions excluded from a headline result. That is our proposed assessment framework. It does not imply that Meta has published all those details, nor does it substitute an estimate for missing operational figures.
The account matters because it places networking, heat and recovery alongside compute. Assessing the next expansion will require actual commissioned capacity and measured useful work. The million-GPU figure remains a stated direction, while the completed first phase is the current result reported by the company.
Editorial verification: 2026-10-19T16:25:17.203539+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.
