More from Atlas this week. This report pairs with our Meta Q3 2026 earnings read, which covers Muse serving economics, the capex pace and how Meta is funding the build. Atlas subscribers can also read the rest of this week’s coverage:
Texas Instruments Q3 2026: Analog Turns Into a Capacity Story, and Our Models Are Behind It.
Lam Research Q3 2026: A Constrained $200 Billion, and NAND Gets Its Own Trade Ratio.
SpaceX Q3 2026: The Newest Hyperscaler Prices Its Watts at Spot.
Micron’s AI Memory Roadmap: Bandwidth Moves Onto the Compute Die, and Wafers Stay Tight.
On October 19, Meta published a detailed engineering account of Prometheus, its gigawatt-class training cluster in New Albany, Ohio, and the same day we sat down with Santosh Janardhan, who leads Meta’s infrastructure, and Ian Buck, who runs NVIDIA’s hyperscale and HPC business. We wanted to understand what it took to build, and more importantly what the two companies think it was for.
Meta did not have one building large enough for what it wanted, so it joined together several structures (scale across), and wrote the software to make them look like one cluster to its researchers. It trained models while it was still adding capacity, and it had to drive down a failure rate that, by our arithmetic, meant something broke every few seconds across a large job at its mid-2025 peak. Teams from both companies rotated through Ohio for months to trace those failures to their source, and Meta says it has since cut the interruption rate by 99%.
The engineering is impressive, and what we find most interesting is the economic argument underneath it. Our models have power holding the energized accelerator fleet to roughly 72% of what demand calls for by 2030. When watts are that scarce, the amount of useful training a company gets out of each one becomes a competitive variable in its own right, and Meta has now published enough to measure it. The payoff shows up in products: a company that gets more research out of the same power can train a better frontier model, adapt it into more products, and distill it into smaller models that are cheaper to serve. Meta’s third-quarter results, reported this week, show both sides of that model. Mark Zuckerberg pointed to Muse Wonder, Meta’s first frontier model and the latest trained on Prometheus, as evidence of “velocity across the board,” and more than 10 million people in the United States and Canada set up a Muse agent in its first six weeks. On the other side, capital spending of $34.9 billion in the quarter pushed free cash flow below zero.
Key insights from the conversation and Meta’s disclosures:
Meta reports goodput, the share of training time that produces useful progress, above 85% for pretraining its Muse Wonder model and above 90% for reinforcement learning. Mean time between interruptions is above 54,000 GPU-hours in the latest runs, against a target of about 100,000.
On a gigawatt-class site, ten points of goodput can deliver substantial energy savings, but the gain does not translate one-for-one into power savings. Depending on operating conditions, the equivalent power benefit could be roughly 50–75 MW, and no operator can bring new power at that scale online quickly in the current market.
Reliability carries a hardware cost. Meta runs large training jobs on 64 of the 72 GPUs in each NVL72 rack and keeps the rest as spares, so roughly 11% of the GPUs in each rack are there for resilience. This enables both optimal power-of-two GPU sizes and resilience for fast failure recovery.
The same cluster that trains the frontier model also produces the product versions of it. Meta post-trains one model differently for agents, coding and the assistant, then distills it down, and as Janardhan put it, “you have to create a bigger model before you distill it down.” Training capacity sets how many of those products Meta can build and how cheaply it can serve them.
What Meta learned does not reset with each cluster. GB300 picked up on the reliability curve where GB200 left off, and Janardhan described starting over with every new cluster as “a fundamental disadvantage.”
We think NVIDIA benefits more than most investors assume, because the fixes that came out of Prometheus land in its platform for every customer at the same time our models have custom silicon gaining unit share.
The full analysis sizes what goodput is worth in dollars on our own unit costs, explains why we think the Street does not model it yet, and works through what Prometheus means for Meta, NVIDIA, Arista and Broadcom, including where Meta’s lead stays in-house and where it spreads to rivals through NVIDIA.
Companies covered in this report: Meta Platforms (META), NVIDIA (NVDA), Arista Networks (ANET), Broadcom (AVGO), CoreWeave (CRWV)
What subscribers get in the full report:
What ten points of training goodput is worth in dollars on a gigawatt-class site, built from our own shell and accelerator unit costs, and why it is the cheapest capacity a power-capped operator can add.
Our 2030 power cap, with how much of the accelerator fleet we think gets energized, how many accelerators end up waiting for power, and why that makes useful compute per dollar of capex the number to watch.
Why the Street does not model goodput yet, and what we would track in its place as Meta spends $135 to $145 billion this year.
The product case in three parts (better models, more of them, and cheaper ones to serve), tied to Muse adoption and Meta’s serving-cost targets from its third-quarter results.
What owning training and serving capacity does to unit costs, with a heavy personal-agent user priced at public API list rates against our modeled cost to serve.
Meta’s energized capacity path through 2030 from our DC Capacity Tracker, including Prometheus and Hyperion, and why the reliability lessons carry to each new site.
Company by company views on Meta, NVIDIA, Arista and Broadcom, including why we think NVIDIA gains more from Prometheus than most investors assume, and where Meta’s lead spreads to rivals.
Three exhibits: Prometheus reliability from the mid-2025 peak to Meta’s target, the power cap and the value of goodput, and the owned versus rented cost of serving an agent.
Go deeper in Atlas. Subscribers get the Atlas companion packet for this report, with the capacity, power, accelerator share and networking tables behind it, Meta’s reported reliability metrics, and the full set of attributed quotes from our conversation. Ask the Atlas Agent to work with any of it, or start from one of these prompts:
