Meta runs Nvidia’s B200 GPUs at a 750-watt power limit, below the 1,000 watts of the general B200 HGX solution, so that it can cool them with air, according to a paper by Meta and AMD engineers posted to arXiv on 15 October. AMD’s MI350X has a 1,000-watt deployed limit on the same air-cooled platform, the paper’s Table I shows.
By our arithmetic, Meta’s B200 limit is 25% below the standard figure, and the MI350X’s limit is 1.33 times the B200’s. As we read the paper, these are the two Meta deployments it compared on ten training models, where AMD was faster on 4 of 10.
A vertical news Short summarising this story. Figures come from the Meta and AMD engineers’ paper on arXiv; our arithmetic is labelled.
Business Pill · GPUs
A short Business Engineer explainer on GPUs. It teaches the general idea only and says nothing about this paper.
The key insight: By our arithmetic from the paper’s Table I, Meta’s 750 W limit for Nvidia’s B200 is 25% below the 1000 W of the general B200 HGX solution, and its AMD MI350X limit of 1000 W is 1.33 times the B200’s; the paper says the B200 is capped to enable air cooling.
The paper’s Table I lists a deployed power limit of 1000 W for the MI350X, 1400 W for the MI355X and 750 W for the B200. It says the MI355X is a 1400 W design using direct liquid cooling, while Meta uses the 1000 W air-cooled MI350X; the MI355X entry is listed as public cloud, the other two as Meta on-premises.
The caption states: “Meta’s deployment of B200 operates at 750 W, which is different from the general B200 HGX solution that operates at 1000 W”. It adds that Meta “operates both GPUs below their vendor-specified TDPs on the same air-cooled platform”.
The reason, in the paper’s words: “B200 is capped at a lower power limit to enable air cooling because its power density is higher than that of MI350X”.
Deployed power limit per GPU as given in Table I of the Meta and AMD engineers’ paper (arXiv 2610.22230, 15 October 2026), with the general B200 HGX figure from its caption. The 25% and 1.33x are our arithmetic. Chart by FourWeekMBA.
How This Fits the Throughput Result
Our first report on the paper covered its Table III: across ten Meta production models, MI350X training throughput was 101.2% of B200’s on the geometric mean, faster on 4 of the 10 models by our count.
As we read the paper, the chips in that comparison are the Meta deployments Table I describes, a B200 with a 750 W limit and an MI350X with a 1000 W limit; Table I itself lists limits, not the settings of each run. The paper lists “higher deployed power” among the MI350X’s principal advantages, alongside HBM capacity: 288 GB against the B200’s 180 GB, about 1.6 times as much, by our arithmetic. The B200 leads on scale-up bandwidth, 900 GB/s per GPU against 538 GB/s.
The authors write that “The B200 keeps an edge on performance-per-watt”, but that power is a small share of capital-dominated total cost of ownership, “so the MI350X’s lower TCO drives provisioning”.
Running Past the Rated Power
The paper describes the MI350X as a 1000 W open accelerator module whose specification permits short power excursions, “from 1.1 TBP over 20 ms up to 2.0 TBP over 20 s”. Meta’s platform, it says, supports average power of up to 1.36 TBP over a 50 ms window for a 1000 W module, about 1,360 W, by our arithmetic.
Tuning how the chip uses that headroom gave up to a 1.33 times gain in a synthetic microbenchmark, and in Meta’s production recommendation-model training it gave end-to-end performance gains of 3–6%, the paper says.
The Business Lever: Data Centre Cooling Sets the Chip Budget
As we read the paper, air cooling shaped the hardware set-up: the denser Nvidia part was held to a 750 W limit to fit air cooling, against a 1000 W limit for the AMD part, with both below their vendor-specified TDPs, the paper says. It adds that Meta runs a mixed fleet of Nvidia GPUs, AMD GPUs and its own MTIA chips “to reduce costs and mitigate vendor lock-in”.
It also cites “Meta’s deployment of hundreds of thousands of AMD GPUs” for inference, and says Meta started using the AMD MI300X for inference in 2024.
The Structural Read
As we read the paper, the power limit is a facility choice: Meta set it so both GPUs could run on the same air-cooled platform, the caption says.
As we read the throughput result, the 101.2% geometric mean was measured on Meta’s deployments, and Table I lists those deployments’ limits as 750 W for the B200 and 1000 W for the MI350X.
As we read the cost passage, the paper puts the B200 ahead on performance-per-watt but says capital, not power, dominates total cost of ownership.
The Meta and AMD engineers’ paper, Table I caption
“B200 is capped at a lower power limit to enable air cooling because its power density is higher than that of MI350X.”
Three Implications
FOR NVIDIAMeta set a 750 W limit for its B200 to fit its air-cooled platform, the paper says, against 1000 W for the general HGX solution.
FOR AMD The paper lists higher deployed power and 288 GB of HBM among the MI350X’s principal advantages.
FOR READING THE BENCHMARK The paper reports no B200 results at 1,000 watts, so the ten-model comparison at the standard limit is not shown.
The framework puts it plainly: “Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.”
As we read it, the Meta paper shows the power equation inside one comparison: cooling set the B200’s limit at 750 W. The framework is the owner’s argument and does not discuss this paper.
As we read the paper, its ten-model comparison was run on Meta’s deployments with these limits; the paper reports no B200 results at 1,000 watts, so how that comparison would look at the standard limit is not shown. Table I gives deployed limits, not measured average draw. The 25%, 1.33 times, 1.6 times and 1,360 W figures are our arithmetic on the paper’s numbers.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Meta and AMD engineers report that Meta runs Nvidia’s B200 at 750 watts, below the 1,000 watts of the general B200 HGX solution, to cool it with air, while its AMD MI350X has a 1,000-watt limit, 1.33 times the B200’s, by our arithmetic.
A note on sourcing. Power limits, memory, bandwidth and quotations come from the paper by Meta and AMD engineers (arXiv 2610.22230, 15 October 2026); the throughput figures come from our 30 October report on the same paper. The 25%, 1.33x, 1.6x and 1,360 W figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
