A megawatt of Nvidia B300 chips serving an open-source model at 60% utilisation brings in about $70 million a year in token sales, SemiAnalysis chief executive Dylan Patel said in a talk at Modal’s Runtime 2026 conference, posted on YouTube on 2 November. On his own cost figures, a $40 million build plus $9 million a year to run it for five years, owning that megawatt averages about $17 million a year, by our arithmetic.
That puts token revenue at about 4.1 times the cost of owning, and pays back the $40 million build in about 8 months, by our arithmetic on Patel’s numbers. Patel used the math to explain why GPU rental prices keep rising: “the reason for that is everyone wants to sell tokens.”
The source · Dylan Patel at Modal Runtime 2026
Dylan Patel, chief executive of SemiAnalysis, at Modal’s Runtime 2026 conference (video posted by Modal on 2 November 2026). The per-megawatt economics start at about 6:00; the Opus 5.5 remark is at about 8:50.
Business Pill · Capex and Depreciation
An evergreen Business Engineer explainer on capex and depreciation: why chip lifetime drives AI profits. It teaches the general idea only and does not discuss Patel’s talk.
The key insight: Dylan Patel put a megawatt of B300s serving DeepSeek V4.1 Flash at 60% utilisation at about $70 million a year in token revenue. On his own cost figures, owning that megawatt averages about $17 million a year, so revenue is about 4.1 times the cost and the $40 million build pays back in about 8 months, by our arithmetic.
Patel gave the cost of owning first. “It costs like $40 million to build the cluster, but to operate it over five years, it’s about $9 million a year,” he said, adding: “that’s about the cost if you build your own B300 cluster.” Modal’s description of the video sums up the section as “$9M/yr to run a megawatt, $70M/yr in tokens”, which ties the figures to one megawatt.
Renting is the other route. “If you’re renting it from the clouds, you know, you’re lucky if you’re paying $14 million a year for a megawatt,” Patel said. So $14 million is roughly the best case, not the typical price.
The build cost fits what Patel said in October. In our 25 October piece Strip Nvidia, Broadcom Margins and a Gigawatt Costs $20B, he put a gigawatt of AI data centre capacity at about $40 billion, rising to more than $60 billion. That is $40 million to more than $60 million per megawatt, by our arithmetic.
Per megawatt per year, in millions of dollars, as SemiAnalysis chief executive Dylan Patel laid out the economics at Modal’s Runtime 2026: owning a B300 cluster averages about $17 million a year ($40 million build plus $9 million a year to run it, over five years, our arithmetic); renting costs $14 million a year at best (“you’re lucky if”); serving DeepSeek V4.1 Flash at 60% utilisation brings in about $70 million a year (Patel). The ~$35 million at 30% assumes revenue scales with utilisation and is our arithmetic, as are the 4.1x and 2.1x ratios.
What a Megawatt Earns Serving Tokens
On the revenue side, Patel used DeepSeek V4.1 Flash, an open-source model. With “60% utilization”, which he defined as “only selling 60% of the tokens that you could generate”, he said: “you’re making $70 million a year.” He added: “the profitability is ridiculous.”
He said the figures come from SemiAnalysis’s open-source InferenceX benchmarks, built on the vLLM and SGLang serving software: “They’re containers you can download and just run.” The InferenceX site runs a profit estimator that, in its own words, “computes yearly revenue for each chip within a gigawatt all-in power budget at a chosen interactivity and utilization”.
Patel said lower usage still pays: “even if you say, oh, utilization is going to be 30%, you’re still going to make a ridiculous sum of money.” He gave no dollar figure for 30%. If revenue scales in step with utilisation, an assumption of ours, 30% would bring in about $35 million a year, still about 2.1 times the cost of owning, by our arithmetic.
The Per-Megawatt Math, Line by Line
Here is how Patel’s figures stack up per megawatt, by our arithmetic. Owning costs $85 million over five years, or $17 million a year on average. At 60% utilisation, the $70 million of revenue leaves about $53 million a year after that average cost, or about $56 million if the megawatt is rented at $14 million.
Payback is the starker number. Revenue of $70 million less $9 million of running costs leaves $61 million a year against a $40 million build, so the build pays back in about 8 months. At 30%, on our proportional assumption, payback takes about 18 months.
On the same assumption, revenue covers the cost of owning at about 15% utilisation and the $14 million rent at about 12%. Those break-even points are our arithmetic, not Patel’s figures.
A different set of assumptions puts break-even at 35%
Late on 2 November, Yingjun Wu, founder of RisingWave Labs, published an X article on the same model and the same benchmark family. On his inputs, an Nvidia B300 renting at $7.40 an hour and selling tokens at $0.0188 per million breaks even at 35% utilisation, and an H200 cannot cover its rent even when fully busy. We set his assumptions beside Patel’s in Nvidia B300 Break-Even: 35% at $7.40 vs 12% on Patel’s Data.
Why Patel Says Prices Keep Rising
Patel said the gap draws in small operators: “there’s like dozens of people that I know who their business is they’ve rented some GPUs and they’re serving inference and they’re making a ton of money.” He said running inference at small scale is “not that difficult”, but harder at the scale of a lab routing work across 25 clouds.
He said the profit on open-source models, not only lab demand, is why rental prices are climbing: “if you can find B300s for $6 an hour, you should sign it and you should sell tokens.” The baseline, he said, comes before the speculative decoding and other techniques that the strongest inference companies use.
Anthropic’s Opus 5.5 and the Training Split
Patel then turned to Anthropic. When it is serving Opus 5.5, he said, “it’s north of 100 million”, and he attributed the figure to materials Anthropic has been releasing to investors through its IPO process. He did not restate the unit. If it is per megawatt per year, like the $70 million, it is more than about 1.4 times the open-source figure, by our arithmetic.
The figure is not in the public filing. We searched Anthropic’s amended Form S-1, filed on 2 November, for “megawatt” and “gigawatt” and found no matches. The filing says “Revenue per dollar of compute has expanded meaningfully”, and, as we reported in Anthropic Gave Up About $600M of Q3 Revenue to Price Cuts, that Opus 5.5 cut cost per task by about 40% from Opus 5.
Earlier in the talk, Patel said the labs spend mostly on training. “All of Claude Code”, he said, “is served by a gigawatt of compute despite the fact that Anthropic has 4 gigawatts today.” That is 25% of Anthropic’s compute, by our arithmetic. In his words, labs are “actually mostly spending on training.”
As we read the talk, the per-megawatt view explains Patel’s pricing argument: when one megawatt of rented chips can be turned into several times its cost in token sales, the renter, not only the lab, bids up the price of the chips.
As we read the numbers, the gap is wide enough that utilisation matters less than it seems. On our proportional assumption, revenue covers the cost of owning at about 15% utilisation, a quarter of the 60% case Patel used.
Dylan Patel, SemiAnalysis, Runtime 2026
“there’s like dozens of people that I know who their business is they’ve rented some GPUs and they’re serving inference and they’re making a ton of money.”
Three Implications
FOR GPU RENTERS At Patel’s best-case $14 million a year, a rented megawatt at 60% utilisation leaves about $56 million a year after rent, by our arithmetic, which is the gap he says keeps rental prices climbing.
FOR BUILDERS The $40 million build is the big number, but on Patel’s figures it is recovered in about 8 months at 60% and about 18 months at 30% (our proportional assumption), by our arithmetic.
FOR READING LAB MARGINS Patel put Opus 5.5 at “north of 100 million” without restating the unit; if it is per megawatt per year, it is more than about 1.4 times the open-source figure, by our arithmetic. The public S-1/A does not state revenue per megawatt.
The framework reads, of the token: “It is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”
As we read it, Patel’s math joins two of those units: the compute a megawatt serves and the price each token fetches. The framework was published in May 2026 and does not discuss this talk.
The figures are Patel’s spoken estimates, quoted from the video’s auto-captions. Patel did not restate the chip for the $70 million line, or the unit for Opus 5.5. Modal’s site dates Runtime 2026 to 1 October in San Francisco, and the video does not give a recording date. We did not read model-level revenue numbers from the InferenceX site, and the 30% and break-even figures rest on our proportional assumption. We did not contact SemiAnalysis or Anthropic.
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.
Dylan Patel put a megawatt of B300s serving DeepSeek V4.1 Flash at 60% utilisation at about $70 million a year in token revenue, against about $17 million a year to own it and at least $14 million to rent it. By our arithmetic, that is about 4.1 times the cost of owning and an 8-month payback on the build.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Patel’s remarks are quoted from the auto-captions of Modal’s video of his Runtime 2026 talk, with caption errors in product names corrected (vLLM, SGLang, InferenceX). His figures are his and SemiAnalysis’s estimates, not company disclosures. Five-year averages, ratios, payback periods and break-even points are our arithmetic; the 30% and break-even figures assume revenue scales with utilisation.
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.
