DeepSeek Says a GPU-Hour Earns $38 While Rent Runs Near $8

DeepSeek says one of its serving clusters earns US$38 per GPU-hour. In the deployment Q&A it published on GitHub with its DeepInfer inference framework on 20 October 2026, DeepSeek wrote that a single cluster behind DeepSeek V4.1 Flash “currently serves 155 trillion tokens per day, with overall revenue of US$38 per GPU-hour (at valley pricing).” Separately, the median posted on-demand rate for an Nvidia B300 was $7.98 per hour in August 2026, according to AIMultiple’s GPU index as reported by Akash Network; DeepSeek does not say which chips the cluster uses.

The economics of renting GPUs and selling tokens are the subject of a post on X on 25 October 2026 by Deedy Das, a partner at Menlo Ventures, who wrote that “there is an absolute barrage of inference providers popping up that get to $1M/mo extremely quickly.” This piece sets Das’s analysis beside DeepSeek’s figure and our own October reporting on rising compute rents and falling token prices. The figures come from different sources and measures, so they are set side by side and not netted against each other.

This Week In The AI Supercycle · VIDEO

The Business Engineer’s explainer for the week of 18 to 24 October 2026: the companies selling AI’s parts say they are booked for years, while the companies paying for the build found money dearer. It was made before Das’s post and does not discuss it.

Business Pill · INFERENCE

A short explainer of inference: the cost paid on every AI answer. It teaches the general idea only and says nothing about DeepSeek.

What Das Wrote

Das’s post sketches the economics of companies that rent GPUs and sell open-model tokens. “Even if you rent on-demand B300s at $8/GPU-hr, many open source models (GLM 5.3, Kimi K3, DeepSeek 4.1 Flash / Pro) can get you >$10-20/GPU-hr at reasonable throughputs (100tok/s),” Das wrote. Das added that “your ability to be making gross margin relies on your utilization and how good your stack is.”

Das wrote that gateways such as OpenRouter and Vercel aggregate demand for these providers, and that “many Chinese open source model labs are now asking for rev share if you’re at a certain scale.” Das also wrote that “we’re seeing 10s of such companies” appear almost every week, with fast revenue growth “at low-ish gross margin,” and raising funding rounds on that revenue.

Das’s profile on X lists investments in Anthropic, OpenRouter and Modal, among others. OpenRouter and Modal are both named in the post. The figures in the post are Das’s own estimates, and the post names no provider’s revenue.

What Was Posted on X

Deedy Das, a partner at Menlo Ventures, posted this on 25 October 2026. The post sets out Das’s own estimates of what inference providers earn per GPU-hour on open models, names DeepSeek’s $38 off-peak figure, and lists the risks. This piece quotes the post and checks the DeepSeek figure against DeepSeek’s own document.

An interview clip posted by @HyperTechInvest on 26 October has investor Gavin Baker telling Jensen Huang that DeepSeek “said that they are monetizing at $38 per GPU hour”, and that he had not heard of anybody renting GPUs anywhere near that high. In the clip Huang answers that “the single problem that nothing can solve is demand”, that “we have to make it easier for people to finance these build-outs”, and that “the residual value of Nvidia’s compute is really, really high.” The post does not say when or where the interview was recorded.

What Can Be Checked

DeepSeek’s $38 figure comes from DeepSeek’s own document, read in full on GitHub. It describes one cluster serving DeepSeek V4.1 Flash, and it is stated “at valley pricing,” DeepSeek’s phrase for its off-peak rates. It is not a fleet average or a margin.

The rent side has two published sources. AIMultiple’s index, as reported by Akash Network, gives “$7.98 for B300” as the August 2026 median on-demand rate across posted prices. The chart attached to Das’s post states that its rented GPU prices come from InferenceX: $8.50 for Vera Rubin and $5.00 for GB300. We could not open InferenceX’s own price table, so those two numbers are attributed to Das’s chart.

Sources: as listed in each row. Revenue and rent figures are different measures and are not netted against each other.

Where It Meets Our Coverage

Our compute squeeze piece set two numbers side by side. Renting an Nvidia H100 cost $2.88 per GPU-hour on Silicon Data’s index on 24 October 2026, against $2.00 on 9 December 2025, a rise of 44% by our arithmetic. Over a similar stretch, YipitData wrote that “across our token-based datasets, we’ve observed a ~50% Y/Y decline in token prices.” Das’s post names the same two forces: “this eventually drives up compute prices and drives down token prices,” Das wrote.

Our Rippling piece reported the buyer side of cheaper tokens: Rippling’s chief executive told The Information that Anthropic models went from 57% of its engineers’ coding work in May to 18%, while OpenAI’s share rose from 16% to 52%, as the company sought to cut costs and OpenAI proved cheaper. Our stranded compute piece reported Elon Musk’s view on X that “we’re going to see a lot of stranded compute in partially complete data centers.”

The Risks Das Lists

Das ends with a list of open questions: “do open source models continue to be competitive? Will their creators rent seek more? Will their shape change compute needs? How do different chips change economics? Who gets compute in a supply crunch?” The revenue-share point is the first of these that touches the arithmetic: Das wrote that it “further pushes the threshold to be in the money.” We found no statement from a Chinese model lab confirming revenue-share terms.

Connecting the Coverage

Our H100 piece traced the Silicon Data rent from $2.00 to $2.88. Our token-price piece reported YipitData’s ~50% decline alongside Gemini list prices set to double. Our compute squeeze piece set the two side by side, and our Rippling piece showed a customer moving work on price. Our stranded compute piece set Musk’s post against the rent data. What this piece adds is DeepSeek’s own revenue figure per GPU-hour, and Das’s account of the providers in between.

The Structural Read

As we read it, the revenue figure describes one cluster: DeepSeek’s US$38 is for a cluster serving 155 trillion tokens a day at valley pricing, and Das wrote that margin “relies on your utilization and how good your stack is.”

As we read it, the two sides moved in opposite directions in our earlier reporting: renters paid more, with Silicon Data’s H100 index at $2.88 per GPU-hour on 24 October against $2.00 in December, while sellers charged less, with YipitData measuring token prices about 50% lower year on year.

As we read it, the open variable is who takes a share of the spread: Das wrote that Chinese open-model labs are asking for revenue share at scale, and no lab has confirmed terms.

Three Implications

FOR INFERENCE PROVIDERS Das wrote that gross margin “relies on your utilization and how good your stack is,” and his chart marks “Fireworks reported ~50%” as a reference line.

FOR COMPUTE RENTERS AIMultiple’s index, via Akash Network, put the August median on-demand B300 rent at $7.98 per GPU-hour, and Silicon Data’s H100 index read $2.88 on 24 October.

FOR TOKEN BUYERS YipitData wrote that it has “observed a ~50% Y/Y decline in token prices,” and Rippling’s chief executive told The Information that the company moved coding work to OpenAI as OpenAI proved cheaper.

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework puts it this way: the token “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, Das’s inference providers sit between the second and third of those roles: they rent the compute and sell the tokens, and DeepSeek’s $38 figure and the $7.98 B300 median are published numbers on each side. The framework was published in May 2026 and does not discuss Das’s post or these figures.

What Is Not Established

DeepSeek’s $38 is one cluster at off-peak prices, not an average, and DeepSeek does not give its cost per GPU-hour in that sentence. Das’s $10-20 range, the $1M-a-month claim and the revenue-share claim are Das’s own estimates and accounts; no provider’s revenue is named. The InferenceX prices are taken from Das’s chart. Rents differ by chip, contract and date. We did not contact DeepSeek, Das or the companies named.

The Bottom Line

DeepSeek says one V4.1 Flash cluster earns US$38 per GPU-hour at valley pricing, and the median posted B300 rent in August 2026 was $7.98. Das, a Menlo Ventures partner whose investments include OpenRouter and Modal, wrote that dozens of inference providers are entering the market on this kind of arithmetic, and that their competition “eventually drives up compute prices and drives down token prices.”