Nvidia B300 Break-Even: 35% at $7.40 vs 12% on Patel’s Data

An Nvidia B300 serving DeepSeek V4.1 Flash breaks even at about 35% utilisation, and an Nvidia H200 cannot cover its rent even when fully busy, Yingjun Wu, founder of RisingWave Labs, wrote in an X article on 2 November. On the figures SemiAnalysis chief executive Dylan Patel gave in a talk posted the same day, a rented B300 megawatt breaks even at about 12%, by our arithmetic.

Both men use the same model, SemiAnalysis’s InferenceX benchmarks and the same definition of utilisation: the share of the tokens the hardware could produce that are actually sold.

The gap is in the other inputs. At full load, Wu’s B300 earns $21.4 an hour against $7.40 of rent, about 2.9 times, by our arithmetic. Patel’s $70 million a year at 60% utilisation implies full-load revenue about 8.3 times his $14 million a year rental figure, if revenue scales with utilisation, by our arithmetic. In the part of the talk that sets out the per-megawatt economics, Patel does not give a token price, a throughput or a number of GPUs per megawatt.

Business Pill · Gross Margin

An evergreen Business Engineer explainer on gross margin: what an AI company keeps from each dollar. It teaches the general idea only and does not discuss Wu, Patel or GPU rents.

What Wu Calculated

Wu posted the article on X at 22:46 UTC on 2 November with a summary: “H200 can’t cover its rent even at 100% utilization, B300 breaks even at ~35%.” He added: “Profitable, yes. Simple, no.” The article says it was originally published on Recurrence Research.

He holds one setup fixed. The model is DeepSeek V4.1 Flash. The traffic is SemiAnalysis’s AgentX workload, 393 real Claude Code sessions replayed, with a median input of 142k tokens, a median output of 444 tokens and about 95% of input served from cache. Each user gets 100 output tokens a second. Throughput is the best public open-source engine result on InferenceX for each card, with FP4 weights.

Price is what OpenRouter providers actually collect on this model, he writes: $0.013 per million input tokens, blended across cached and uncached, and $0.57 per million output. With a 99/1 input/output mix, that is $0.019 per million total tokens. “Not list price. Not DeepSeek’s off-peak price. What the market pays.”

Rent is the lowest publicly visible on-demand price: $3.99 an hour for an H200 and $7.40 for a B300, both from Hyperstack. His formula is tokens per second times 3,600 times price times utilisation, minus rent, with no bandwidth, people or spare capacity.

On those inputs, Wu’s table gives the B300 a throughput of 315,201 tokens a second per GPU and full-load revenue of $21.4 an hour, so it breaks even at 35% and runs a 31% gross margin at 50% busy and 51% at 70%. The H200 manages 24,124 tokens a second and $1.6 an hour, a break-even of 244%. A GB300 NVL72 rack needs from 24% to 84%, depending on how it is configured.

What Was Posted on X

Yingjun Wu, founder of RisingWave Labs, posted this on 2 November with a link to his X article on whether hosting open-source models is profitable. It gives his two headline results for DeepSeek V4.1 Flash: an H200 cannot cover its rent even at 100% utilisation, and a B300 breaks even at about 35%. This piece sets his assumptions beside the figures Dylan Patel gave in his Runtime 2026 talk.

What Patel Said a Megawatt Earns

Patel, speaking at Modal’s Runtime 2026 conference in a video posted on 2 November, put the cost of building a B300 cluster at $40 million, plus about $9 million a year to operate it over five years. “If you’re renting it from the clouds, you know, you’re lucky if you’re paying $14 million a year for a megawatt,” he said.

On revenue, he used DeepSeek V4.1 Flash at 60% utilisation, which he defined as “only selling 60% of the tokens that you could generate”, and said “you’re making $70 million a year.” He said the numbers come from SemiAnalysis’s open-source InferenceX agent benchmarks. He added: “even if you say, oh, utilization is going to be 30%, you’re still going to make a ridiculous sum of money.”

If revenue scales with utilisation, his figures put break-even at about 12% for a rented megawatt and about 15% for an owned one, by our arithmetic, as we reported this morning.

Where the Two Calculations Part

Rent is the first input that differs. Patel said that “if you can find B300s for $6 an hour, you should sign it and you should sell tokens”. Applied to Wu’s full-load revenue of $21.4, a $6 rent moves the B300 break-even from 35% to about 28%, by our arithmetic. Wu’s own sensitivity table gives 26% at a reserved rent of about $5.50, which he labels his estimate, and 51% at the SAVRN median of $10.80.

Price is the second. Wu’s table moves the B300 break-even to 52% at DeepSeek’s off-peak price of $0.0126 per million total tokens and to 26% at list price, $0.0252. SemiAnalysis’s 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”.

Speed per user is the third. Wu’s break-even is 35% at 100 tokens a second per user, about 20% at 50 and 93% at 200. He writes that with all three inputs at his “good” setting a B300 breaks even at 10%, and with all three “bad” it loses money at 100%. The 12% from Patel’s figures sits closer to that 10% than to the 35% base case. Wu’s own reading of the split: “Each is correct for its own assumptions.”

The two read 30% differently. On Wu’s base case, a B300 at 30% earns about $6.42 an hour, about 87% of its $7.40 rent, by our arithmetic; at a $6 rent, it keeps about 7% of revenue. On Patel’s figures, scaled in proportion, 30% brings in about $35 million a year against the $14 million rental figure, about 2.5 times, by our arithmetic.

What Wu Says the Listings Leave Out

Wu writes that a new provider relying on OpenRouter routing alone lands at 10% to 30% utilisation, below the B300’s 35%, so the setup works only if the provider brings one or two direct customers. OpenRouter’s default routing weights providers by the inverse square of price, he writes, so a provider at $1 gets roughly 9x the traffic of one at $3.

He also writes that AgentX throughput counts cache hits, and that about 95% of the tokens flowing through a box are cache hits billed at $0.003 to $0.006 per million. On rent, he cites an a16z chart of Ornn’s index: H200 on-demand rent is up 163% since January. At Ornn’s executed-transaction rents, he writes, the B300 break-even moves from 35% toward 50%.

Connecting the Coverage

Our related pieces on GPU rents and token prices:

• Nvidia B300 Rent Up 52% Since May as AI Token Prices Fall: our piece on the squeeze between rent and token prices.
• AI Token Prices 2026: What Every Major Model Charges per Million Tokens: our price tracker.
• Traders Can Now Bet 10x on Nvidia GPU Rents, H100 at 40%: our piece on GPU rent futures.

The Structural Read

As we read the two, the hardware is not what separates them: both run DeepSeek V4.1 Flash on Nvidia B300s and measure throughput with SemiAnalysis’s InferenceX. What differs is what a token sells for, how fast each user is served and what an hour of the chip costs.

As we read Wu’s sensitivity table, each of his three inputs can move the B300 break-even by more than the full gap between his base case and the 12% we derived from Patel’s figures; the speed setting alone runs from about 20% to 93%.

As we read the 30% case, the two sets of inputs point in opposite directions on the same utilisation: about 2.5 times the rental figure on Patel’s numbers, and about 87% of the rent on Wu’s base case, by our arithmetic.

Three Implications

FOR TOKEN SELLERS Wu writes that a new provider relying on OpenRouter routing alone lands at 10% to 30% utilisation, below his 35% B300 break-even, and that the setup works only with one or two direct customers.

FOR GPU RENTERS Patel said a B300 at $6 an hour is worth signing; on Wu’s throughput and price that rent puts break-even at about 28%, by our arithmetic.

FOR H200 OWNERS Wu’s table gives the H200 $1.6 an hour of full-load revenue against $3.99 of rent, a break-even of 244%.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, Wu and Patel are arguing over the size of that marginal cost against the price a token fetches. The framework does not discuss Wu, Patel, DeepSeek or the B300.

What Is Not Established

Neither calculation is a reported profit. Wu’s figures come from public benchmarks, listed rents and OpenRouter prices, and he asks what others see in production. Patel’s $70 million is his estimate on InferenceX benchmarks. The 12%, 15%, 28% and 2.9-times and 8.3-times ratios are our arithmetic, and the Patel-side figures assume revenue scales in proportion to utilisation, which he did not state.

Patel’s figures are per megawatt and Wu’s are per GPU-hour; without a GPU count per megawatt, the dollar figures cannot be converted, so we compare only ratios and break-even rates. We did not re-run the InferenceX benchmarks or contact Wu, Patel or SemiAnalysis.

The Bottom Line

Yingjun Wu puts the Nvidia B300 break-even for DeepSeek V4.1 Flash at 35%, at $7.40 an hour of rent and $0.0188 per million tokens, and says an H200 cannot cover its rent. Dylan Patel’s per-megawatt figures imply about 12%, by our arithmetic. The model, the benchmark family and the definition of utilisation are the same; rent, token price and speed per user are where the two sets of assumptions differ.