Yingjun Wu's H200 math: hosting DeepSeek V4.1 Flash can't cover rent

Builders on X are pushing back on the idea that hosting open-weight models is easy money. Yingjun Wu, founder of inference startup Recurrence Research, posted an analysis on November 2 claiming that hosting DeepSeek V4.1 Flash on H200 GPUs cannot cover its rent even at 100% utilization, while an 8×B300 node breaks even around 35%. The post drew 177 likes, 46 replies and a linked long-form article within roughly seven hours.

Key facts

• The analysis claims H200 hosting of DeepSeek V4.1 Flash earns about $1.60 per GPU-hour at full load against $3.99/hr rent (Hyperstack on-demand), so no utilization level breaks even
• It puts an 8×B300 node at ~35% break-even utilization ($7.40/hr rent), with 31–51% gross margin at 50–70% utilization
• The market price used is OpenRouter's realized blended price for DeepSeek V4.1 Flash (1017 update): about $0.013 per million input tokens blended across cache, $0.57 per million output, roughly $0.019 per million total at a 99/1 mix
• Engagement: 177 likes, 14 reposts, 192 bookmarks and 32,000+ views in the first ~7 hours, per the post's public counters and the morning Grok sweep
• Our recompute: 3.99/1.60 gives an H200 break-even of 249% utilization (impossible), and B300 margins of 30% at 50% and 50% at 70% utilization — matching the article's 31–51% range within a point

What's driving the conversation

Wu's post, and the long-form X article it links, argues against what he calls "a common claim on X and LinkedIn ... that anyone who can rent GPUs can sell tokens at a profit." The method: hold everything fixed except the GPU card. One model (DeepSeek V4.1 Flash, which he cites as roughly 30% of all OpenRouter tokens for the week of October 19–25), one traffic pattern (SemiAnalysis's AgentX agentic workload, ~95% cache hits), one price (OpenRouter realized, not list), one rent (lowest public on-demand rates). Then compute gross margin as revenue minus rent per GPU-hour.

His three headline results, all from public inputs:

• H200 loses at any utilization. The card lacks FP4 compute units, so it runs the FP4-weights model, by his figures, at roughly one-thirteenth of B300 throughput while costing a bit more than half the rent. Full-load revenue of $1.60/hour against $3.99 rent means break-even would require 249% utilization.
• B300 is "the practical starting point": break-even near 35%, clearly profitable above 50%. But 35% is above where new providers typically land on OpenRouter routing alone, which he puts at 10–30%, citing DeepInfra's estimate of 10–20% for typical self-hosted inference.
• GB300 NVL72 is "the best and the worst option, depending on who tuned it": an 84% break-even at standard configuration versus 24% with the vLLM team's disaggregated rack setup, a 3.5× spread from software.

The replies have largely been infra practitioners arguing about the sensitivity table rather than the conclusion — Wu himself shows the B300 break-even swinging from 10% to 93% depending on price, rent and throughput assumptions. Darkbloom's warning to Mac Mini providers, reported by The Register via XenoSpectrum on October 30, that sagging payouts will not recover quickly, is the same margin-pressure beat arriving from a different network.

The substance

What is verified: the method reconciles. We recomputed the break-even arithmetic from the thread's own stated inputs — $3.99/hr H200 and $7.40/hr B300 rents, the ~$0.019 per million token blended realized price, and the article's full-load revenue figures. H200 break-even works out to 249% utilization, which is to say never. B300 margins compute to 30% at 50% utilization and 50% at 70%, inside the article's stated 31–51% band. Nothing in the math is internally inconsistent.

What is not verified: the inputs. The throughput numbers come from SemiAnalysis's InferenceX benchmark runs and a vLLM blog post; the realized price is a single-day (October 26) OpenRouter reading; the rents are the cheapest publicly visible on-demand rates, and the article itself notes Ornn's executed-transaction index ran H200 at $6.30 and B200 at $9.70 that day, which would push the B300 break-even from 35% toward 50%. This is an analysis claim built from public estimates, not a market measurement — our verification is a recompute, not an audit.

The frame matters as much as the numbers. Wu is not saying open-weights hosting is doomed; he is saying it is a hardware-selection and sales-channel business in which the model and the engine are free and therefore advantage nobody. His closer: "Hosting an open-source inference engine produces tokens, not profit. Getting paid for those tokens is a separate business, and most providers on OpenRouter are not really in it." That is analysis and commentary, and we label it as such — but it is analysis with its work shown.

Why builders are watching

Anyone deciding whether to supply or to buy inference should see break-even math rather than marketing. For suppliers, the post argues the deciding variables are the card (Blackwell or nothing, in his telling), the configuration, and whether you bring direct customers instead of relying on OpenRouter's price-ranked routing. For buyers, it explains why endpoint quality varies so much at the bottom of the price table: when the market pays ~$0.019 per million blended tokens, the margin lives in output caps, quantization and cache behavior, which is exactly what our earlier margin-claims and OpenRouter token-share coverage tracked. Whether a marketplace can hold quality while competing on realized price is the open question underneath all of it.