A rented RTX 3090 beats the cheapest no-retention Qwen API above 12% load on OpenRouter's 8 October prices, and two B200s need 76% at $3.70 per GPU-hour. On CTOL's arithmetic, a tenth of a developer's time lifts the 12% hurdle to about 47% at those prices.
OpenRouter's 12 November post says a rented RTX 3090 beats the cheapest Qwen3.8 27B zero-data-retention (ZDR) API above 12% load, on 8 October prices. Two B200s running DeepSeek V4.1 Flash need 76% at $3.70 per GPU-hour. A new 19 October endpoint had already moved a separate 3090 threshold from 8% to about 29%.
OpenRouter sells API routing, and it writes that it did not rent these GPUs itself, so the GPU inputs come from published benchmarks. Its own recommendation is to self-host a base load and use APIs for peaks. CTOL's view: treat the 12% as a floor, not a buying rule, and rent a GPU only for steady base load.
The 12% belongs to an 8 October price list
The break-even load is GPU rent divided by what the same hour of tokens costs through the API. For the chat mix, OpenRouter estimates 565 output tokens a second, or 2.03 million an hour. OpenRouter's table uses RunPod's $0.50 an hour for the card (RunPod's Community Cloud lists $0.22 as of 9 November). The 8 October cheapest ZDR route prices that hour at $4.33 and the median ZDR route at $6.14.
On CTOL's arithmetic, $0.50 divided by $4.33 is 11.5%, and divided by $6.14 is 8.1%. These are OpenRouter's 12% and 8%. The post adds that its break-evens are lower bounds, because they leave out engineering time and spare peak capacity.
The price list then moved. On 8 October, Qwen3.8 27B output cost $1.49 to $5.25 per million tokens across 18 endpoints, 14 of them ZDR. OpenRouter says a new ZDR endpoint charged $0.50 per million output tokens on 19 October, which undercut that range.
The short-prompt 3090 row produces 942 tokens a second, or 3.39 million an hour. At $0.50 per million, those tokens cost about $1.70, so $0.50 divided by $1.70 gives about 29%, on output cost alone. The same $0.50 price applied to the chat mix gives $1.02 an hour and 49%. On CTOL's arithmetic, the 49% is a sensitivity that assumes the endpoint serves equivalent traffic.
A tenth of a developer lifts the hurdle to 47%
The US Bureau of Labor Statistics puts median software developer pay at $135,980 a year, from May 2025 data. Suppose running one GPU takes a tenth of a developer. That costs $1,133 a month, or $1.55 an hour over a 730-hour month, before benefits and overhead.
The effective cost per hour becomes $2.05. Against the 8 October $4.33 hour, the chat-mix break-even rises from 11.5% to 47.4%. Spread the same tenth over ten identical GPUs and it falls to 15.1%. The staffing share is CTOL's assumption, and scale decides how much of it each card carries.
Speed is a second cost. The 942 tokens-a-second figure comes from a heavily patched vLLM build that OpenRouter calls a best case. At 64 concurrent users it gives about 15 tokens a second per user, while Qwen3.8 27B API endpoints had a 45-token median. The conditions differ, so this is a reason to test the application and not a latency finding.
Privacy routing also limits the API side. In OpenRouter's 8 October tests, a DeepSeek-only EU request returned 404, while an EU-domain ZDR route for GLM 5.3 returned 200. The permitted route sets the comparator.
B200s need three-quarters of capacity at the lowest rent
| Setup, 8 October prices | Cheapest ZDR | Median ZDR |
|---|---|---|
| Qwen3.8 27B, RTX 3090 chat mix, RunPod $0.50 | 12% | 8% |
| DeepSeek V4.1 Flash, 2x B200, InferenceX retail $3.70 | 76% | 37% |
| DeepSeek V4.1 Flash, 4x H200, InferenceX retail $2.90 | Never | 82% |
"Never" means rent exceeds the API cost of the same hour even at full load. The 76% is $3.70 divided by $4.87, the cheapest ZDR hour for the B200 mix. A two-B200 node costs $5,402 a month at retail rent and needs 2.7 billion output tokens to beat that route. The rented 3090 costs $365 a month and pays back after about 170 million chat-mix output tokens.
RunPod's October B200 rate of $6.79 never beats the cheapest route. RunPod's pricing page, as of 9 November, lists B200 at $5.98 on Community Cloud and $7.99 on Secure Cloud. Both exceed $4.87, so no load clears it on those rents.
Demand is the other limit. Office-hours traffic, 40 of 168 hours, keeps a GPU busy at most 24%. Round-the-clock batch can reach 70%, which clears the median B200 hurdle at $3.70 and misses the cheapest. In a September VentureBeat Pulse survey of 71 enterprises, 81% put average GPU utilization at 50% or less. About three in ten said a quarter of the time or less. The answers are self-reported and mix owned and rented GPUs.
Blackwell and hybrid cases cut the other way
Yingjun Wu of Recurrence, which sells inference, says "an 8x B300 node breaks even at 35% utilization". He writes that small teams "can make money on 8xB300 at 50%+ utilization with a couple of direct customers". He also writes: "A new provider relying on OpenRouter routing alone typically reaches 10-30%." On H200, he says full load earns $1.6 an hour against $3.99 rent, consistent with OpenRouter's "never" for that hardware.
Jonathan Knoop and Hendrik Holtmann argue in a January 2026 arXiv paper that consumer GPUs can reliably replace cloud inference for most small-business workloads, with an exception for time-sensitive retrieval over long contexts. They report payback under four months at 30 million tokens a day. Their costs are electricity only, on owned cards and models older than Qwen3.8 and DeepSeek V4.1.
Dor Amir of Nadir, which sells routing software, models a synthetic week whose peak-sized fleet ran at 32% utilization. His hybrid of five GPUs plus API overflow cost $14,054 a month, against $20,075 for eleven peak-sized GPUs and $21,875 for API only. He says "you shouldn't self-host for cost" if the cheapest API clears the quality bar. The figures are modeled.
Lambda's 16 November B200 prices set the next test
Lambda's pricing page says B200 on-demand rates of $7.69 to $7.99 per GPU-hour take effect at 00:00 UTC on 16 November. At 70% load, break-even needs an API-equivalent hour of $10.99 to $11.41, against $4.87 in OpenRouter's 8 October cheapest-ZDR comparison. The next check is whether the cheapest eligible DeepSeek route that day prices the same hour of tokens below that line.
The evidence favours renting only steady base load that clears a staffed break-even. If the cheapest route prices the hour below Lambda's line on 16 November, a rented B200 does not clear it.