A Russian-language news item said DeepSeek had released DeepInfer, its own inference code for DeepSeek V4.1 Flash, and bundled it with a broader argument about GPU inference economics. The only source attached to the item, however, was an October 22, 2026 opinion post by Evan Conrad, CEO of SF Compute, titled 'The end of startups', which argues that compute is sold out, that non-investment-grade buyers are being asked for ever larger upfront payments, and that SF Compute's forwards market for GPU hours is the fix. That post never mentions DeepSeek or DeepInfer.
The two threads matter to builders for different reasons. An official, production-grade inference engine from a frontier open-weight lab changes the cost of self-hosting that lab's models. Claims about rental prices, inference margins and credit terms shape whether a team rents, reserves or buys API tokens. This analysis checks both, separately.
What we checked
- Can the cited GPU rental rates, inference revenues, and reported upfront-payment requirements be verified through primary contracts or market data? partly confirmed. The post's figure of about $6 per hour for a B300 sits inside the public on-demand range on October 25, 2026 (cheapest on-demand $4.99 to $6.17 depending on tracker, medians $8.02 to $9.25). The roughly $30 per hour of inference revenue per B300 and the 'as high as 100% upfront' contract terms are the author's own assertions; no contract, filing or market dataset I opened confirms them. [10][6][7][8]
- Did DeepSeek actually publish DeepInfer, its inference code for DeepSeek V4.1 Flash? confirmed. The deepseek-ai/DeepInfer repository exists, was created on October 16, 2026, is Apache-2.0 licensed, and describes itself as the framework that powers DeepSeek's production services, with DeepSeek-V4.1-Flash as the only supported model. The linked SF Compute post does not mention DeepInfer at all. [1][2][10]
What we found
- DeepSeek published DeepInfer on GitHub. The repository was created on October 16, 2026, carries an Apache-2.0 license, and had commits as recent as October 25, 2026. Its README calls it 'the high-performance inference framework that powers DeepSeek's large-scale production services' and lists DeepSeek-V4.1-Flash as the only supported model. [1][2]
- The open-source release is offline inference only. The README says a complete service still needs a KV cache system, prefill/decode disaggregation and load balancing across data-parallel replicas, none of which ship in this release. Supported hardware is NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, built against CUDA 13 or CANN 9.2. [2]
- Features listed include attention data parallelism, MoE expert parallelism with load balancing, DSpark draft-token decoding with confidence-scheduled verification, batch-invariant inference, deterministic sampling and structured outputs. The examples default to the deepseek-ai/DeepSeek-V4.1-Flash-1017 checkpoint and show expert, data and context parallel flags. [2][3]
- Third parties are already wrapping it. GPUStack's DeepInfer-server repository adds a Chat Completions API with streaming and tool-call parsing on top of a pinned DeepInfer commit, for a single multi-GPU SM100 node. Separately, the DeepSeek-V4.1-Flash model card on Hugging Face carries a minimal reference implementation that is explicitly not a production serving engine. [4][5]
- Public trackers put B300 rental prices close to the post's $6 per hour. On October 25, 2026, gpurentalprices.com listed B300 on-demand from $4.99 with a provider median of $9.25; priceofcompute.com showed a median of $8.02 across four providers; gpueconomy.com listed cheapest on-demand at $5.89 and a median rate card of $8.99. So $6 is a low-end but real on-demand price, not a typical one. [6][7][8]
- The $30 per hour of inference revenue per B300 cannot be checked from public data. DeepSeek's own API lists deepseek-flash, served by DeepSeek-V4.1-Flash-1017, at $0.30 per million cache-miss input tokens and $1.20 per million output tokens at peak, with off-peak at half price. Turning those prices into revenue per GPU-hour would need throughput and utilisation numbers that neither DeepSeek nor the post provides. [9][10]
- The upfront-payment and credit-crunch claims rest on the author alone. The post says SF Compute has seen contracts with up to 100% paid upfront and forecasts that most AI startups will be unable to buy compute within six months, then presents SF Compute's forwards market as the solution and links to hiring. No contract, lender statement or filing backs those figures in the material I opened. [10]
What it means
If you serve DeepSeek-V4.1-Flash yourself on Blackwell-class NVIDIA hardware or Ascend 950, DeepInfer is now the reference engine from the people who run the model in production, under a permissive license. Treat it as an engine, not a service: you still need your own KV cache layer, prefill/decode split and replica load balancing, or a wrapper like GPUStack's. Compare its throughput against your current stack before switching, because DeepSeek's own API pricing for the same model is low enough that self-hosting only pays at high, steady utilisation.
On the economics, use the public price trackers rather than the opinion post. B300 on-demand prices range widely, from about $5 to more than $11 an hour, so shopping around and committing for a term matter more than any headline figure. The claims about 100% upfront contracts and a coming startup credit crunch may be true for large reserved clusters, but they come from a company selling a fix, so treat them as a forecast to watch rather than a fact to plan on.
Still open
- Will DeepSeek release the missing service pieces (KV cache system, disaggregation, load balancing) or support for other DeepSeek models?
- What throughput per GPU does DeepInfer reach on B300 versus vLLM or SGLang for V4.1-Flash, measured independently?
- Is there any public contract, lender or filing evidence of non-investment-grade buyers paying 100% upfront for GPU clusters?
- What is a realistic inference revenue per B300-hour at current API prices and utilisation?