“DeepSeek makes GPUs 3x more valuable.” Sounds huge. Sounds like free money.
Before you rush to buy or sell anything, let’s look at what DeepSeek actually released. Spoiler: it’s real, it’s interesting, and it’s not a magic drop-in upgrade.
Quick answer
DeepSeek open-sourced DeepInfer on October 20, 2026. It’s the in-house inference engine DeepSeek uses to serve its models. It’s Apache 2.0 licensed. It runs on NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 chips. The “3x” figure is DeepSeek’s own revenue claim: about $38 per GPU-hour versus roughly $12.91 from its early-2025 setup. It isn’t a verified 3x speedup on any GPU, and it’s not a drop-in tool for NVIDIA and AMD.
What is DeepInfer?
It’s DeepSeek’s inference framework. Inference means the part where a trained model answers your questions.
DeepSeek built it mainly on vLLM, with ideas from SGLang and TensorRT-LLM. Then it put the code on GitHub. By the first afternoon, the repo had over 8,000 stars.
The code includes a scheduler, speculative decoding, expert load balancing for mixture-of-experts models, and structured outputs. It also has deterministic sampling, so you get the same output on repeated runs.
Where does the “3x” come from?
Not from a benchmark in the repo. The repo has no speedup multiplier at all.
The number traces back to DeepSeek’s own figures. Reportedly it earns about $38 per GPU-hour with this engine. Back in early 2025, it calculated about $12.91 per GPU-hour for its V3/R1 setup. Divide one by the other and you get roughly 3x.
Now the fine print:
- It’s DeepSeek’s own number. No outside team has verified it.
- It reflects revenue per GPU-hour, not profit.
- It’s a comparison across two different years and setups.
- It says nothing about what a GPU is worth on the market.
So “3x more valuable GPUs” is a headline, not a finding.
Does it work on NVIDIA and AMD?
Not as claimed. The confirmed hardware is NVIDIA SM100 (Blackwell) data-center GPUs and Huawei Ascend 950. One codebase, switched by a setting.
AMD? I couldn’t find any confirmation. H100 and A100 fleets aren’t supported yet either, according to one hands-on review.
It’s also not a “drop-in replacement.” Right now it supports one model: DeepSeek-V4.1-Flash. That checkpoint is around 510 GB. And some benchmark runs used 64 Blackwell GPUs for decoding. This is not a weekend project on your gaming card.
What’s missing from the release?
A lot, if you want to run a real service.
- No KV cache system.
- No prefill/decode disaggregation setup.
- No load balancing across replicas.
- Offline inference only.
- Several kernel folders are empty in a default clone.
DeepSeek’s own README says those production pieces aren’t included. So the best numbers come from a full internal stack that you don’t get.
Is there a catch for US companies?
Possibly. Reports say the release landed one day after the US ordered civilian federal agencies to remove DeepSeek products. One analysis flags that the order may reach third-party tools built on DeepSeek code. If you work with US government clients, check with a lawyer first.
Why should you care?
Because inference cost is where AI money gets made or lost. Better scheduling and speculative decoding mean fewer idle GPU cycles. That’s a real trend. Open code lets other engine teams learn from it.
But for most small teams in Pakistan, India or the UAE, the practical move is simple. Don’t buy hardware on this headline. Watch for independent benchmarks. Test on your own workload.
What we still don’t know
- Independent speed tests.
- Support for AMD, H100 or older GPUs.
- Whether DeepSeek will release the missing production parts.
- How the 3x holds up outside DeepSeek’s own setup.
FAQ
What is DeepSeek DeepInfer?
It’s DeepSeek’s open-source inference engine, released October 20, 2026 under Apache 2.0.
Does DeepInfer make GPUs 3x faster?
Not proven. The 3x comes from DeepSeek’s revenue per GPU-hour ($38 vs about $12.91). It’s self-reported and not an independent speed test.
Which GPUs does DeepInfer support?
NVIDIA SM100 (Blackwell) and Huawei Ascend 950. AMD support isn’t confirmed.
Is DeepInfer free?
Yes. The code is Apache 2.0. But you need big hardware and the model weights, and a full production setup needs extra parts you’ll have to build.
Can I use it with other models?
Not yet. Right now it supports DeepSeek-V4.1-Flash only.
Bottom line
DeepInfer is a genuinely useful open-source release. It’s not a 3x miracle for every GPU.
The smart take: learn from the code, wait for independent tests, and ignore the “3x more valuable” hype.