At 09:30 UTC on Tuesday, DeepSeek pushed DeepInfer to GitHub under the Apache 2.0 license: the inference engine that runs DeepSeek V4.1 Flash in production, shipped alongside a component called deepseek-recipe that converts API requests into tokens and back. By noon UTC a release tracker had the repository past 6,400 stars and 500 forks. Hacker News did what Hacker News does.

The number that matters wasn’t the throughput claim. It was the price. DeepSeek says one cluster serves 155 trillion tokens a day, and it volunteers a figure of roughly $38 per GPU-hour at its off-peak rate.

Companies do not publish their unit economics out of humility. They publish them when they want to be the denominator.

The Note From April 2025 Said No

Anyone who wants to read this as a come-to-Jesus moment about openness should go back eighteen months. In April 2025 the same team posted a short GitHub note explaining why it would not release the engine. Three reasons: the code was an old fork, it was bolted to internal infrastructure, and a small research team had no capacity to maintain a large open-source project. All three were true, and all three were also the standard reasons every lab gives before it decides it has something better to do with the code than keep it.

What changed since then wasn’t a change of heart about community. It was the hardware.

DeepSeek V4 was trained on Huawei silicon, and the surrounding ecosystem has been rebuilt around what Chinese labs can actually buy rather than what they’d prefer. Reporting cited by policy trackers has Huawei producing on the order of 750,000 Ascend 950PR units in 2026. Nvidia, for its part, has told investors it is effectively out of the China data center market, now served by Huawei and its competitors.

When you are serving models on a stack the rest of the world has never optimized for, the most valuable thing you own is the code that makes it fast. And the most valuable thing you can do with that code is give it away, so that every engineer who touches it bends your stack into the shape of the default in places you will never sell a chip.

Why You Publish Your Own Price

Consider the strategic content of the $38 figure. It is a claim, unverifiable from the outside, and it is being planted in every procurement meeting and board deck on earth. A CFO with a model to serve now has a reference number to wave at American vendors. Those vendors can dispute it, but disputing a benchmark is a far worse position than setting one.

That is the part the open-source cheerleading misses. Free weights are a gift. A free serving stack is a standard. Weights get commoditized by everyone eventually; the layer that turns weights into tokens at a defensible price is where the margin actually lives. DeepSeek handed out the second thing and called it the first.

The Moat Washington Is Still Measuring

Here is where the export-control debate has gotten lazy — and where the people who have been right about the chip war need to update their priors.

The argument in Washington remains a fight about silicon: how many H200s, whether the Chip Security Act closes the smuggling routes, whether the December 2025 loosening was a mistake. Fine. But the CUDA moat always had a software twin, and it is the serving layer — vLLM, TensorRT-LLM, SGLang — the unglamorous middleware that American engineers treat as plumbing and American labs treat as a trade secret.

Nobody in the policy conversation is counting that. There is no Bureau of Inference Stacks. And the American labs, busy guarding weights that will be replicated in eighteen months anyway, keep their serving code closed — which means the reference implementation for the post-CUDA era is being written in Hangzhou and handed out with a permissive license.

I asked a staff engineer at a mid-sized inference startup what his team made of it, on the smoking patio outside a San Jose hotel bar during a serving-stack meetup. His answer was blunter than the discourse: “We stopped benchmarking against each other. Now we benchmark against their number.”

That is what standards capture looks like from the inside. Not a takeover. Just a quiet, near-unanimous agreement about which ruler to use.

The Gift Is Free. The Ruler Isn’t.

None of this requires DeepSeek to be sinister, and the reflex to call it theft is mostly a way of avoiding the harder thought. DeepSeek released real code that will genuinely lower costs for developers from Lagos to Louisville. It is also running the oldest play in the standards business: commoditize the layer beneath you, own the layer everyone measures themselves against. That $38 number will appear in more procurement documents this quarter than any benchmark MLPerf publishes.

The people who should feel least comfortable this week are not the export-control hawks; they have a coherent, if incomplete, theory of the problem. It is the open-source triumphalists who will spend the week calling this a victory for humanity, and the American labs that will keep treating their serving stack like a state secret while congratulating themselves on a moat that has already been forded.

DeepSeek gave the engine away. The benchmark was the point.

Sources