DeepSeek's open-source push entered a louder phase this week. A day after the Chinese AI lab published DeepInfer — the inference framework behind its production AI services — on GitHub under the Apache 2.0 license, its researchers escalated the surrounding debate with unusually combative language, declaring the contest between open and closed AI "a decisive battle between two futures, two destinies for humanity."
The framework release itself drew rapid attention, collecting more than 5,500 GitHub stars within days after a Hacker News discussion about it drew hundreds of upvotes on Monday. But it is the rhetoric and the economics claims that accompanied the release — reported by the South China Morning Post and Bloomberg — that signal how DeepSeek now positions itself: not as a low-cost alternative, but as the standard-bearer of an open-source movement fighting for the industry's future. For more on this contest, see our AI models coverage.
'Total Victory': DeepSeek Draws Battle Lines
In a social media post on Tuesday, senior researcher Chen Deli framed open-source and closed-source AI as representing two fundamentally different futures for humanity. "Open source vs closed source … A decisive battle between two futures, two destinies for humanity," Chen wrote. "However long they want to fight, we'll fight. Until total victory."
It is some of the strongest language yet from a Chinese industry figure amid intensifying China-US competition in AI technology. Chen's colleagues struck a more commercial note: Cui Tianyi, who leads DeepSeek's Harness team, said the company hoped more businesses and organizations worldwide would "deploy DeepSeek models and profit from doing so" — a pitch aimed directly at the cloud providers and enterprises deciding where to spend their inference budgets.
The Claim: $38 Per GPU-Hour
Alongside the rhetoric came a concrete number. In a technical article published on the Chinese Q&A platform Zhihu, DeepSeek claimed that a single computing cluster running its V4.1-Flash model on DeepInfer could generate revenue equivalent to $38 per GPU per hour, based on off-peak pricing rates.
The framework's efficiency techniques underpin that figure. According to the company, DeepInfer improves cost-efficiency by retaining previously processed information for more than 72 hours, reducing the repeated computation that normally inflates serving costs. DeepSeek did not disclose its operating costs or profit margins, so the per-GPU figure remains a company claim rather than an audited result — but it is an unusually specific one, and independent cloud operators can now test it directly, since both the model and the serving stack are public.
What DeepInfer Runs On
DeepInfer is tailored to DeepSeek-V4.1-Flash, the open-weight model released in September with a new architecture that reduced the memory required to run it while outperforming the company's earlier V1 Pro model. An update released on Saturday made it one of the world's most powerful and cost-efficient open-weight models, according to third-party benchmarking firm Artificial Analysis.
Notably, the framework targets two hardware ecosystems at once: Nvidia's SM100 Blackwell GPUs and Huawei's Ascend 950 chips. That dual support lets operators serve a frontier-class open model on Chinese silicon as well as on chips subject to US export controls — a flexibility closed providers cannot match, and one that softens the impact of any single hardware restriction. As an open-weight model, V4.1-Flash can be served by third-party cloud providers on their own infrastructure, and DeepSeek said open-sourcing DeepInfer would help the community lower compute costs, reduce energy use, and make AI more affordable.
An Escalation in the Open-Closed Fight
The framing sets up a direct collision with leading US labs. Anthropic CEO Dario Amodei has repeatedly argued that open-source AI poses risks to humanity, and has accused Chinese companies of stealing American AI technology — claims Chinese industry figures openly reject. DeepSeek's executives are now arguing the opposite case in public: that open models are cheaper to run, freer to adopt, and better for the countries and companies that cannot absorb closed frontier pricing.
DeepInfer itself is not a from-scratch product. As the repository documents, it ties together the specialized kernels DeepSeek has released piece by piece over the past two years — including FlashMLA, DeepGEMM, and DeepEP — into a single runtime with model and sampler execution built in. The strategy is consistent: give away the full stack, from kernels to serving framework, and compete on efficiency and openness rather than exclusivity.
What to Watch
Three things will show whether the "total victory" rhetoric has substance. First, adoption: whether Western and third-party cloud providers integrate DeepInfer for V4.1-Flash hosting at scale. Second, verification: whether independent operators reproduce the $38 per-GPU-hour revenue claim under real workloads. Third, the response from closed providers, whose pricing power depends on the efficiency gap that DeepSeek is working aggressively to close. Bloomberg noted the tool is aimed at cutting AI operating costs broadly — if the claims hold, the pressure on closed AI pricing only intensifies.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
