Anthropic's April preview of Claude Mythos, a next generation frontier model, made the next capability plateau feel unreachable for one of China's largest open source AI labs.
When Anthropic released the first Claude Mythos preview in April 2026, the reaction inside one of China's largest open-source AI labs was immediate: DeepSeek founder Liang Wenfeng saw a model trained at a scale of compute and data that no algorithmic-efficiency shop could match by cleverness alone. Within days, the first financing rumors leaked. Three years after telling venture capitalists to walk away because DeepSeek would never build a product roadmap, Liang had decided the capability gap was real, and the only way to close it was to stop running his lab like a university lab and start running it like a serious infrastructure company.
That framing of Mythos as the trigger for the raise comes via QbitAI's summary of a recent report from The Information. The numbers behind it tell the scale of the pivot. DeepSeek is raising roughly $7.4 billion, and Liang wrote the largest single check himself, about 20 billion yuan (roughly $3 billion) of personal cash, or close to two-fifths of the round, according to QbitAI. Post-financing, the company said publicly this week that all of its departments, including AI systems, infrastructure, product, and deep-learning research, will at least double headcount, from a current base of around 300. The core Harness team, which wraps DeepSeek models into autonomous agents, is actively recruiting for multiple roles in Beijing, and is run by Cui Tianyi, who left Jane Street in March and has been posting multiple job listings on X.
That hiring push is the obvious story. The harder one is what the money is buying, which is the same thing that already cost DeepSeek 15 months.
In 2025, the lab quietly rebuilt its training and inference stack off Nvidia's CUDA software platform to run natively on Huawei Ascend chips. The migration moved DeepSeek off the standard Nvidia software path just as the rest of the industry was racing to ship a new flagship model every two or three months. Peer labs went through a full generation while DeepSeek sat out roughly fifteen months without a release. The gap is now visible on the calendar: DeepSeek's V4 flagship shipped in April 2026, and Reuters reported the same week that V4 had been adapted to Huawei's hardware, with Fortune framing the release as a price-performance play aimed at developers looking for cheaper inference. The same gap also cost DeepSeek the H2 2025 Anthropic Claude Code wave, the developer-facing coding agent that has become one of the most lucrative monetization paths the industry has found so far.
Liang has not pretended that gap is closed. He told investors during the roadshow that coding tools and AI chatbots are transitional artifacts on the road to AGI, per QbitAI, defining AGI in his telling as machines capable of human-level understanding, reasoning, learning, planning, and adaptation across broad tasks. It is a defensible research position, but it also helps explain why DeepSeek has been less embedded in the developer toolchain than the openness of its model releases would have suggested, even as the rest of the frontier shipped developer tools.
The strategic bet underneath the raise is the chip one. Liang told the same investor audience that he expects Huawei chips to catch up to Nvidia within a few years, and that he wants DeepSeek to be the first major lab to finish the native adaptation, QbitAI reported. Huawei only began working with the lab directly after learning, in 2025, that DeepSeek had been quietly testing Ascend hardware on its own. Additional context from Chinese outlet Guancha and Huxiu frames the round against a tighter domestic funding environment for Chinese AI labs and emphasizes the unusual structure of Liang's personal contribution.
Liang, for his part, has told people around him that his strategy does not change: keep open-sourcing, keep prices low, stay focused on AGI. He has also said that AI should not be controlled by a small group of people, according to QbitAI's summary of his recent public and roadshow comments. On open-sourcing, DeepSeek continues to stand out as one of the only major AI labs that publishes the underlying code for all of its model releases, rather than just the model weights, though that positioning should be read as competitive self-description rather than as an independently audited claim against peers like Meta's Llama team and Mistral.
What the round actually closes on is the question the next twelve months will answer: whether doubling headcount to roughly 600 and putting the largest single check in the hands of an open-source, no-roadmap, Huawei-first founder is enough to close the Mythos-scale compute gap that triggered the raise in the first place. DeepSeek's claim to being a different kind of AI lab depends on it.