As DeepSeek prepares to roll out the official version of its V4 model — expected around mid-July/early August 2026, following an April preview, with legacy models retiring July 24 — the Hangzhou-based lab is doing so from a position of new financial strength and unusual philosophical clarity. In June, DeepSeek closed its first external funding round, raising roughly $7.4 billion at a post-money valuation of $50–59 billion, led by a personal $3 billion commitment from founder Liang Wenfeng alongside Tencent, CATL, and China's state-backed AI investment fund. A second round, reportedly targeting a $74 billion valuation, is already said to be in motion.
For an industry accustomed to founders chasing valuation and user growth, Liang's public reasoning offers a striking counter-narrative — one investors evaluating the Chinese AI sector would do well to understand.
The Roadmap, Not the Product
Liang has been explicit that DeepSeek's mission is not iterative product-building but a staged march toward artificial general intelligence: Coding Agents, then continuous learning, then AI self-iteration, then embodied intelligence. Coding agents are the current priority; multimodal capability, 3D, video generation, and world models are deliberately deprioritized as tangential to raw intelligence. Even hallucination reduction, while acknowledged as necessary, is treated as a "product issue" rather than a research priority.
The logic: AI's central bottleneck isn't taste or intuition but the inability to learn continuously — a capability no lab, DeepSeek included, has cracked. Solving it, Liang argues, is the precondition for models that can meaningfully accelerate AI research itself, and only then for embodied intelligence.
A Crowded Field, A Deliberate Pass
DeepSeek operates in China's densely populated frontier-model field, competing against Zhipu AI's GLM-5.2 — which triggered its own "mini DeepSeek moment" in June by matching global coding benchmarks at a fraction of the cost — and Moonshot AI's Kimi K3, a roughly 2.8-trillion-parameter open-weight model closing gaps with Western systems on long-horizon tasks. Globally, OpenAI and Anthropic remain the reasoning and safety benchmark, though Chinese labs increasingly rival them on cost-adjusted performance.
Liang's response to this fragmentation is notably passive: he expects Chinese labs to eventually consolidate around a handful of players, and explicitly declines to compete for the "super app" prize that rivals are chasing through aggressive commercialization. DeepSeek, he says, has no interest in becoming "the next ByteDance" or "the next Tencent."
Restraint as the Real Strategy
The most consequential claim in Liang's worldview — and the one executives should sit with longest — inverts standard competitive doctrine. Liang argues that pursuing maximum market share or profit is not merely unnecessary but self-defeating: "If your vision is to gain more, you've already lost." His reasoning is structural, not sentimental. AI, he believes, will eventually represent roughly 10% of global GDP; any company that tries to monopolize a prize that large will be "abandoned by history." Restraint — open-sourcing frontier models rather than gating them, cutting prices to a quarter of their original level despite internal objections, and capping ambition at "reasonable profit" rather than maximization — is therefore not philanthropy. It is, in Liang's framing, the strategy most likely to raise DeepSeek's probability of winning the only race that matters.
This has concrete governance consequences: even after taking $7.4 billion in outside capital, Liang structured the round so most investors hold no voting rights and face a five-year lock-up, preserving his control precisely so short-term investor pressure cannot erode the long game.
For executives and investors watching from outside, the real question is not whether Liang's bet is altruistic — it is whether it is rational. In a market where Liang himself identifies cost, not brand or feature velocity, as the primary competitive axis, a founder willing to sacrifice margin and market share for AGI probability may be making the more calculated wager, not the more idealistic one.