July 24, 2026, (Inside AI) — DeepSeek will likely keep its most advanced AI models open-source, founder Liang Wenfeng told investors, doubling down on a strategy that prioritizes long-term artificial general intelligence development over near-term revenue.
The commitment emerged from a leaked transcript of a four-hour meeting reported by Yicai. Liang argued that money is merely fuel for the deeper mission, and that open development and commercial monetization are not mutually exclusive.
This stance sharply contrasts with Western labs like OpenAI and Anthropic, which have locked their best models behind paid APIs. For DeepSeek, open weights act as a moat, building developer gravity that turns a model into an ecosystem. The approach has already proven effective: the lab stunned Silicon Valley last year when its low-cost V3 and R1 models matched far pricier US systems, despite Washington's curbs on advanced-chip exports.
Computing Power Is the Real Gap
Liang was blunt about constraints, describing computing power as his single biggest limitation. He framed the gap with the US as one of resource access rather than technical know-how. On chips, he expressed a preference for buying over building.
"I hope to be able to buy chips at a reasonable price so that I don't have to make chips myself," Liang Wenfeng, founder, DeepSeek
He confirmed plans to eventually build large computing clusters, but for now, the focus remains on securing external supply. This reflects a pragmatic acknowledgment of the hardware bottleneck that defines much of China's AI race.
The fundraising landscape has shifted. DeepSeek recently completed its first round at a roughly $52 billion valuation, reversing years of refusing outside capital. Pressure came from domestic rivals like ByteDance, Moonshot, and Z.ai, all racing in the same open-model space. The cash infusion may help bridge the compute gap without compromising the lab's open ethos.
AGI Through Narrow Focus
Liang outlined a roadmap from chain-of-thought reasoning to AI agents, then continual learning, and eventually embodied intelligence. He plans to leave areas like video generation and world models to others, keeping DeepSeek narrowly focused on the path to AGI.
This disciplined scope contrasts with the sprawling ambitions of many competitors. By avoiding distraction, DeepSeek aims to deliver the frontier's most capable open models, free to download. For developers globally, that could mean sustained access to cutting-edge tools without licensing fees.
The strategy aligns with research suggesting open-source models can accelerate innovation and safety through broader scrutiny, as explored in a recent paper on open foundation models. DeepSeek's own technical reports, such as the DeepSeek-V2 architecture, detail how efficiency gains can rival closed-source performance.
Liang's vision extends beyond chatbots, but his immediate bet is that openness will attract the developer ecosystem needed to sustain the AGI pursuit. Whether that bet pays off may redefine how the industry balances profit and progress.
