DeepSeek’s AGI Obsession: Rejecting Super Apps, Betting on Open Source, and the $74 Billion Plan to Beat US Chip Curbs

DeepSeek founder Liang Wenfeng has never been shy about his ultimate target. In a multi-hour investor meeting covered by Yicai and reported by Reuters on July 23, 2026, he made clear that DeepSeek’s singular goal is artificial general intelligence (AGI) — and that nearly everything else is a distraction .

The Hard No: What DeepSeek Will Not Build

Liang explicitly rejected building a consumer “super-app,” ruling out 3D generation, video generation, and world models . He described multimodality as “just a component, not intelligence itself” . This restraint is deliberate: “giving up certain things in exchange for a higher probability of achieving AGI” .

Development Roadmap: From Reasoning to Embodied Intelligence

Liang outlined a five-stage progression: chain-of-thought reasoning → AI agents → continual learning (AI that learns on its own) → AI that builds better AI → embodied intelligence. Coding agents are the near-term priority . He characterized this as a “gradual singularity” .

Open Source Is a Core Belief, Not a Strategy

Liang told investors DeepSeek is likely to keep its most advanced models open-source, arguing that open-source development and commercial monetization are not mutually exclusive . He stated that AI is “big enough (potentially 10% of global GDP) that you cannot monopolize it” and that open-sourced models are identical to those deployed in production . This is framed as a core belief, not a forced move .

First External Funding Round: $7.4 Billion from Tencent and CATL

DeepSeek closed its first-ever external funding round in June 2026, raising approximately 50 billion yuan ($6.9–$7.4 billion) . Major investors include Tencent Holdings and battery giant CATL . The post-money valuation was reported at over 350 billion yuan (~$52 billion) . Notably, founder Liang Wenfeng himself contributed roughly $3 billion of the total, and most investors received no voting rights .

IPO Plans: Shanghai STAR Market

DeepSeek is preparing for an initial public offering on the Shanghai Stock Exchange’s STAR Market . Per the Wall Street Journal, the company is targeting an IPO as early as Q2 2027 . Multiple outlets report plans to submit a filing by the end of 2026 with a potential listing in 2027 . The WSJ and Reuters note that DeepSeek has begun discussions with investors and banks .

Fresh Funding Round at ≈$74 Billion Valuation

Weeks after closing the June round, DeepSeek began talks for a fresh fundraising round at a valuation of approximately 500 billion yuan ($74 billion) — a ~37% jump from the June round valuation . Reuters reported the round at about $74 billion, and Bloomberg/FT reported a pre-money valuation target of $71 billion .

The Compute Gap: “The Biggest Constraint”

Liang described computing power as the biggest constraint, saying the main gap with the US is access to resources rather than technical know-how . He stated: “I hope to be able to buy chips at a reasonable price so that I don’t have to make chips myself” . He acknowledged that the performance gap between available Chinese chips and US chips remains significant .

Betting on Huawei: The Domestic Chip Ecosystem

Liang expressed optimism that China can break its Nvidia dependence . He called the replacement of Nvidia chips with domestic alternatives “a historic inflection point,” stating there are no hardware or ecosystem compatibility barriers preventing adoption . DeepSeek has optimized its V4 model specifically for Huawei’s Ascend 950 chips (including the 950PR super-node) . The Huawei Ascend 950PR “greatly surpasses” Nvidia’s H20 (the most advanced chip Nvidia could previously sell in China) but still trails Nvidia’s H200 . Huawei expects ~$12 billion in AI chip revenue in 2026, driven partly by DeepSeek demand .

Efficiency: Adaptation to Scarcity, Not Philosophy

Liang framed DeepSeek’s celebrated efficiency — building models at a fraction of US competitors’ cost — as an adaptation to scarcity rather than a philosophical choice . The company’s API pricing aims only to recover hardware costs within ten months . As one account summarized: “Restraint is a strategy — giving up certain things in exchange for a higher probability of achieving AGI” . DeepSeek’s weighted average cost for a standardized intelligence task (V4 Flash) is just $0.02 versus $2.75 for Anthropic’s Claude Fable 5 .

Key Caveats

Many of these details come from a single investor meeting as reported by Yicai and relayed by Reuters, plus supplementary reporting from Bloomberg, WSJ, and FT. Some of the most granular quotes (e.g., the exact five-stage roadmap, “gradual singularity” framing, and specific rejections of 3D/video/world models) are sourced from second-hand summaries rather than a direct transcript. The IPO timeline and fresh round valuation are still described as “plans” and “discussions” — not finalized.