"Money has never been the problem for us. Bans on shipments of advanced chips are the problem.
"Liang Wenfeng
DeepSeek: The Lab That Changed Everything
In January 2025, a small Chinese AI lab released an open-source reasoning model that erased $593 billion from Nvidia's market cap in a single day. The model matched OpenAI's best for a reported training cost of $5.576 million. Marc Andreessen called it "AI's Sputnik moment." For 18 months after that shock, DeepSeek refused every investor, every cheque, every overture from China's biggest tech firms. Then, in June 2026, it opened the door and raised $7.4 billion in one round at a $52 billion valuation. The AI race just tilted again.
First external funding round
$7.4 billion (50 billion yuan), June 2026
Post-money valuation
approximately $52 billion
Second round in talks
targeting $71 to $74 billion pre-money
Employees
approximately 300, plans to double to 400
Founded
July 2023, Hangzhou, China
IPO target
Shanghai Star Market, possible debut Q2 2027
The Round: Who Wrote the Cheques
The investor list reads like a who's who of Chinese technology and industry. Tencent led with 10 billion yuan. CATL, the world's largest EV battery maker, contributed 5 billion yuan, a bet that AI and energy storage are converging. JD.com, NetEase, and IDG Capital each put in 3 billion yuan. China's national AI industry fund invested 1 billion yuan directly into DeepSeek's core entity and, unusually, retained voting rights. Every other external investor accepted a lock-up of up to five years with no votes. Liang Wenfeng himself contributed 20 billion yuan of his own money, nearly half the round, cementing his 78 percent equity stake.
DeepSeek trained a world-class AI model for $5.6 million while US rivals spend \
billions. Now it has $7.4 billion more. Does algorithmic efficiency plus capital make it unstoppable, or does the West's chip advantage still matter?
R1 training cost (reported)
$5.576 million vs billions for US rivals
AIME 2024 math benchmark
79.8 percent (vs OpenAI o1: 79.2 percent)
MATH-500 score
97.3 percent (vs o1: 96.4 percent)
API input price
$0.28 per million tokens (vs GPT-4o: $2.50, roughly 9x cheaper)
China market share (early 2026)
89 percent of AI API and app usage
Nvidia market cap erased on R1 launch day
$593 billion (January 27, 2025)
The Architecture Edge DeepSeek-R1 uses a Mixture-of-Experts design: 671 billion total parameters but only 37 billion activate per inference pass. That keeps compute proportional to a much smaller model. The technique is not new, but DeepSeek applied it at a scale and cost efficiency that rivals had not. DeepSeek-V4, released April 2026, added a native 1 million-token context window and training on 32 trillion tokens.
The Control Structure External investors face a lock-up of up to five years and hold zero voting rights. The sole exception is China's national AI fund, which retained a vote. Liang Wenfeng holds 78 percent equity through a special structure. One analyst noted: "When decision-making power is this concentrated, the cost of correcting a strategic obsession is extremely high." Liang told investors: open-source research comes before short-term commercialisation. Always.
What the Money Is For
The immediate priorities are threefold. First, computing capacity: DeepSeek's internal chip stockpile remains a genuine bottleneck even after recent GPU buys, and Liang has told the Chinese premier as much. Second, the team: headcount will roughly double from 300 to 400, a tiny number by US frontier-lab standards. Third, a longer-term bet on building its own AI chip, reducing dependence on Nvidia hardware subject to US export controls. The IPO on Shanghai's Star Market, expected no sooner than Q2 2027, would give DeepSeek access to public capital markets and a currency for future deals.
The Export Control Shadow US export restrictions on advanced chips have directly shaped DeepSeek's story at every turn. R1 was trained on Nvidia H800 chips, which were not restricted until October 2023. The House Select Committee on the CCP has accused DeepSeek of stockpiling chips before controls took effect and recommends expanding restrictions further. A recurring irony in analyst reports: the pressure to innovate around chip scarcity is precisely what drove the algorithmic breakthroughs that
