DeepSeek reaches $1B annual revenue, plans to raise $7.5B at $75B valuation

DeepSeek, the Hangzhou-based AI lab that sent shockwaves through Silicon Valley when it emerged as a serious contender in early 2025, has hit a $1 billion annualized revenue run rate as of September 2026. That figure doubled from $500 million reported just months earlier, powered almost entirely by a single revenue stream: API access.

The company is now finalizing a second external funding round targeting approximately $7.5 billion (50 billion yuan) at a $75 billion (500 billion yuan) valuation, expected to close by the end of October 2026.

How DeepSeek doubled revenue without doubling customers

The revenue surge wasn’t driven by a sudden flood of new users. It was driven by price hikes. DeepSeek raised its API pricing by 2.3x to 4.5x, a move that would typically send enterprise customers scrambling for alternatives.

Except it didn’t. CEO Liang Wenfeng noted that customer demand remained strong despite the increases, and the company still positions itself as one of the most competitively priced providers for major AI models.

DeepSeek’s consumer chatbot, which competes with ChatGPT and similar products, remains completely free. Every dollar of that $1 billion run rate comes from developers and enterprises accessing DeepSeek’s models through its API.

Gross margins paint an equally compelling picture: 82.9% through July 2026.

From hedge fund to AI juggernaut in three years

DeepSeek was founded in July 2023 by Liang Wenfeng, who previously built High-Flyer, a quantitative hedge fund.

The company closed its first external funding round in June 2026, raising approximately $7.4 billion at a post-money valuation of around $52 billion. The current round, if completed at the reported terms, would represent a 44% valuation jump in roughly four months.

DeepSeek has also enlisted CITIC Securities as the lead underwriter for a planned initial public offering on the Shanghai STAR Market, targeted for 2027.

More than 70% of DeepSeek’s computing capacity is currently dedicated to training new models.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.