DeepSeek ARR hits $1B; >70% compute to training

Opening

DeepSeek’s annualised revenue run rate has hit about $1 billion — more than double from a few months ago — The Information reported Thursday, citing two people with direct knowledge, per Reuters. Reuters (Abu Sultan) says CEO Liang Wenfeng shared the figure at a recent investor meeting. Growth was partly driven by raising model pricing 2.3 to 4.5 times, The Information said. The firm is pushing a second funding round targeting 50 billion yuan (~$7.45 billion) at a 500 billion yuan valuation by end of October and preparing a potential Shanghai Stock Exchange listing, per The Information via Reuters. Reuters could not verify; DeepSeek was not immediately reached. Separately, Reuters had reported this month that DeepSeek hired CITIC Securities for a potential STAR Market IPO with timing, valuation, and size still undecided. Liang told investors the company still allocates more than 70% of computing capacity to training and less than 30% to inference, The Information added via Reuters. The company released DeepSeek-V4.1-Flash earlier this month.

The numbers

ARR / run-rate — ~$1 billion; more than double vs a few months earlier — The Information via Reuters (two people with direct knowledge).

Model price hikes — 2.3× to 4.5× — The Information via Reuters.

Fundraise target — 50 billion yuan (~$7.45B) at 500 billion yuan valuation by end October — The Information via Reuters (correction: fundraise, not IPO, for this round).

Compute split — >70% training / <30% inference — Liang to investors, The Information via Reuters.

Why it matters

A dated China lab ARR print with an explicit training-heavy compute split and price-driven revenue step-up. Does not by itself settle hyperscaler capex or U.S. lab margins; it does put a falsifiable China-side revenue and allocation claim on the tape next to export-control / domestic-chip narratives.

Positioning

AI capex durability — Training share >70% while ARR doubles is consistent with continued training demand; the theme still breaks on a top-four hyperscaler sequential capex cut or inference cost destroying new-build need — neither is in this print.

Inference margin inversion — Price increases (2.3–4.5×) are the opposite of U.S. frontier list cuts already in the archive; China lab pricing and U.S. lab pricing are not the same market.

Connects to

Alibaba Zhenwu V900 / >20GW — Same week’s China domestic-chip / capacity tape (news).

Anthropic / OpenAI cheaper models — Contrast on list-price direction (news).