DeepSeek's $74B STAR Market Bet Resets the Bar for Every Chinese AI Listing

CITIC Securities is not a bit-player in this story. The firm ranked first across Asia-Pacific ex-Japan for investment-banking fees in 2025 (US$1.06 billion), led72 A-share underwriting mandates worth2,706 billion yuan of cash and asset deals, and just reported its strongest half on record: H1 2026 revenue of 49.69 billion yuan (+50% year over year) and net profit of 23.34 billion yuan (+69.6%), with investment-banking fees up 44% on the year. That record isn't coincidence — it is the by-product of a Chinese IPO market that has opened its arms to loss-making AI and 'future industry' issuers since the Shanghai Stock Exchange rewrote its STAR Market rules on June 17, 2026.

Two of the four underwriters already on the DeepSeek mandate are reportedly CITIC-led; the other three are unnamed in public reporting. The engagement follows the standard 'tutoring' (辅导) phase, after which a formal filing would land with the Shanghai Stock Exchange. The tutoring phase typically runs three months, which puts a credible application window in early 2027.

Strip out the narrative and DeepSeek's asking valuation does something the Chinese AI board has not seen before: it makes the country's flagship model lab worth roughly the same as its flagship AI chip designer — on a fraction of the revenue. Cambricon trades at a 673 billion yuan market cap (~$94 billion) after first-half revenue of 5.996 billion yuan (+108% year over year) and net profit of 2.311 billion yuan. Hygon Information Technology sits at 555 billion yuan (~$76 billion). DeepSeek's 500 billion yuan price tag would put it 79% the size of Cambricon on7% of the revenue.

The benchmark is even harsher against the model peers. Z.AI — Zhipu — listed in Hong Kong in January 2026 at a US$7.3 billion valuation and has since run to roughly US$52 billion on no profits. SenseTime at US$6.8 billion is the only listed Chinese generative-AI software name with a multi-year revenue track record, and it trades at 8x sales. DeepSeek's implied multiple on a US$70 million seven-month revenue run rate is the outlier: roughly 530x trailing sales, on a business still loss-making after11 billion yuan of infrastructure spending.

On June 17, 2026, the Shanghai Stock Exchange activated the 'fifth listing standard' for unprofitable AI large-model companies — a path that lets firms without positive net income in their most recent reporting year qualify if they can demonstrate 'strategic technology' status and meet a minimum revenue threshold tied to the broader STAR Market regime. The change also widened access to quantum technology, nuclear fusion, and brain-computer interface companies. DeepSeek will be the most-watched test of the new rule.

It is a deliberate Beijing policy choice. CSRC Chairman Wu Qing said the same week that regulators would 'guide more long-term patient capital into equity investments' to fund tech innovation — language that turns the DeepSeek mandate into a flagship for the broader AI capital-markets strategy rather than a one-off.

DeepSeek's infrastructure-spending bill — 11 billion yuan in just seven months — is the part of the story the listed names will price. Every additional training cluster DeepSeek brings online, and every API repricing pass (V4 Flash off-peak rates climbed over 10x on August 16, 2026), feeds demand for domestic inference silicon. The names sitting on that demand curve are the same ones DeepSeek's IPO will be benchmarked against: Cambricon, Hygon, Moore Threads, and MetaX Integrated Circuits.

The arithmetic is unforgiving. Cambricon's first-quarter growth ran at roughly 160% YoY; second quarter decelerated to 76% YoY, with Q2 revenue up only 7.8% sequentially from Q1. DeepSeek's own V4 pricing reset — cache-hit pricing up over 1,100% effective mid-August — confirms that frontier-model compute is moving from training-bound to inference-bound, where the silicon mix tilts even harder toward domestic suppliers under US export controls.

Six days before Reuters broke the DeepSeek story, Moonshot AI — the Kimi developer — confidentially filed for a Hong Kong IPO targeting a US$3 billion raise at a US$50 billion pre-IPO valuation. The two largest private Chinese model labs are now running dual-track listings on opposite sides of the border, and the disclosure gap between STAR and HKEX will start to look very wide very fast.

Hong Kong has the disclosure depth, the offshore capital pool, and a fresh comparable in Z.AI that has rallied over 24x since its January debut. STAR has the policy air-cover, the patient-capital mandate, and access to mainland retail flow. The two listings will be a real-time experiment in how each market values the same business model — and the spread on day one will probably be the most-watched number in Chinese tech since the Shein listing.

Reuters and the South China Morning Post both stress that no final size, valuation, or timing has been set. Until a prospectus prints, the work that matters is the work the underwriter does in the tutoring phase. Three numbers will decide whether the $74 billion tag holds in public markets.

If the prospectus clears those bars — and if CITIC's tutoring wraps before mid-Q1 2027 — the actual filing would land the same quarter Moonshot's Hong Kong book is expected to price. Two flagship AI listings on opposite exchanges, three weeks apart, will not be quiet affairs.