China's AI IPO Gold Rush: DeepSeek and Moonshot Race to Go Public

*China's foundation-model labs spent years refusing venture capital. In four months, they have raised tens of billions of dollars and hired underwriters. (Image: Unsplash)*

The email arrived at a Hong Kong family office on a Tuesday morning in late August 2026, one of eight similar messages the investment team had received in a single week. The subject line was bland — "DeepSeek access vehicle, final allocation" — but the terms inside were anything but. To buy a stake in the hottest AI company on earth, the family office would need to commit at least ¥10 million (US$1.4 million) into a special-purpose vehicle already two layers removed from the actual company. The first-layer SPV charged a 6% entry fee; the second layer, 8%. Some outer-layer vehicles were quoting front-end fees above 15% and profit participation of up to 40%. Investors had no voting rights. Their capital would be locked for five years.

They subscribed anyway. Within days, the allocation was oversubscribed.

This is what the next phase of China's AI race looks like from inside the money. Not benchmark charts or GPU clusters, but a scrum of bankers, sovereign funds, state-backed industrial capital, and desperate family offices fighting for a place on a cap table that did not exist four months ago. On September 9, Reuters reported that DeepSeek — the Hangzhou lab whose V4 models now serve trillions of tokens a week — had hired CITIC Securities, China's largest investment bank, to prepare an initial public offering on Shanghai's STAR Market, targeting a filing within the year. Days earlier, Beijing rival Moonshot AI had quietly submitted a confidential A1 filing to the Hong Kong Stock Exchange. Zhipu and MiniMax are already listed in Hong Kong. Unitree Robotics went public as the "humanoid robot first stock" this year.

The labs that once dismissed capital markets as a distraction have decided that the war will be won by whoever can finance compute, talent, and iteration fastest. The gold rush is on — and like every gold rush, the people selling shovels are doing arithmetic the miners would rather not see.

The Refusenik Opens the Vault

For two years, Liang Wenfeng ran DeepSeek like a research institute that happened to have a hedge fund attached. He turned away venture capitalists while rivals like Moonshot and Zhipu raised billions. The lab's stated philosophy — that profit pressure corrupts long-horizon research — became part of its brand, and its sudden dominance in January 2025 was celebrated partly because it seemed to prove that capital wasn't the point.

That era lasted exactly until the economics of frontier AI caught up with the philosophy.

Between April and September 2026, DeepSeek completed one of the most dramatic corporate reversals in Chinese technology history. The details matter, because they reveal how quickly the ground has shifted under the entire industry:

The speed is the story. DeepSeek's first round closed in June at a valuation above US$50 billion. Its second round, opened weeks later, was pitched at roughly ¥500 billion — a ~50% step-up before a single new product shipped. In the four months between April and September, the company's enterprise value grew faster than its own model iterations.

The 28-Day Flagship and the Burn Beneath It

Those model iterations are now measured in weeks. DeepSeek V4 Pro launched on August 13, 2026. On September 10 — 28 days later — every request to the flagship was silently rerouted to the new V4.1 Flash, a 552-billion-parameter mixture-of-experts model with native multimodal vision, billed at the cheaper Flash rate with cached-input prices cut by as much as 60%. Chinese developers have already written the obituary memes: *born August 13, passed September 10, survived by its own replacement.* For context, OpenAI's GPT-5.6 family and Anthropic's Claude Fable 5.1 were both still flagship-current on the same day.

The product velocity is only possible because the burn is enormous. Leaked disclosures from August sketch a company growing into its valuation at a pace that has no precedent outside wartime logistics:

The hiring surge is the tell. Cui Tianyi, who leads DeepSeek's Harness team, explained that data volume, machine count, training tasks, agent environments, and request load have grown so complex that the existing backend must be "upgraded, maintained, and rewritten" — hence senior engineers with two to ten years of experience, tested with a newly designed exam that, by Cui's own account, stumped gold-medal-level fresh graduates. This is not a research boutique dabbling in deployment. It is an industrial operator staffing up for permanent war — and it is spending like one.

What changed? Money, at industrial scale. The lab that trained R1 on scraps now competes against ByteDance (¥160 billion AI budget for 2026, half earmarked for chips), Alibaba (a record HK$80 billion share placement in August), and Tencent simultaneously. Research purity is a luxury that Liang Wenfeng, whatever his preferences, can no longer expense to a quant fund.

The Shadow Market and the Real One

The desperation inside that Hong Kong family office's inbox is a direct product of scarcity engineering. DeepSeek's second round comes with terms that would be laughed out of a Silicon Valley term sheet: five-year lockups for first-round indirect investors (now being restructured to three years to satisfy listing rules), no voting rights for external capital, and a founder personally auditing the ultimate identities behind every limited partner to keep the shareholder registry clean for regulators.

Yet demand has manufactured its own shadow economy. According to the Financial Times, multiple vehicles holding DeepSeek allocations began reselling access to outside money at escalating markups — layer upon layer of SPVs, each taking a slice, some quoting 40% profit participation on top of 15% entry fees. People close to the company say DeepSeek did not authorize any of it, and that Liang Wenfeng is now personally reviewing the final investor list to prevent shares from ending up with "unknown entities" — a governance cleanup that normally happens with a seasoned CFO and investor-relations team. DeepSeek has neither. It has a founder with a spreadsheet and a grudge against opacity.

*Shanghai's Lujiazui financial district. The STAR Market's revised fifth-set standard gives loss-making hard-tech companies a direct path to public capital. (Image: Unsplash)*

The irony is that the shadow market exists because the real market is being deliberately widened for exactly these companies. The Shanghai Stock Exchange's revised fifth set of STAR Market listing standards — designed for leading, temporarily unprofitable hard-tech firms — removes the revenue-profit thresholds that would have blocked DeepSeek for years. Hong Kong's Chapter 18C regime has already absorbed Zhipu and MiniMax, and is now preparing to absorb Moonshot. China's regulators, who spent 2021 teaching tech platforms a lesson about capital discipline, have spent 2026 building on-ramps for the AI labs.

The question regulators are quietly betting on: would you rather have DeepSeek's compute war financed by transparent public markets, or by a teetering stack of offshore SPVs charging 40% carry? On September 9, the answer became obvious to everyone holding an allocation.

The Kimi Moment, Priced in Dollars

If DeepSeek's IPO is a story about capitulation to capital, Moonshot AI's is a story about capital capitulating to a model.

On July 16, Moonshot released Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model with native 1-million-token context, released with open weights under a modified MIT license. Hours later it topped the Arena frontend-coding leaderboard with 1,679 points, ahead of Anthropic's Claude Fable 5 (1,631) and OpenAI's GPT-5.6 Sol (1,618). Within 30 minutes of its Hugging Face upload, K3 had logged over 4,000 likes — the fastest-rising model release the platform had ever recorded. OpenAI's own president reportedly told investors the US-China model gap might now be measured in months, not years.

The revenue curve followed the benchmark curve with an almost eerie correlation:

Caixin's mid-September cover story — literally titled "The Kimi Moment" — framed the sequence as China's second DeepSeek shock. The capital markets agreed. Moonshot completed its ¥31.5-billion-valuation round in late July, immediately opened a Pre-IPO round targeting US$50 billion, and in the first week of September filed confidentially with the Hong Kong exchange. People familiar with the process say the listing could come within six months. At US$50 billion, Moonshot would price at rough parity with Zhipu's entire listed market capitalization — a remarkable statement about how quickly benchmark dominance converts into valuation when the buyer is an exchange full of momentum capital.

The cap tables of China's two most valuable model labs are now nearly mirror images of each other: Alibaba, Tencent, Meituan, IDG, and state-affiliated funds appear on both. The competitive frontier in Chinese AI is increasingly financed by the same handful of balance sheets, which means the rivalry that matters — the one between labs — is becoming a contest of engineering execution rather than capital access. That, more than any single model release, is the structural shift of 2026.

The Class of 2026: Everyone Is Going Public

Step back and the pattern stops looking like a series of deals and starts looking like a sector-wide liquidity event. The table below tracks the capital-markets wave that has swept through Chinese AI in 2026 alone:

Two observations stand out. First, the window is policy-engineered: STAR Market's relaxed standards and Hong Kong's 18C rules were rewritten precisely so this cohort could list before profitability. Second, the giants are financing in parallel — ByteDance chose debt (a US$29.6B syndicated loan) while Alibaba chose equity, and both explicitly earmarked proceeds for AI compute. Whatever else China's AI buildout is, it has become the largest coordinated capital-allocation event in the history of the country's technology sector.

The public-market step changes the psychology of the race in ways private rounds never could. A listed Zhipu must answer to analysts every quarter about GLM-5.3's gross margin on tokens that lose value weekly. A listed DeepSeek would disclose the economics of the V4 family to the world — including, presumably, to competitors in San Francisco. Listing buys compute, but it also buys transparency, and transparency is a weapon your rivals can load too.

The Global Mirror: Shanghai and Hong Kong vs. The SEC

The parallel nobody in Beijing or Hangzhou discusses publicly: America's frontier labs are running the same play, on nearly the same clock. OpenAI and Anthropic have both filed confidential S-1 registration paperwork with the US Securities and Exchange Commission, with listing windows reportedly targeted between late 2026 and early 2027. The two superpowers' flagship AI companies are about to do something that would have been unthinkable in 2023 — compete for public capital at the same time, in public.

Underneath the valuation headlines, a quieter metric keeps validating the Chinese cohort: usage. OpenRouter data for the week of September 7–13 shows global model call volume up 10.4% to 127 trillion tokens — and Chinese models occupying eight of the top ten slots. GPT-5.6 Luna led at 15.8 trillion tokens, but DeepSeek V4.1 Flash (11.8T) and Tencent's Hy4 preview (11.6T) followed immediately, with Xiaomi and Zhipu also in the top ten. The market that global developers actually pay for at scale is, increasingly, a Chinese-dominated league table — which is precisely the revenue evidence underwriters need to sell these IPOs to institutional investors.

The difference is architectural. America's listings will concentrate AI capital in two closed-model giants. China's listings are spreading capital across a half-dozen model labs plus an entire semiconductor supply chain that IPO'd earlier in the cycle. One model bets on moats; the other on redundancy. The public markets of 2027 will render a verdict one quarter at a time.

Why Capital Velocity Now Matters as Much as Compute

Strip away the fees and valuations and the September gold rush resolves into a simple strategic logic: in a war where model generations turn over in 28 days, the winner is whoever can close the loop between research breakthrough and funded infrastructure fastest.

Private markets were too slow. DeepSeek's own experience is the case study — it needed two rounds, a shadow market, and an IPO to assemble the compute budget that ByteDance allocates in a single annual budget cycle. Public markets offer something venture cannot: repeatable, near-instant equity issuance (as Zhipu's two post-IPO placements in nine weeks demonstrate), employee equity that actually prices (critical when your researchers are being poached by cash-rich giants), and a currency — listed shares — for acquiring the agent-tooling startups now flooding Shenzhen and Hangzhou.

There are costs, and they are real. A listed DeepSeek must disclose the economics behind the token price war it is currently winning; a listed Moonshot must explain to Hong Kong retail investors why K3's pricing power survives when DeepSeek keeps cutting. Quarterly earnings season will do to Chinese AI labs what it did to the internet platforms after 2021: reward the profitable and expose the burning. And the governance cleanup — Liang Wenfeng personally vetting LPs, lockups shortened to three years to satisfy listing rules, no CFO in place at either DeepSeek or Moonshot — suggests these companies are listing at the precise moment their organizational adolescence ends.

What Comes Next: The Pipeline and the Risks

The deal calendar now writes itself. DeepSeek is expected to submit its STAR Market application before the end of 2026, with listing in 2027; market speculation on post-IPO valuation runs from ¥1.5 trillion to ¥2.5 trillion, numbers that would instantly make it one of China's most valuable listed companies and would, incidentally, likely crown Liang Wenfeng among the country's richest individuals. Moonshot's Hong Kong listing could beat it to market if the six-month timeline holds. Behind them wait the chip suppliers (already public and repricing), the agent-framework startups (too small, for now), and the embodied-AI cohort led by Unitree.

The risks are equally legible. Valuation compression is the obvious one: paying ¥500 billion for a company with ¥475 million in seven-month revenue requires believing the token economy compounds for years without a margin crash. Geopolitics is the tail risk — a listed DeepSeek is a DeepSeek whose share registry, customers, and suppliers become transparent to sanctions architects. And burnout is the quiet risk: labs that went from 200 employees to 1,000 in a year, now hiring 150 senior engineers in a single week with rewritten exams, are betting that culture scales faster than cap tables.

But the direction is set. The labs that changed China's AI standing in 2025 did it with research. The ones that intend to hold it in 2027 will do it with research funded by the public — literally, share by listed share. The family offices paying 15% fees for shadow access are betting the window stays open. Every week of 127-trillion-token usage says it does.

*What the internet is saying*

@quantdance_ on X (Twitter):

"Wall Street spent 2025 asking 'where's the China AI revenue?' The answer this week: DeepSeek did ¥475M in 7 months with ¥11B in infra spend and it's still the most coveted IPO in Asia. The revenue isn't the product. The capacity is." — *A quant investor noting that investors are buying compute optionality, not current revenue.*

知乎用户 · 如何评价 DeepSeek 冲击科创板 IPO？

"融资5000亿估值、收入4.75亿，这个市销率放在任何市场都是天文数字。但换个角度：全国产算力链路上唯一能打的前沿模型，稀缺性本身就是估值。" — *"A ¥500B valuation on ¥475M revenue is an astronomical price-to-sales ratio by any measure. But flip the perspective: it's the only frontier model that runs on fully domestic compute. Scarcity is the valuation."*

小红书网友 · 金融民工日记：

"中信的朋友说 DeepSeek 的项目组已经连轴转一个月了，尽调材料堆成山，梁老板亲自盯股东名单，据说是怕'不明身份的钱'混进来。这届 IPO 尽调跟反洗钱似的。" — *"A friend at CITIC says the DeepSeek deal team has been working flat out for a month, due-diligence materials piled like a mountain. Boss Liang is personally watching the shareholder list, reportedly afraid of 'unidentified money' slipping in. This IPO diligence is basically anti-money-laundering work."*

微博网友 · @科技瞭望台：

"DeepSeek上市、Kimi上市、智谱上市、MiniMax上市——国产大模型全员证券化。希望募资真的砸向算力和人才，而不是砸向PPT和市值管理。" — *"DeepSeek listing, Kimi listing, Zhipu listing, MiniMax listing — the whole domestic foundation-model cohort going public. Hopefully the raised money actually goes to compute and talent, not to PowerPoints and market-cap management."*

GitHub · deepseek-ai/DeepSeek-V4.1-Flash discussion:

"Everyone in this thread arguing about whether V4.1 Flash beats Opus 5 on Terminal-Bench is missing the real story: this model launched at up-to-60%-off pricing because the company behind it is about to file audited financials. The price war IS the IPO roadshow." — *A developer connecting the aggressive V4.1 Flash pricing directly to pre-IPO financial positioning.*

豆瓣小组 · 上班这件事：

"150个社招HC、应届金牌选手都做懵的笔试，看着像技术理想主义者的公司也开始抢人大战了。上市之后的DeepSeek，还会是那个'慢公司'吗？" — *"150 open positions and an entrance exam that stumps gold-medal graduates — the company that looked like a haven for technical idealism is now in the talent war too. After the IPO, will DeepSeek still be the 'slow company'?"*