DeepSeek Is Going Public. Bet the Monopoly, Not the Multiple.

Hi {{first_name|there}},

Quick note before the issue.

Yesterday you got an accidental early send of this draft. That was on us. iki, my overzealous AI chief of staff, pushed a test live. Sorry for the noise. This is the final version.

This is the first East-West AIpha built end to end without human assistance. I work directly with Kimi on East-West research, and with Grok Bot and Muse AI on this issue.

Thank you for being an existing subscriber. You get East-West AIpha free for one year. If this is useful, send it to a friend. The next 80 referrals also get a free year. After that, paid, $38/month.

Why this issue matters: DeepSeek's IPO is not just another listing. China is effectively crowning DeepSeek as the national AI champion, an avant-garde pick that reprices the stack for Chinese and Western model builders. That is the bet inside.

San

Automate Work, Amplify Life.

East/West AIpha — for people putting real capital and careers on it

(Tuesday, September 16th)

DeepSeek has tapped 中信证券 CITIC Securities to prepare a STAR Market IPO. Nothing is filed yet: no tutoring agreement, no Zhejiang regulator filing. But the machine has started, and Shanghai rewrote its listing rulebook in June with AI model companies in mind.

Everyone will cover the valuation number: 5,000亿元 pre-money, roughly $71 to $75B depending on the FX print you use. That is not the story.

The story is that DeepSeek makes money in a way no Western lab does, on terms no Western founder could demand. Five-year lockups. Zero voting rights. Identity-vetted investors paying double-digit fees for the privilege. And it is listing into a market that just watched 智谱 Z.AI round-trip a trillion-Hong Kong-dollar valuation in nine months.

This issue is about the money. How each Eastern lab actually earns it, how that differs from the Western model, and what DeepSeek's prospectus will reveal for the first time.

DeepSeek picked CITIC to prepare the STAR Market listing. As of mid-September (21财经, 财联社, 澎湃), no tutoring agreement is signed and no filing sits on the Zhejiang CSRC site. The internal target is filing in 2026, but mandatory tutoring runs at least three months. Year-end is tight. SCMP (26 Aug) sees earliest filing end-2026, listing in 2027.

The funding walk, per Chinese sources: the June first round closed at over 3,500亿元 post (~$52B) with 500亿元 raised. 腾讯 Tencent put in 100亿元, 宁德时代 CATL 50亿元. The current round prices the company at 5,000亿元 pre-money and targets another 500亿元, with 中芯聚源 SMIC Capital, 博裕 Boyu, CPE源峰 and 合肥国资 at the table.

The reported financials (The Information, Jan to Jul 2026, unaudited as reported): revenue 4.75亿元 (~$70.7M), roughly ten times all of 2025. Net loss 7.15亿元. Blended gross margin 44.6%, API gross margin 82.9%. AI infrastructure spend of ~110亿元 in seven months, nearly ten times 2025's full-year spend. About 300 staff, hiring 150 more. (~$450M annualized from the seven-month run-rate is arithmetic, not a company guidance figure.)

The Western frame is binary: closed labs sell tokens (OpenAI, Anthropic); open labs give models away and monetize something else (Meta). China has produced at least four distinct monetization machines. DeepSeek is a fifth.

DeepSeek. 深度求索 DeepSeek, the third path: open the weights, charge for the pipe. DeepSeek open-sources everything under MIT and still runs a paid API priced at roughly one-tenth of Western flagships. The flywheel: open weights make DeepSeek the standard; most enterprises cannot self-host, so traffic returns to the official API. Liang's pricing doctrine: peg API prices to recovering hardware cost in about ten months. He has said that even at double the price, usage barely drops. The company deliberately refuses profit maximization. In March 2025 DeepSeek disclosed a 545% theoretical inference cost-margin. In August it moved to peak/off-peak pricing and cut Flash prices up to 60%. V4-Pro off-peak now runs $0.66 per million input tokens, $1.98 output, versus Claude Sonnet 5 at $2/$10 and GPT-5.6 Sol at $4/$20 (Sept price pages). Not a company failing to monetize. It monetizes at the infrastructure layer, where the ceiling is higher.

Z.AI. 智谱 Z.AI, from government contractor to token merchant. In 2025, 73.7% of Z.AI's revenue came from heavyweight private deployments for government and enterprise clients. By H1 2026 the mix had inverted: 86.5% of revenue is now API, ARR passed $1.6B by end-August, 7.4 million developers on the platform, token volume up 40x year-to-date. The cost of the pivot: blended gross margin fell from 50% to 26.4%, though API gross margin itself turned positive at 24.6%. Z.AI raised API prices 83% this year and volumes rose anyway; nine of China's ten largest internet companies call GLM daily, and it sells token bundles on Tmall like phone credit. (H1 2026 interim report)

MiniMax. 米奈 MiniMax, the overseas subscription company. MiniMax is the mirror image: 73% of revenue from outside mainland China, driven by consumer subscriptions (Talkie, Hailuo video), API a secondary line. First of the cohort to break even (2024, on Talkie), 17.9% gross margin trending up, over $1B in cash. The closest Chinese analog to the Anthropic shape: product-led, global, consumer-plus-API.

Moonshot. 月之暗面 Moonshot (Kimi), premium pricing under a compute ceiling. Kimi K3 output is priced around ¥100 per million tokens, the most expensive domestic tier, a bet that benchmark leadership converts to pricing power. The constraint is physical: domestic AI compute demand up 417% year-on-year against 128% supply growth. Moonshot paused new consumer sign-ups and reallocated GPU to paying traffic; a heavy user's compute cost can exceed the ¥59 monthly subscription. ARR went from $100M (March) to a reported $400 to $500M mid-year; Nomura projects $1B by year-end. An A1 filing went in confidentially the week of 2 September; a Hong Kong-plus-STAR-Market dual listing is under exploration. Full disclosure, as always: WorkOptional is a heavy Kimi user. We watch this one from the inside.

Qwen. 阿里通义千问 Alibaba Qwen, open source as a cloud funnel. Alibaba gives away Qwen weights because the model is not the product. The cloud is. AI-related product revenue has grown triple-digits for eight straight quarters; Alibaba Cloud grew 26% last quarter on ¥38.6B of AI capex, up 220% year-on-year. Every company deploying Qwen open weights is a future cloud customer.

The Western model, for contrast. Anthropic: 75 to 85% of revenue from usage-based enterprise API, net revenue retention of 500%, inference gross margin up from 38% to a reported 70 to 85%, first $1B+ quarterly profit in Q3. OpenAI: 65%+ from consumer subscriptions, still loss-making. The subscription model's marginal-cost problem in reverse: heavy users eat the light users' fees. Google sells the full stack, chips to Workspace.

Subscribe to East/West AIpha

This newsletter is normally $38/month. For the next 88 subscribers, it's free for a year.

Yes, with one firewall.

DeepSeek V4-Pro at peak pricing is still cheaper than every Western flagship's standard tier; off-peak it is half of that. For the roughly 80% of enterprise workloads that need "good enough" (email triage, customer service, first-pass code, document processing), a model at one-tenth the price and ~90% of the capability collapses the price umbrella every closed lab charges under. Anthropic's September letter, 151M+ distillation exchanges traced to accounts linked to Chinese labs, backhandedly acknowledges the capability gap is being arbitraged.

The firewall is trust, not technology. Western enterprises with sensitive data will not route workflows through Chinese servers at any price, and Washington is building the wall higher with blacklist language around Chinese model companies. Likely equilibrium: a bifurcated market. Western labs keep the premium tier; Eastern labs take the price-sensitive global majority and all of domestic China. The premium tier can thrive, on a smaller share of a bigger pie.

For investors the trade is not "East beats West." It is this: the East is repricing what intelligence costs, and every Western AI multiple is implicitly long the old price.

The capital structure is the tell. The state AI fund invested 10亿元 directly, the only outside investor with voting rights, lockup-exempt. Everyone else enters through the 宁波程恩 partnership where Liang is GP: five-year lockup, zero votes, LP identities vetted. Demand so far exceeds access that a shadow market of layered SPVs has formed: 6% front fees at layer one, 8% at layer two, up to 18% entry plus a 35% profit cut at the outer edge, minimum tickets ¥10M. DeepSeek itself is not involved (FT, 9 Sep). New-round lockups may shorten to three years to fit STAR Market rules.

No Western frontier lab looks like this. OpenAI and Anthropic sell governance for capital. DeepSeek sells nothing but exposure, with the state as the only direct voter. At listing, public investors buy a claim on a national champion whose control is deliberately unreachable. Price that.

Subscribe to East/West AIpha

This newsletter is normally $38/month. For the next 88 subscribers, it's free for a year.

Two of DeepSeek's cohort are already public in Hong Kong, and the tape is the lesson (Wind close 15 Sep 2026; prices and market caps as reported — cross-check live quotes):

智谱 Z.AI (02513.HK): HK$665.5, market cap ~HK$310B (~$40B). Up 406% from January, after peaking at HK$2,980 in June: a ~HK$1T round trip in nine months.

米奈 MiniMax (00100.HK): HK$232.8, market cap ~HK$81B (~$10.4B). Down 32.5% from January, after peaking at HK$1,330 in March.

Both fell hard on 15 September (Eastmoney: Z.AI −5.69%, MiniMax −7.05%) as the DeepSeek IPO news cycle ran. The unlisted leader is already the index. Hong Kong's IPO market: H1 2026 raised HK$209.9B (+92% YoY), second globally.

Chinese finance media saw the repricing before we did. 碧山 Bi Shan, writing in 博望财经: "它的万亿市值从来不是业绩撑起来的，是'大模型第一股'这个稀缺身份撑起来的。当稀缺性消失，市值自然会找到它该有的位置" — "The trillion-dollar valuation was never supported by performance but by the scarcity of being the first listed model stock. When the scarcity disappears, the market cap finds its true level."

(WorkOptional private disclosure — not a public filing.) Recently we were offered Moonshot AI's Series G at a $50B pre-money valuation (raise sized around $3B in Reuters coverage, up to about $5B in Bloomberg reporting; up to $60M allocation, 3% fee, 20% carry). Kimi K3 is the largest open weights model ever at 2.8T parameters, with revenue it told investors growing from $300M in June to $1B in August (not audited).

When the 招股说明书 drops, eight things to check:

Order Book 1 · Book 2

Commandment Four: Don't be a dead fish. Dead fish go with the flow. Live fish swim.

The Western consensus on Chinese AI is a strong current. Uninvestable. Copycat. State-run. Drifting with it feels safe. It is also how you miss every turn that matters.

水涨船高 (shuǐ zhǎng chuán gāo). When the water rises, the boat rises. Beijing is raising the water: the rulebook rewritten, the state fund seated, the exchange rebuilt for listings like this. The question was never whether the water rises. It is whether you are in the boat.

(The dead fish had a thesis too. It was called consensus.)

East reads: a policy-ratified coronation. The exchange rewrote the rulebook, the state fund holds the only outside vote, and the sector gets repriced off DeepSeek's multiple. As 36氪's “World Model Factory” column put it, once DeepSeek lists, Chinese model companies stop being priced as OpenAI-at-a-discount and start as fractions of DeepSeek. Chinese press names the risk: a mega-raise siphoning capital from every smaller AI listing.

West reads: either “the next Ant IPO, policy risk with a ticker,” or the first audited look at frontier-lab unit economics from the price-disrupting side.

Our read: the IPO matters less than what the IPO forces into the open. For the first time, a frontier lab built on cost destruction will publish audited numbers. If the reported 82.9% API gross margin survives scrutiny, the price umbrella is not leaking. It is gone.

Here is our position. We are long — not because the multiples work.

They do not. $75 billion on a reported ~$450 million of annualized revenue is 167 times sales. On the revenue ruler, it is overvalued.

We are long for the same reason we invested in SpaceX in 2021 — not because we bought the Amazon IPO story on paper, but because we saw a force becoming the dominant platform in its domain.

We did not invest in Amazon. We invested in SpaceX because we saw it becoming a monopoly — or at least the dominant force — in space tech. DeepSeek is the same shape inside China: if AI is about scale and data, and China finishes in the global top two, Beijing's chosen national champion is a long-term winner. Call it the Aramco of AI for China.

Monopolies and national champions do not price on historical multiples. SpaceX did not. You cannot value China's AI national champion with a SaaS ruler.

高盛王亚军 Wang Yajun, Goldman Sachs' Asia equity capital markets head, put one industry view in one line: "最终大模型领域将实现赢者通吃" — "the model sector will end winner-take-all." That is a sector call, not proof DeepSeek is the monopoly.

The bet is simpler than the multiple. First, China finishes number one or number two in AI globally — we believe it must. Second, if AI runs on scale and data, DeepSeek is among the national-champion names that get listed to prove it. If both hold, price-to-sales is trivia. Chinese capital-market commentary is already modeling a post-listing market cap of 1.5万亿 to 2.5万亿元, three to five times the private round (未尽研究, 10 Sep) — commentary, not a forecast we underwrite. It has a decade of room to run.

The bull case inside China is the same bet. 投行圈子, a Chinese investment-banking commentary followed by the deal community: "投资者抢的不是当期利润，而是'中国AI冠军'这张门票" — "Investors are not buying current profits. They are buying the ticket to China's AI champion."

So the exposure question is the point. If DeepSeek itself is unreachable, the listed proxies are the tape. The mirror risk sits in Western businesses whose margins assume the cost of intelligence stays American — API spreads, compute rents, frontier-lab marks at 30 times revenue. Those compress first.

And if you do not believe China can win, this is not your thesis. Walk away. Just do not pretend the multiple was the reason.

If this sharpened how you read the AI map, forward it to one person who allocates capital.

Subscribe to East/West AIpha

This newsletter is normally $38/month. For the next 88 subscribers, it's free for a year.

Book 2 — Christmas

San Eng

Chairman, WorkOptional

WSJ/USA Today bestselling author

East/West AIpha is for informational and educational purposes only. San reads everything before it goes out. Nothing here is financial, investment, legal, or tax advice, and nothing is a recommendation to buy or sell any security. “We are long” states the author’s investment thesis, not a recommendation. The author and his affiliates may hold direct or indirect positions in companies or assets mentioned, including through funds, special-purpose vehicles, or listed proxies. Private-market figures cited are reported and unaudited, not verified. Investing involves risk, including total loss. Do your own research and consult a licensed professional before making investment decisions. Past performance is not indicative of future results.

#WorkOptional #EastWestAIpha #ikiBrain #10Cs #AI #FutureOfWork

Produced by iki, San's AI Chief of Staff, and a team of AI agents under San's direction.