Moonshot AI, the Beijing lab behind Kimi, is closing a pre-IPO round on 27 August at about $50 billion. That is not a listing yet. It is the last private mark before a Hong Kong filing that the company has told investors it wants in by 30 September, with shares possibly trading in late 2026 or early 2027. The number that matters for anyone who actually uses the product is simpler: Moonshot did about $300 million of annual recurring revenue in June. MiniMax listed in Hong Kong in January on similar ARR and opened around $9 billion. Anthropic last raised at $965 billion on roughly $47 billion of run-rate revenue. Same industry. Three completely different prices for a dollar of sales.
Key takeaways
- $50 billion on $300 million ARR is about 167× sales. Anthropic is ~20×. MiniMax listed at ~30×. DeepSeek is chasing ~$71 billion on $400–500 million of ARR. Moonshot is pricing itself as the next DeepSeek, not the next MiniMax.
- Two LLM companies are already public — Z.ai (Zhipu) and MiniMax, both Hong Kong, January 2026. Anthropic has a confidential US filing and is aiming at autumn. OpenAI filed a week later and is leaning 2027.
- Users and revenue have decoupled. ChatGPT has ~800 million weekly users. Claude has perhaps 30–50 million consumers and more revenue than OpenAI. Kimi's consumer app peaked near 36 million and then shrank; Moonshot now says "tens of millions of professional users" and that API is ~70% of revenue.
- On the product you actually buy, Kimi vs Claude has not changed: subscriptions still sit at ~$19 vs ~$20. On the API, Anthropic made Sonnet 5's $2/$10 rate permanent on 10 August, so Kimi K3 at $3/$15 is no longer the cheaper flagship.
What Moonshot is selling this week
The round is a Series G, pre-money about $50 billion, closing 27 August according to IPO Zaozhidao as reported by KrASIA. The previous private mark was roughly $30–35 billion. The jump is Kimi K3, the 2.8-trillion-parameter open-weight model Moonshot released on 16–17 July. Independent leaderboards put K3 first on Arena's frontend coding arena, ahead of Claude Fable 5. President Zhang Yutong said enterprise ARR had grown "severalfold" after the launch, without giving a new absolute number.
Revenue is already on a steep slope that we have been tracking on the site: $100 million ARR in March, $200 million in May, $300 million in June. About 70% of that is API, not the consumer subscription. That is the same shape as Z.ai's GLM coding business and the opposite of MiniMax, which listed on companion apps and video. Moonshot is asking public-market money to believe K3 turns a developer API into a frontier lab. The 2026 API price war is the backdrop: Chinese labs win volume; Western labs still win the invoice.
A note on sourcing, because this category invents numbers for sport. Moonshot denied an IFR/Reuters report that it would file in August and raise $3 billion. The $50 billion figure is what investors in the current round have been told, not a priced IPO. Treat it like any other private mark — real until the next one, and not a quote you can trade.
Every LLM listing, side by side
There are now two kinds of "LLM IPO". The ones that already happened, in Hong Kong, at single-digit billions. And the ones that have not happened yet, which bankers are discussing in the hundreds of billions or the trillions. Moonshot is trying to jump from the first group into the second without waiting to list.
| Lab | Status (Aug 2026) | Last mark | ARR / run-rate | Multiple | Venue |
|---|---|---|---|---|---|
| Anthropic (Claude) | Confidential S-1, 1 Jun | $965B (May Series H) | ~$47B (May); later briefings higher | ~20× | Nasdaq, targeting autumn 2026 |
| OpenAI (ChatGPT) | Confidential S-1, 8 Jun | ~$852B | ~$25B early 2026 | ~34× | US, leaning 2027, $1T floor |
| Moonshot (Kimi) | Pre-IPO close 27 Aug | $50B pre-money | $300M (June) | ~167× | HKEX, file by 30 Sep |
| DeepSeek | Private; first outside round June | ~$52B done; talks at ~$71B | $400–500M | ~140× at $71B | Shanghai STAR, 2027 |
| Z.ai (Zhipu / GLM) | Listed 8 Jan 2026 | IPO ~$4–9B; peaked ~$128B in June | $1B by July (from ~$100M in Jan) | Wild — tiny free float | HKEX 2513 |
| MiniMax (Hailuo, Talkie) | Listed 9 Jan 2026 | ~$9.2B day one | ~$300M by May; targeting $1B | ~30× at listing | HKEX 0100 |
Figures are reported private marks, company briefings and contemporaneous press as of 21 August 2026. ARR is not audited revenue. Z.ai's market cap is retail-driven and moves tens of percent in a session — do not treat the June peak as a stable comparable. We mapped the wider private-market zoo in May in the valuation race; several of those marks have already doubled.
Read the table left to right and one pattern is loud. The US labs are selling a cash-flow story that already exists. Anthropic's last round was twenty times a number that is already in the tens of billions. OpenAI is more expensive on a multiple, has more consumers, burns more cash, and is in no hurry. The Chinese labs that already listed sold a much smaller story — MiniMax opened at $9 billion with the same $300 million ARR Moonshot now has. Then Z.ai's stock went feral, GLM coding revenue went to $1 billion, and the next cohort (DeepSeek, Moonshot) decided the January prices were a rounding error.
Moonshot is not comparing itself to MiniMax. It is comparing itself to DeepSeek: open weights, developer API, a model that scared Western labs for a week, a three-digit sales multiple. If K3 keeps converting, $50 billion is the ticket into that race. If it does not, Hong Kong will remember that MiniMax listed the same revenue for a fifth of the price.
Who actually uses each model
User counts in this industry are a mess on purpose. Vendors mix weekly and monthly, count a WhatsApp embed as a "user", and publish the number that makes the last round look cheap. The useful split is not the headline MAU. It is who shows up, from where, and whether they pay.
| Product | Users (best sourced) | Who they are | How the lab gets paid |
|---|---|---|---|
| ChatGPT | 800M+ weekly (OpenAI); ~1B monthly app users (Sensor Tower, May) | The default, everywhere ChatGPT is legal. Students, offices, phone-in-line users. | Plus/Pro subs + API + enterprise. Lots of free users; the paid minority funds the lab. |
| Gemini | ~400–750M depending on whether Workspace/Android is counted | People already inside Google. Many never chose an AI; it arrived in Search or Gmail. | Bundled. Hard to separate from Alphabet. |
| Doubao | 260M+ monthly — China's most-used AI app | Chinese consumers on ByteDance rails. The ChatGPT of a market ChatGPT cannot enter normally. | Cheap API on Volcengine; consumer is a distribution play, not a $20/month ladder. |
| Claude | ~30–50M consumer monthly (third party; Anthropic does not publish a MAU) | Developers, knowledge workers, teams. Small consumer base, heavy users. | API + Claude Code + Cowork + Team/Enterprise. This is why $47B ARR sits on ~40M people. |
| Kimi | Company: "tens of millions of professional users" a month. Consumer app peaked ~36M in 2024, then fell. Android ~873K MAU in August. | China professionals who wanted long context; now developers on the API. Not the mass-market chatbot it was in 2024. | API ~70% of revenue. Consumer ladder still priced like Claude: $19 to $199. |
| DeepSeek | ~30–80M depending on tracker; 81.6M weekly in China (Reuters, Feb) | Coders and bargain-hunters, China-heavy, a real global web-share tail. | Almost all API. Cheapest frontier-class tokens we track. |
| Grok | 40–117M — the high number is X-integrated (SpaceX IPO disclosure) | People already on X, plus a smaller standalone-app group. | Bundled into SuperGrok / X Premium+ / Cursor. Hard to know what "a Grok user" paid. |
| GLM / MiniMax | Z.ai: ~15M consumer, 8,000+ institutions, ~100k overseas API users. MiniMax: 200M+ cumulative across Talkie and Hailuo. | GLM is Claude-shaped (devs, enterprises). MiniMax is entertainment and video, not a daily coding assistant. | Z.ai just printed $1B ARR on coding. MiniMax listed on consumer apps. |
Two sentences fall out of that table, and they are the whole valuation argument.
ChatGPT has twenty times Claude's consumers and less revenue. Anthropic's users are fewer, denser and on invoices. That is the business Moonshot is trying to copy with K3: stop counting chat MAU, start counting API tokens. Doubao still owns China the way ChatGPT owns the rest of the world. Kimi lost that fight. The 76% drop from 36 million monthly users was not a rounding error — it was Doubao and DeepSeek taking the consumer slot, which is why Moonshot now talks about "professional users" and why API is 70% of the bill.
Kimi's remaining footprint looks like Claude's, not ChatGPT's — at 1/150th of the revenue. That gap is the $50 billion question. Either K3 (and Kimi Work, and the agent swarm we walked through in the plan-by-plan comparison) pulls Kimi onto the Anthropic path, or Moonshot is a $300 million API shop wearing a DeepSeek multiple. Z.ai already proved a Chinese coding model can hit $1 billion ARR in five months. That is the precedent Moonshot's bankers will put on slide two. The precedent on slide three is MiniMax, which listed the same ARR for $9 billion.
What this changes if you already pay for one of them
Almost nothing about the product. A Hong Kong ticker does not give you more agent credits on Vivace or a larger Claude Max pool. It does change three practical things.
Price. We have been saying Kimi K3 and Claude Sonnet 5 were a tie at $3/$15 per million tokens. That stopped being true on 10 August, when Anthropic made Sonnet 5's introductory $2/$10 rate permanent. Put the same workload through the token calculator: the Western mid-flagship is now cheaper than K3, while Kimi's mid-tier (K2.5 at $0.60/$3, K2.6 at $0.95/$4) still undercuts every Claude model. The IPO story needs Kimi to charge a premium. The price list is moving the other way. Full Chinese ladder is in Chinese AI models pricing.
Jurisdiction. A Hong Kong listing makes Moonshot easier for international funds to hold. It does not move your prompts out of Singapore, and it does not change Chinese content filters. India is the second-largest country on this site, almost entirely on Kimi and the Chinese-model cluster. If Washington or Beijing tightens access to Chinese weights — both have been talking about it — those users feel it first. The safety answer is still the one in are Chinese AI models safe: the weights are just software; the hosted API is the risk. Claude still does not train on paid conversations by default. Kimi still does.
The second-opinion seat. Kimi is already inside Cursor and GitHub Copilot's model pickers, which is how a lot of Western developers meet it without a Moonshot account. That distribution is worth more to a $50 billion story than another 10 million consumer MAU in China. It is also why a listing does not trap you: if Kimi gets worse, or dearer, or geo-fenced, you switch the dropdown. We mapped those plans in AI plans that let you choose the model.
Moonshot is asking to be priced like DeepSeek and used like Claude. The users it still has look like Claude's. The revenue does not. That gap is the entire $50 billion.
If you are choosing this week: keep Claude as the native seat if you care about files, Cowork and not training the lab on your repo. Keep Kimi (or DeepSeek, or GLM) as the cheap API or the second opinion. Do not pay $199 for Vivace because a banker wrote $50 billion on a term sheet. That number is for the cap table. Yours is still $19 or $20 a month, and the wizard will send you to the same place it did last month.
Frequently asked questions
When is Moonshot AI's IPO?
Moonshot is closing a pre-IPO round on 27 August 2026 at about $50 billion pre-money. That is not the listing. The company has told investors it aims to file with the Hong Kong Stock Exchange by 30 September, with a listing expected in late 2026 or the first quarter of 2027. Moonshot has already denied one Reuters/IFR report that it would file in August and raise $3 billion, so treat any date as a plan, not a booking.
How does Moonshot's $50 billion valuation compare to Anthropic and OpenAI?
It is a different sport. Anthropic last raised at $965 billion with roughly $47 billion of run-rate revenue as of May 2026 — about 20 times sales — and has confidentially filed for a US IPO targeting autumn 2026. OpenAI last marked at about $852 billion, with a larger consumer base and a 2027 listing bias. Moonshot is asking $50 billion on $300 million of June ARR, about 167 times sales. The $50 billion tag is huge for a three-year-old Chinese lab and small next to the two US frontier names.
Which LLM companies are already public?
Two pure-play Chinese labs listed in Hong Kong in January 2026: Zhipu AI (Z.ai) on 8 January and MiniMax on 9 January. MiniMax opened around $9 billion of market cap on similar ARR to Moonshot's current $300 million. Z.ai later re-rated violently — its market cap peaked near HK$1 trillion ($128 billion) in June — and reported $1 billion of ARR by July. DeepSeek is still private, targeting Shanghai's STAR Market in 2027. Anthropic and OpenAI have confidential US filings; neither has priced yet.
How many people use Kimi compared with Claude and ChatGPT?
ChatGPT is the default: OpenAI has reported 800 million-plus weekly active users, and app trackers put monthly users near a billion. Claude's consumer base is much smaller — third-party estimates cluster around 30 to 50 million monthly users — but Anthropic's revenue is larger than OpenAI's because enterprises pay. Moonshot says Kimi has tens of millions of professional users a month; the consumer app peaked near 36 million monthly users in 2024 and then shrank as Doubao and DeepSeek took the China chat market. Kimi now looks more like Claude (fewer users, more API) than like ChatGPT.
Does Moonshot's IPO change whether I should pick Kimi or Claude?
No. A listing does not make the model better at coding or safer with your files. On subscriptions they already sit in the same band — Kimi Moderato at $19, Claude Pro at $20. On the API, Kimi K3 is still $3/$15 per million tokens while Anthropic made Sonnet 5's $2/$10 rate permanent on 10 August, so the flagship-for-flagship price edge has flipped. Pick Claude for privacy, Cowork and the top end (Opus 5). Pick Kimi for cheap mid-tier tokens and agent swarms. The IPO is a story about Moonshot's owners, not about your next prompt.