According to LatePost, Moonshot AI has confidentially submitted its A1 filing to the Hong Kong Stock Exchange, formally kicking off its IPO process. The A1 is one of the formal application forms for companies seeking a HKEX listing.
At nearly the same time, it is pushing ahead with a new funding round at a $50 billion pre-money valuation. This is likely the last private round before Kimi goes public.
Taken together, these two moves mean this is not simply another large-model company heading for an IPO.
2026 is likely the final IPO window in which independent large-model companies can still be priced on tech dreams.
Let's recap what happened this year.
On Jan. 8, Zhipu listed on the Hong Kong Stock Exchange. The next day, MiniMax went public. On June 1, Anthropic confidentially submitted a draft S-1 to the SEC, with the latest plan to list as soon as late September to early October. On June 8, OpenAI also confidentially filed IPO documents, though it later leaned toward delaying its listing to 2027. On June 12, SpaceX officially debuted on Nasdaq. Now it's Kimi's turn.
The private market can still price Kimi based on what foundation models might become. But once the listing documents are out, the price enters a different vocabulary. Those deciding its worth shift from a few long-term private investors to institutions and retail investors in the public market.
The $50 billion valuation in its final pre-IPO round may be the last time Kimi is priced mainly by people who believe in the future.
Here's how prices were driven up over the past three years.
From 2023 to 2026, each round of price increases by independent foundation-model companies reflected a shift in what buyers were actually purchasing.
At first, they were buying a bet that large models would become the next computing platform.
Back then, many companies didn't even have a real product. Valuations depended mainly on the team, the model's direction, and who might get there first.
Moonshot AI was founded in April 2023. Two months later, it raised more than $200 million in seed funding at a $300 million post-money valuation.
Later, the market started paying for model capability.
Whoever was closest to the frontier and climbing leaderboards fastest commanded the highest price. In February 2024, Alibaba led Moonshot AI's Series A+ round, raising over $1 billion at a valuation of about $2.5 billion. In August of that year, a Series B brought the valuation to $3.3 billion.
Then came users, APIs, and agents. As capabilities got called and paid for, products finally developed something measurable.
In the past year, an even bigger assumption has been layered into valuations:
Today's large-model companies will become the platform companies of the AI era.
Kimi's fundraising curve this year is almost the shape of that assumption being marked up repeatedly.
Between January and February, three consecutive rounds took the valuation from $10 billion to $18 billion. In May, it raised about $2 billion more, for a post-money valuation above $20 billion. A new round began in June at a pre-money valuation of $31.5 billion. On July 29, the Series F closed, raising more than $3.5 billion at a post-money valuation of about $35 billion. A week later, Series G launched at a pre-money valuation of $50 billion.
Across the ocean, the prices are even more striking.
Anthropic completed a roughly $30 billion Series G in February at a post-money valuation of $380 billion. On May 28, it raised another $65 billion in a Series H, lifting its post-money valuation to $965 billion — surpassing OpenAI's then-valuation of about $852 billion for the first time.
The most distinctive feature of this pricing is that it doesn't require today's profits to prove anything, or even require today's revenue to match today's price. As long as investors believe few foundation-model companies will be left standing, they can bet on the endgame in advance.
For the past three years, large-model company valuations have run ahead of business models. What capital is really buying is the 'what if it wins' scenario.
The question used to be who would win. Now it's becoming: once you win, what business are you actually in?
Models are no longer a scarce commodity.
Whether that $50 billion can hold up ultimately hinges on a basic question:
What exactly does Kimi have that rivals can't quickly match?
For years, the high valuations of large models rested on an implicit assumption: model capability itself is an asset.
With few able to train frontier models, clear capability gaps, and high training barriers, owning a strong model alone can constitute much of a company's value.
By 2026, gaps between leading models are increasingly hard to sustain, and ordinary users see less and less difference between first and third place.
Open-source models are closing in on closed-source frontiers faster. At K3's release, an open-source model beat closed-source models on Code Arena for the first time. Similar milestones have become increasingly common over the past year.
The lifespan of each model generation is getting shorter.
A model that leads tech news today may be left with only a half-generation advantage in a few months. Meanwhile, inference prices keep falling. Agents have further changed the relationship between users and models. Users ultimately care about whether the task is completed, not necessarily which model is running behind the scenes.
But that doesn't mean models have become worthless. Zhipu raised API prices by 83% in the first quarter of this year, yet call volume grew 400%. Some research reports interpret this as the industry's pricing shifting from traffic consumption to monetizing computing power.
Raising prices without driving away customers is a strong signal in any mature industry. So the more precise change is that scarcity is shifting.
The scarcest thing used to be 'I have a strong model.' Now model capability hasn't lost value, but it no longer automatically translates into company value. Owning a frontier model is becoming the starting point of valuation, not the conclusion.
Chinese peers are starting to diverge, and the divergence itself is good news.
Let's look at what each of China's leading companies is relying on to sustain a long-term war.
Kimi is a model, a proprietary product, and an agent, relying on independent funding to maintain frontier research and development.
MiniMax focuses more on multimodal capabilities and overseas revenue, with a broader product matrix. It is also an independent company that must solve its own funding problems.
Zhipu takes a different path. With B2B, government-enterprise, and MaaS, its business structure is closer to a model platform and enterprise AI.
Looking further, ByteDance and Alibaba are a completely different class of players. ByteDance has Doubao and Seed, backed by a massive advertising and content cash machine. Alibaba's Qwen can sustain long-term investment, open-sourcing, and price cuts, because the models ultimately return value to Alibaba Cloud and the broader ecosystem.
For independent model companies, the model war is a matter of survival. For big tech, models can remain a cost center for a long time.
Big tech can even accept models not being profitable, as long as they help defend larger businesses like cloud, advertising, e-commerce, content, and office software.
Independent companies lack this buffer. If they cannot quickly generate strong enough cash flow, their options are actually limited.
Keep raising capital, find a super giant, or enter the public markets.
Kimi is now taking the third path. Two companies have already gone ahead of it this year.
Zhipu listed on Jan. 8 at an offer price of HK$116.20, rising 13% on debut. MiniMax listed on Jan. 9 at HK$165, surging 109% on its first day. By June 8, both companies had been added to the Hang Seng Tech Index.
Both companies are solving the same problem: how to translate burned R&D spending into a revenue trajectory that public markets will believe.
Zhipu's answer is closer to selling capabilities to businesses that can do the math, while MiniMax is closer to scaling up products, users, and overseas markets first.
Rewind to 2025, and the latter answer almost perfectly matches the entire experience of China's internet over the past two decades.
In the mobile internet era, users themselves were assets. As user bases grew, customer acquisition and infrastructure costs could be spread out. Each new user often meant future revenue from ads, subscriptions, transactions, and more.
Large language models are different. Every additional user opening a chat window adds computing load in the background. Scale does bring efficiency gains, but GPU electricity costs and depreciation won't be diluted by more users.
The previous generation of internet companies treated users as currency. This generation of model companies has suddenly discovered that the exchange rate for that currency has changed.
The market has begun to distinguish the quality of revenue, and even after doing so, it remains willing to pay high prices for the industry.
Zhipu's MaaS API annual recurring revenue reached about 1.7 billion yuan by March this year, yet the market assigns it a valuation of hundreds of billions of Hong Kong dollars.
No one is really valuing the company on 1.7 billion yuan in revenue. The market is still paying for the future, but now it's starting to demand that the future at least take shape.
The regime shift has happened, but it is far from complete.
This is the moment Kimi most wants to catch up to.
Capital markets pricing the future in advance is nothing new.
The first comparison is the internet in the 1990s.
When Amazon went public in 1997, it was certainly not yet the giant spanning e-commerce, cloud computing, and logistics infrastructure that it would later become.
The market was buying a future: the internet would transform commerce, online transactions would expand, and the digital world would eventually produce several super-companies.
Many internet companies entered public markets while their business models were still unproven. Capital markets gave them time to establish a foothold before demonstrating how they would make money.
AI today bears clear similarities to that era, but with one notable difference. Many internet businesses saw unit service costs fall as they scaled. AI still carries heavy real computing costs.
More users often means more chips, larger data centers, and higher electricity consumption. So public markets will inevitably focus more on growth quality than they did for the previous generation of internet companies.
A closer comparison is biotech. Many biotech companies went public before their lead drugs had even received approval. Markets were willing to price research capabilities, pipelines, and probability of success. The Hong Kong Exchange's Chapter 18A, launched in 2018, did something similar, allowing companies without commercialized products to access public market funding early.
Foundation model companies and biotech firms share a common trait: they burn through R&D spending quickly, with highly uncertain outcomes. Today's most important assets may still be a group of engineers, a training methodology, and a generation of technology not yet fully commercialized.
There is an even more recent example. Around 2021, the previous generation of AI companies—SenseTime, Megvii, CloudWalk, and YITU—also caught a window when markets were willing to pay for the future.
SenseTime's market value once exceeded HK$300 billion after its listing, as markets similarly believed AI would become infrastructure.
But that generation's problem was that revenue remained tied to a project-based model. Contracts were signed one by one, deliveries made project by project, with staff following the work. Gross margins were often eroded by customization and labor costs.
This generation of foundation-model companies is indeed different. API, subscription and token-based billing models are inherently closer to recurring revenue, creating a new valuation language for large-model companies.
Once that language truly matures, the market will no longer be as forgiving as it is today.
$50 billion
Now let's revisit that $50 billion. There's no need to rush to calculate the price-to-sales ratio, or to immediately judge whether it's expensive or cheap.
First, consider what this price demands:
Kimi's model capabilities must remain in the top tier over the long term.
The product must gradually evolve from a high-frequency AI tool into a stable entry point.
Agents must drive new payments and retention.
Enterprise and API revenue must continue to scale.
Inference costs must continue to decline.
Revenue growth must outpace compute consumption and R&D investment over the long term.
Most importantly, today's model capabilities must not depreciate significantly within six months due to open sourcing, price wars, and technological iteration.
Kimi has already supplied part of the answer.
Annual recurring revenue crossed $100 million for the first time in March, hit $200 million in May, and topped $300 million in June.
API accounts for roughly 70% of enterprise revenue. The Kimi Hosted Agent platform is also being deployed within enterprises, aiming to integrate the model into workflows including office, R&D, and investment research.
Following K3's release, Huang Zhenxin, Moonshot AI's enterprise business head, said market demand for K3 has far exceeded expectations, straining compute resources — a 'happy problem.'
That is good news in a tech team's weekly report. In an earnings release, there is a cost column to watch beside it.
That is the clearest line between two pricing regimes.
In the private market, if Kimi ultimately becomes the next-generation AI platform, $50 billion may prove cheap.
The public market will ask how much more capital and how long it will take to become a company truly worth $50 billion.
Epilogue.
Capital markets are most generous not when a technology has already been proven.
It is the years when everyone is certain it will change the world but no one knows who will end up making money.
In those years, all company boundaries are blurred. Amazon could be imagined as the future commercial infrastructure. A biotech could hold a multi-billion-dollar market cap before its drug ever reaches the market.
Since 2023, large-model companies have enjoyed the same grace period.
Model capability, user growth, and agent potential can all be converted directly into valuation. If the future is big enough, revenue that hasn't yet occurred can be priced in ahead of time.
The catch is that every technology wave reaches the same juncture.
Eventually, the market stops asking how important a technology is and starts asking how much of the spoils each company will capture.
The internet settled on advertising, e-commerce, subscriptions, and cloud. Biotech built clinical-stage and pipeline valuations. The previous AI generation was ultimately repriced on project revenue, gross margins, and cash collection cycles.
Large models are now approaching their own reckoning.
Zhipu and MiniMax have already presented their first earnings to public markets. Anthropic and OpenAI are at the threshold. Kimi has arrived at the door with a $50 billion private-market offer.
Before long, the term 'foundation-model company' may itself lose its valuation significance.
Some will be valued as enterprise software companies, others as consumer internet gateways, some as cloud infrastructure, and others will be left with only an expensive R&D team.
At that point, the market will no longer assign the same valuation to all of them simply because they all train large models.
What is really disappearing is the era when 'if you could become the future, you could be priced as the future upfront.'
The A1 that Kimi is offering right now is a bid to seize that window.
Once public markets truly learn to price models, agents, tokens, inference costs, and revenue quality item by item, dreams will still be worth something.
But after that, there will be a calculator sitting next to the dream.