Six months ago, Kimi founder Yang Zhilin said something.
He said that on the dark side of the month, there is sufficient cash in the account and no sense of urgency to go public in the short term.
What was the result?
On June 30th, the Science and Technology Innovation Board Daily exclusively reported that Kimi’s previous round of financing had just been completed, and a new round had already begun. The valuation after the previous round of investment was $20 billion, and before the new round of investment, it was $31.5 billion.
Calculate, this is the fourth round in six months. Over the past six months, we have accumulated more than 3.9 billion US dollars in financing.
You should know that six months ago, Kimi’s post investment valuation was only $4.3 billion.
In today’s Chinese primary market, it has been assumed that Kimi should continue to raise funds, and OpenAI and Anthropic should also continue to raise funds.
This matter itself is worth discussing more than a valuation of $31.5 billion.
1、 The big model has become a heavy asset business
The logic of Internet company financing is simple: exchange money for market share. It’s okay to lose today, we’ve reached a turning point and the return index has returned. Over the past twenty years, VCs have all followed this logic.
Big model companies are an exception, as AI companies are essentially capital intensive businesses.
For every generation of models pushed forward, thousands of high-end GPUs, billions of dollars in computing power costs, hundreds of researchers’ salaries, and a complete set of data cleaning and post training pipelines need to be prepared in the account.
Take OpenAI as an example, its annual revenue for 2025 is $13.07 billion. Listen a lot, right?
But the total cost and expenses are 34 billion US dollars. Just developing one project cost 19.18 billion US dollars, which is more than the annual revenue. The computing power bill paid to Microsoft is $10.59 billion. In addition, the inference cost is 7.5 billion US dollars, and the sales and marketing expenses are 5.73 billion US dollars. The annual operating loss was 20.92 billion US dollars.
By the first quarter of 2026, data indicates that losses are continuing to expand. Revenue of 5.7 billion US dollars, cash expenditure of 3.7 billion US dollars, burned off more than half of the revenue, and R&D expenses of 8.6 billion US dollars.
OpenAI predicts that its annual cash consumption will be $25 billion in 2026 and $57 billion in 2027.
If OpenAI proves why AI companies burn money, then Meta and Google prove that the entire industry is entering a capital expenditure race.
Meta’s annual capital expenditure guidance for 2025 is between $70 billion and $72 billion. By 2026, it will be directly raised to $125 billion to $145 billion, almost doubling. These increments are all deposited on heavy assets such as AI servers, self-developed acceleration cards, and global data centers.
The same goes for Google. The capital expenditure guidance for 2025 is $75 billion to $78 billion, which was raised to $180 billion to $190 billion in the latest communication in June 2026. The CFO’s original statement is that there will be significant improvements in 2027.
This money is not spent and gone, most of it is converted into fixed assets with computing power. The AI industry does not consume one-time cash flow, but continuously expanding computing resources. The money you invest directly corresponds to the hardware and racks in the computer room.
Internet companies address information distribution. The product is ready, and the marginal cost of serving one more user is approaching zero. The scale expands, and the increment is almost entirely profit.
AI companies are solving the problem of continuous computing power supply. Once the infrastructure stops investing in additional computing power, model iteration and service carrying capacity begin to decline on the same day.
Kimi raised $3.9 billion in six months, seemingly breaking the conventional rhythm of the primary market. But from a financial perspective, this money needs to cover team salaries, data center electricity bills, and high-end GPU procurement, to support the next round of model development and commercialization.
These investments will eventually become visible assets. What you are buying are cards, computer rooms, and one kilowatt hour of electricity.
2、 A broken chain, very short
Having sufficient cash in the account does not mean it can be stopped.
In the big model industry, every link in the entire chain is burning money, including base model training, online reasoning services, and agent ecosystem building. Any reduction in investment in any link will cause the growth curve to turn downward directly.
The cruelty of this industry lies in the fact that competitors invest billions of dollars a year in iterative base models, while you only invest a few hundred million. The gap is not a matter of iteration speed, but directly forms an intergenerational divide.
The growth of large model capability is non-linear. Only when the computing power exceeds a certain threshold will new capabilities emerge. Your computing power hasn’t reached that point, and your model and top manufacturers are not on the same competitive curve at all.
In this industry, technological gaps are almost immediately transmitted to business outcomes.
Insufficient computing power → Model lagging behind by one generation → Developer migration to competitors → End user loss → Revenue interruption → Narrowing financing channels.
If any link in the chain goes wrong, the entire business operation will stall.
Kimi’s continuous financing is not for pursuing any high-speed expansion. The core is one thing, to hold the bottom line of technological iteration and not be left behind by industry agents.
The top AI companies worldwide are the same. OpenAI and Anthropic both maintain high-frequency and large-scale financing. Kimi only fully exposed the survival logic of heavy capital in the Chinese market.
3、 $300 million ARR is the key
The truly interesting figure in this round of financing is not the valuation of 31.5 billion.
It’s ARR, annual recurring income.
In March 2026, Kimi ARR exceeded $100 million. In May, it exceeded 200 million US dollars. In mid June, it officially exceeded 300 million US dollars.
It tripled in three months. More than 70% of them are called by API developers.
The valuation of $31.5 billion is the capital market’s pricing of future imagination. The $300 million ARR is a proven and verifiable real gold and silver.
The valuation of the domestic big model track in the past three years has been supported by industry stories. From Zhipu to MiniMax to the Dark Side of the Moon, the universal narrative consists of three sentences: Chinese version of GPT, short-term catching up with OpenAI, and Chinese Anthropic.
This round, Kimi did something different by directly presenting quantifiable and continuously growing real financial revenue to the primary market: $300 million ARR, 70% contribution from API, 400% year-on-year growth in overseas paying users, and product landing in over 200 countries.
This is the first time that the domestic AI industry has used real commercial revenue to support valuation, rather than relying on technical stories. This means that the big model industry is entering a new stage of revenue verification technology.
The impact of this change is much greater than the valuation of $31.5 billion itself.
Starting from mid-2026, domestic large model enterprises will be completely divided into two categories. One type is still competing for model scores and technical parameters. The other type already has stable and sustained payment income.
Companies that generate revenue are also undergoing internal differentiation:
Zhipu will be listed on the Hong Kong Stock Exchange in January 2026 and will simultaneously promote the Science and Technology Innovation Board A+H. In 2025, the annual revenue will be 724 million yuan, with a gross profit margin of 41%, mainly focusing on the localization deployment of government and enterprise MaaS.
MiniMax， Following the landing of Zhipu on the Hong Kong Stock Exchange. In the first three quarters of 2025, the revenue was 53.43 million US dollars, with overseas revenue accounting for 73%, driven by C-end subscription products Talkie and Conch AI.
Step up Star, Pre IPO round led by operator industry capital, with the goal of submitting Hong Kong stock returns within 2026. The annual revenue for 2025 is approximately 500 million.
Zero One Thing voluntarily withdrew from the universal base competition and in June, jointly established Wanfeng Intelligence with Zhengda Group to engage in egg chicken farming.
Baichuan Intelligence shrinks its general model and focuses all resources and funds on the medical vertical field.
The former “AI Six Tigers” are now completely different in terms of race track, market, and commercialization path, and are no longer in the same competitive group.
Kimi has taken a path that no one in China has fully implemented before, with API priority, developer ecosystem driven, and overseas paying users driving revenue. This route is basically based on Anthropic’s script.
Anthropic’s valuation after H-round investment in May 2026 is $965 billion, and its revenue growth curve is highly informative: ARR at the end of 2024 was only $1 billion, rising to $14 billion in February 2026, and officially disclosed an annualized operating revenue of over $47 billion at the end of May; From the end of 2024 to May 2026, the annualized revenue scale will expand to 47 times its original size.
If Kimi can run this overseas API route, the pre investment valuation of $31.5 billion is just the starting point. If the commercialization growth is less than expected, this valuation is the periodic peak of the primary market foam.
4、 The money for AI was exchanged for a group of people
The traditional primary market has a perception that the higher the valuation of a company, the more difficult it is to raise funds, and the fewer institutions that can accept large investments.
The AI track is completely reversed. The higher the valuation of top companies, the easier it is to obtain large long-term funds.
The core change is that the people who pay have changed.
The proportion of traditional VC is shrinking. Industrial capital, state-owned industrial funds, and global sovereign funds have become the main force.
DeepSeek recently completed its first round of external financing, exceeding RMB 50 billion, with a post investment valuation of approximately RMB 400 billion (nearly USD 59 billion). The main investors are national AI industry funds and leading physical industry groups.
Kimi’s list of investors for the past six months includes Meituan Dragon Ball, China Mobile, CPE Yuanfeng, Singapore Temasek, and Abu Dhabi MGX. In Anthropic’s G and H rounds, GIC, MGX, Qatar Investment Authority, and others have also appeared.
These long-term institutions are not looking for short-term returns of 3 to 5 times. What they want is a ticket to the next generation of digital infrastructure.
The logic behind Tencent’s investment in Kimi is different from its earlier investments in JD.com, Meituan, and Pinduoduo. In the early years, it was driven by financial returns, but now the layout of the top models is a strategic bottleneck for the underlying industrial chain.
For national capital, AI has already been placed on the same level as electricity, railways, and 5G, serving as the infrastructure for the next generation of the national economy. Having a globally competitive base model in hand is equivalent to mastering the discourse power of computing power distribution and data flow in the future digital economy.
So now AI financing is no longer just a simple financial investment, but a long-term strategic resource allocation for major countries and industrial groups.
Enterprise valuation begins to reflect both industry value and strategic value simultaneously. Industry weight and geopolitical weight are both included in pricing.
$31.5 billion is not the result of free market bidding. It is a quota jointly allocated by multiple industries and sovereign capital.
This also explains why Kimi is not in a hurry to go public.
Zhipu and MiniMax went public on the Hong Kong stock market in January this year, while Step Leap Star is rushing to submit its application. DeepSeek has secured a large amount of financing from the national team. Only the dark side of the moon remains in the primary market.
What is Kimi waiting for? Most likely, we are waiting for ARR to rise from $300 million to $3 billion, or even billions of dollars.
The money in the primary market can tolerate long-term non profitability and wait for compound income growth. The secondary market is not good, it values short-term indicators such as quarterly growth rate, market to sales ratio, and computing power liabilities. Once listed, the compound interest on future revenue growth will be discounted in advance by the market.
By staying in the primary market, Kimi can complete the commercial growth cycle and master the narrative rhythm on its own.
5、 Anti Internet Cost Curve
The core logic of making money on the Internet platform is that the marginal cost is close to zero.
The same set of products costs 10 yuan to serve the first user and 10 million users, with almost no increase in single user costs. The larger the scale, the wider the profit. In the past two decades, consumer Internet platforms have been rewarded by this model.
The cost curve of AI companies is completely opposite.
The more users and API calls, the higher the operating costs. The call volume has increased, but the losses have actually widened.
If we break down the costs of AI companies, it can be roughly written as follows:
Comprehensive computing power cost ≈ number of active users × amount of single inference adjustment × unit computing power price
If any one of the three items increases, the total cost will also increase accordingly. If all three items rise together, the cost is exponential expansion.
OpenAI previously disclosed that when ChatGPT was first launched, the daily inference cost was approximately $7 million. That’s still the small-scale user stage. The number of users and calls has already multiplied many times, and the actual daily cost will only be higher.
AI is not traditional software, it is more like a heavy asset utility that continuously consumes energy.
After the establishment of the computing power data center, the expansion of user call scale will not dilute fixed costs such as hardware depreciation, electricity bills, and bandwidth. The more people serve, the more computing resources are consumed.
Heavy assets, high growth, and sustained funding gaps are typical characteristics of top tier large model enterprises.
Before achieving steady-state profitability, OpenAI, Anthropic, and Kimi will maintain a high-frequency financing rhythm.
6、 Kimi is not that special
Looking back at Kimi’s core data, the pre investment valuation was $31.5 billion, ARR was $300 million, API revenue accounted for 70%, overseas paying users grew by 400%, and a new round of large-scale financing was launched.
These numbers do not describe a company that has just achieved profitability, but rather a heavy asset enterprise that, although its business is running smoothly, has simultaneously given rise to a larger computing power funding gap.
The higher the scale of API calls, the greater the inference hardware and power consumption. The faster the expansion of overseas markets, the heavier the capital investment in cross-border computing nodes and multi regional compliance. ARR often faces double fixed capital expenditures for every additional dollar of revenue.
The income curve is moving upwards, while the capital expenditure curve is steeper.
On a larger scale, this is not a problem for the Kimi family. OpenAI and Anthropic are the same.
They are all in the same business state, with faster income growth and capital expenditure growth.
The financing of top AI companies is not a temporary expansion tool, but a survival bottom line to maintain the iteration of core computing power.
The real core contradiction is not whether we can generate revenue or make profits, but whether we can continue to receive money and withstand this computing power arms race.
If it’s outside of the page:
3.9 billion US dollars were received, with a valuation of 31.5 billion US dollars hanging there. These two numbers do not represent Kimi’s safe landing, they are more like a countdown.
The higher the valuation, the higher the threshold for the next round. The pre investment of $31.5 billion means that at least someone is willing to take on the next round at a valuation of $40 billion or even higher. The institutions that can offer this price can be counted globally.
With each round of melting, Kimi is one step closer to the ceiling that cannot be melted. And the computing power bill will not be suspended just because you cannot raise money.
What Yang Zhilin is really racing against time is not technology, but pushing ARR to a number that can support a market value of billions of dollars before the money in the primary market runs out. Arrived, listing is coronation. Failing to make it to the market is a failure.
The time left for him may be shorter than the market thinks.
相关文章
