On October 6, a report by Bloomberg once again drew the AI community's attention to The Dark Side of the Moon.

Sources with knowledge of the matter disclosed that The Dark Side of the Moon has concluded its final round of funding, achieving a valuation of approximately $50 billion, and has plans to go public on the Hong Kong Stock Exchange in the first quarter of 2027. This news comes just a month after it was reported that the company had confidentially submitted its A1 listing application to the HKEX.

By the end of 2025, Yang Zhilin, the founder of The Dark Side of the Moon, had stated in an internal memo, "There's no hurry to go public in the short term; the future is in our hands."

However, the transition from "no hurry" to "quietly accelerating" took less than a year. Yet, when considering the global landscape of large model industries, The Dark Side of the Moon's change in stance is not hard to fathom.

In January of this year, Zhipu and MiniMax both listed on the HKEX; in October, DeepSeek was rumored to be finalizing a new round of financing, with plans to go public in early 2027; across the ocean, OpenAI and Anthropic have also entered the countdown to their listings.

Leading companies in the large model sector are flocking to the capital markets, and their "capital siphon effect" has become increasingly pronounced. The capital frenzy surrounding large models has also begun its countdown.

01 The Pressure to 'Deliver Results' Looms Large

Over the past few months, The Dark Side of the Moon has been making a series of strategic moves across models, products, and commercialization.

In July of this year, The Dark Side of the Moon unveiled K3, its most advanced large model to date. Within 30 minutes of its launch, it topped the Hugging Face Trending list, and its annualized recurring revenue (ARR) also surged multiple times, skyrocketing from $100 million in March to $1 billion in August. Bloomberg, citing sources familiar with the matter, reported that ARR is expected to double from August to reach $2 billion by the end of the year.

In September, The Dark Side of the Moon's activities intensified.

On September 10, it launched the "Moon Landing Plan" enterprise partnership program; on the 11th, K2.8 Preview went live; on the 17th, it released a financial industry solution; on the 18th, Kimi K3 made its debut on Amazon; and on the 21st, Kimi Code Desktop was officially released...

At the same time, The Dark Side of the Moon's valuation has been on a meteoric rise. From January to May this year, it completed four consecutive funding rounds, with its valuation soaring from $10 billion to over $20 billion; in June and July, it secured Series E and F funding, reaching a post-money valuation of $35 billion; according to market reports, after completing its latest Pre-IPO round, its valuation has reached approximately $50 billion.

These moves all seem to point to one thing: The Dark Side of the Moon is making its final push before the IPO. Perhaps it is not short on cash, but in the AI industry, where advancements occur weekly, there is no reason to remain idle.

Globally, The Dark Side of the Moon is not the only one accelerating its IPO plans.

DeepSeek, which had previously stated it would not seek external funding, completed its first round of financing exceeding $7 billion in June this year, setting a record for domestic AI single-round financing. In October, it was rumored to be raising at least RMB 80 billion in a new round to pave the way for its 2027 IPO.

In May this year, StepFun was reported to be finalizing nearly $2.5 billion in financing and had already dismantled its red chip structure in preparation for a full-scale sprint for a Hong Kong IPO; 01.AI is also advancing its Pre-IPO financing, planning to list on the HKEX in 2027.

Meanwhile, across the ocean, Anthropic and OpenAI have also submitted their listings, with foreign media reporting that Anthropic plans to launch its IPO as early as mid-November; OpenAI is expected to follow suit next year, both aiming for a $1 trillion valuation.

As more high-quality targets emerge, the logic of valuing large model companies based on "storytelling" is becoming increasingly untenable. Already-listed Zhipu and MiniMax serve as a cautionary tale.

In January this year, the two companies successively listed on the HKEX, with peak market capitalizations exceeding HK$1 trillion and approaching HK$400 billion, respectively. However, their market caps have now fallen to around HK$300 billion and over HK$70 billion, respectively.

The decline in their market caps can be attributed to market adjustments and the expiration of lock-up periods, but the more fundamental reason lies in the companies' financial fundamentals—revenues are rising, but their profitability has not kept pace.

According to Zhipu and MiniMax's 2026 mid-year reports, both companies' revenues for the past six months exceeded their full-year revenues from the previous year. However, after excluding accounting factors such as the revaluation of preferred stock fair value brought about by the listing, both companies' adjusted net losses widened.

The reasons are strikingly similar: the high training costs of next-generation models, coupled with the escalating computing power and inference costs driven by increasing user numbers and API calls, are eroding profits.

The reality is that stronger and larger models do not necessarily translate into easier profitability.

This also underscores The Dark Side of the Moon's current dilemma. As the number of AI targets in the market increases, investors have more references for valuing large model companies, and this open book is continuously squeezing the valuation space for unlisted large model companies.

This means that the later large model companies go public, the smaller the slice of the pie they may secure.

As market funds further converge toward high-quality AI targets, they prefer tangible cash flows over lofty narratives, which is why The Dark Side of the Moon is rushing to go public.

02 The Industry Narrative is Evolving

The narrative logic of the large model industry is undergoing a transformation.

During the "Hundred Models Battle" phase, companies vied on who had larger model parameters, faster iteration cycles, and smarter models; however, as large models become more widespread, the market is realizing that operational capabilities are more crucial than raw intelligence.

Consequently, intelligent agents have emerged as the new trend in the large model industry. After the "lobster craze" erupted early this year, the industry officially entered the era of long-context agents capable of proactively handling complex tasks.

The challenge is that an agent's proactive capabilities are not solely dependent on "intelligence" but also on the information density of the context and its ability to remember and understand that context—the more complex the task, the more dialogue it needs to remember and the more reasoning it triggers.

Moreover, agents can operate continuously 24/7, and each complex task they handle triggers multiple model inferences, with each interaction generating a longer token bill.

In July this year, The Dark Side of the Moon released the Kimi K3 model with 2.8 trillion parameters, currently the world's largest open-source model. However, less than 48 hours later, The Dark Side of the Moon issued an open letter stating that due to a shortage of computing power, it would suspend new C-end user subscriptions for Kimi K3.

While smarter and more usable large models sell more tokens, they also consume more computing power. If the gross margin per token is not substantial enough, the larger the scale, the steeper the losses may be.

As a result, investors have also learned to be prudent in their calculations. AGI remains the ultimate goal of the industry, but until then, which vendor can balance its computing power bills is equally important.

After all, neither the primary nor secondary markets can indefinitely sustain "dream-feeding." As the race nears its end, the industry needs to select the final survivors.

An industry consensus is gradually emerging: the true competitive edge for large model companies lies in possessing more computing power resources, producing tokens at lower costs, and establishing smoother business models.

Over the past six months, the entire industry has quietly shifted away from the "muscle show" of token consumption and toward improving token production efficiency through technological upgrades, investing in computing infrastructure, and binding industry chains.

In May this year, Yang Zhilin proposed the goal of "pursuing ultimate Token Efficiency," stating, "The higher the Token Efficiency, the better the results (companies) can achieve with as few tokens as possible."

Meanwhile, Zhipu, DeepSeek, and others are also advancing full-stack Infra strategies. Zhipu has launched the construction of a 1GW-scale domestic AI computing power data center; DeepSeek is planning a 1GW self-built intelligent computing base in Ulanqab.

On the other hand, the business models of large models are also quietly evolving.

Recently, according to multiple media reports including CCTV Finance, based on a financial overview provided by OpenAI to its investors, its ARR in September was $50 billion, about 30% lower than the $70 billion previously consensus in the capital markets.

Yet, at this juncture, OpenAI announced that it would make GPT-6 Luna available to free users.

Almost simultaneously, Anthropic introduced the lightweight model Claude Haiku 5.5, with its pricing reduced to one-tenth of the previous generation.

On one hand, the capital markets expect revenue scale; on the other hand, large model companies are actively lowering income thresholds. Both OpenAI and Anthropic have chosen to adopt tiered pricing for their models.

Their logic is consistent: companies cannot just pursue the strongest models but must also learn to calculate costs—first attract users with lightweight, low-cost models, then recoup the conceded profits through paid models, enterprise services, etc.

Only by building an ecosystem can subsequent commercial monetization occur.

03 Enhancing Commercialization Capabilities

Facing changes in industry trends, The Dark Side of the Moon is also keeping pace.

First, model tiering.

After launching the K3 model, The Dark Side of the Moon found that K3's sensitivity to historical thinking content led to longer task completion times and increased token consumption, drawing criticism from many users.

A smarter model is not necessarily the best choice for all users. Therefore, The Dark Side of the Moon launched K2.8 Preview in early September.

In terms of accessibility, only Kimi members paying $99/month or more can access K3, while K2.8 Preview is available to all members and directly offers 1M context. Externally, K2.8 is seen as a "lite K3."

The Dark Side of the Moon has come to realize that after occupying the "capability mindshare" with K3, it also needs a model with lower unit costs and access thresholds to expand its user base and provide a foundation for commercialization.

Second, accelerating commercialization.

On September 10, it officially launched the "Moon Landing Plan," collaborating with five system integrators to build a team of frontline deployment engineers (FDEs), deepening and expanding its services to secure a group of long-term clients with stable cooperation intentions.

On September 18, the K3 model landed on Amazon Bedrock, allowing global enterprise developers to directly access it. According to foreign media reports, The Dark Side of the Moon will receive a commission of up to 30% from K3's usage revenue.

For The Dark Side of the Moon, the commission model can convert model usage into predictable, recurring cash flows and provide the capital markets with a clearer valuation reference.

In addition, The Dark Side of the Moon's accumulated know-how in vertical domains has also become a deeper competitive edge.

It recently released a financial industry solution, integrating 10+ authoritative data sources, 9 financial skills, and 5 security compliance measures, with initial clients including Industrial and Commercial Bank of China and CITIC Securities.

Finally, ecosystem refinement.

Currently, The Dark Side of the Moon has launched the chatbot product Kimi, as well as professional productivity products Kimi Work and Kimi Code. It also recently introduced Kimi Code Desktop, providing developers with a complete programming assistance toolchain.

With the foundational model as the center, The Dark Side of the Moon's ecosystem radius is expanding, and the more diverse the product scenarios it reaches, the higher its revenue ceiling can be pushed.

From pursuing C-end scale to focusing on technological depth and B-end commercialization, The Dark Side of the Moon has quietly switched its revenue engine. The yardstick for valuation no longer solely depends on user numbers but on how indispensable users find it.

Looking at the global AI industry, this is also a highly instructive example.

When "stronger models" are no longer an absolute competitive edge, every large model company must reanswer the same question: how to turn technological leadership into a sustainable business.

What The Dark Side of the Moon is doing is comprehensively catching up in model capabilities, commercial scenarios, and ecological infrastructure. The more diversified its revenue structure, the more sustainable its technological capabilities can be monetized.

Yang Zhilin once said, "An IPO is a means to accelerate AGI; timing is everything." Now, the time has come, but it is also fleeting. For The Dark Side of the Moon to stay at the table, it must first squeeze onto this train.