Following the release of Moonshot AI's Kimi K3, demand far exceeded expectations, leading to compute constraints and the suspension of new consumer subscriptions on July 19. With a parameter scale of 2.8 trillion, Kimi K3 is the world’s first open-source model of its 3-trillion-parameter class. On the Arena frontend development leaderboard, Kimi K3 ranked first with a score of 1,679, outperforming Claude Fable 5 and GPT-5.6 Sol. Although it leads in certain areas, its overall performance still lags behind the top proprietary models. The company holds cash reserves exceeding RMB 10 billion and has completed multiple funding rounds, with a valuation surpassing $20 billion. After initiating the dismantling of its VIE structure in May, Moonshot AI has sent a listing proposal to investors and expects to complete its Hong Kong stock exchange listing within as little as six months.

Article author and source: 36Kr

After Kimi K3 became a sensation, Moonshot, hailed as one of the "Six Little Tigers of AI," faced a "compute shortage."

On July 19, the Moonshot AI Kimi team announced that, due to unforeseen computational challenges, they have decided to immediately suspend new consumer subscriptions and allocate all available computing resources to serving existing subscribers.

At the same time, the Kimi team stated that it is fully accelerating its computing power expansion; as new computing resources come online gradually, more subscription slots will be opened until full subscription services are restored.

Kimi K3, developed by Moonshot AI, is a next-generation model with a parameter scale of 2.8 trillion, making it the world’s first open-source large model in the 3-trillion-parameter class.

According to Arena's latest frontend development ranking, Kimi K3 scored 1679 points, surpassing Claude Fable 5's 1631 points and GPT-5.6 Sol's 1618 points to claim the top spot.

However, Moonshot AI openly acknowledged that Kimi K3's overall performance still lags behind the top proprietary models, Claude Fable 5 and GPT-5.6 Sol.

It is worth noting that Yang Zhilin, founder and CEO of Moonshot, stated in last year’s internal letter that the company held over RMB 10 billion in cash and could raise significantly more capital from the primary market, adding that “there is no urgent need to go public in the short term.”

However, entering 2026, the company's capital strategy has rapidly shifted. Moonshot has already sent a listing proposal to investors and expects to complete its Hong Kong stock exchange listing in as little as six months.

Kimi is experiencing a computing power shortage and has paused new consumer subscriptions.

On July 19, the Kimi team announced that since the release of Kimi K3, the company has received far more support than anticipated, but also faces unforeseen compute capacity challenges—over the past 48 hours, user request volumes have significantly exceeded the team’s estimates and are approaching the current cluster’s capacity limit.

To ensure the experience of existing subscribers, the team has decided to suspend new consumer subscriptions effective immediately, dedicating all available computing power to serving current subscribers and fully protecting their rights and benefits.

In response to insufficient computing power, the Kimi team stated that the company is fully accelerating its efforts to expand computing capacity. As new computing resources come online, it will gradually open more subscription slots until normal subscription services are fully restored.

For all new subscribers going forward, the Kimi team will separate the core Kimi benefits (including Kimi Web, Kimi App, and Kimi Work) from the Kimi Code benefits to enable more precise allocation of computing power and ensure an optimal user experience.

According to Ban Yue Tan, computing power is a new form of productivity that integrates information processing capability, network transmission capacity, and data storage capacity, primarily categorized into general-purpose computing (general computing), intelligent computing (intelligent computing), and supercomputing (supercomputing).

Among these, intelligent computing power is primarily used for complex data analysis and artificial intelligence tasks. For a long time, China's computing power market has faced a structural imbalance characterized by relatively excess general-purpose computing power and relatively insufficient intelligent computing power.

In fact, it's not just the Kimi team—the shortage of computing power may have become one of the key factors limiting the development of large models worldwide.

Back in January this year, Zhipu, known as the "first listed company in large models," issued a lock-up announcement for the GLM Coding Plan, stating that with the launch of GLM-4.7, the number of GLM Coding Plan users experienced rapid growth, leading to a temporary strain on the company's computing resources.

At that time, Zhipu stated that it would temporarily limit the sale of the GLM Coding Plan, reducing daily sales to 20% of the current level, thereby freeing up more computing resources for Zhipu’s long-time users and providing a smoother programming experience.

Additionally, in April this year, major model providers such as MiniMax, Kimi, and Zhipu all experienced API overload or service outages.

In addition to domestic large model players, overseas AI giants also face similar challenges. As reported by The Wall Street Journal in March, due to computing power constraints, OpenAI was forced to scale back its operations and abandon the highly anticipated Sora video generation application.

Additionally, Anthropic, another industry leader, has also faced challenges with frequent system outages and a significant decline in service stability. In March of this year, Anthropic announced restrictions on the number of tokens users could use during peak hours from 5 a.m. to 11 a.m. Pacific Time on weekdays.

Kimi K3 tops the front-end development ranking, but its overall performance still has room for improvement.

As the source of this "compute shortage" from Moonshot, Kimi K3 has recently been making headlines.

According to Moonshot AI, Kimi K3 is a 2.8-trillion-parameter model built on the KDA hybrid linear attention mechanism (Kimi Delta Attention) and Attention Residuals technology, natively supporting visual understanding and featuring a 1-million-token context window.

The Dark Side of the Moon also stated that Kimi K3 is the world's first open-source model at the 3-trillion parameter scale, designed for cutting-edge intelligent scenarios such as long-range programming, knowledge work, and reasoning.

In terms of pricing, Kimi K3 uses a pay-as-you-go model: input is priced at ¥20 per million tokens (cache miss) and ¥2 per million tokens (cache hit), while output is priced at ¥100 per million tokens.

According to NetEase Intelligence, in Arena’s front-end development leaderboard updated on July 16, Kimi K3 ranked first with a high score of 1679, surpassing Claude Fable 5’s 1631 and GPT-5.6 Sol’s 1618.

Arena stated that in the frontend domain, Kimi K3 ranked first in six out of seven categories—branding and marketing, reference-based design, data and analytics, consumer products, simulation, and content creation tools—and came in second in gaming, behind Fable 5.

Kimi K3's impressive test performance quickly sparked widespread discussion overseas. Russ Salakhutdinov, the advisor of Moonshot AI's founder Yang Zhilin, posted a message praising the new version of Kimi for bringing a major breakthrough to the open-source community.

Additionally, Tesla CEO Musk commented directly in the comments section of the related review, calling it “Impressive.”

However, Arena explicitly stated that its ranking measures frontend web development tasks, including agent programming workflows requiring multi-step reasoning and tool usage, rather than the model's overall capabilities.

The Dark Side of the Moon also acknowledged that Kimi K3's overall performance still lags behind the top proprietary models, Claude Fable 5 and GPT-5.6 Sol, but it demonstrated cutting-edge capabilities across the full suite of evaluations, consistently outperforming all other models.

According to a ranking published by the independent evaluation agency Artificial Analysis, Kimi K3 has an intelligence score of 57, lower than Claude Fable 5’s 60 and GPT-5.6 Sol’s 59, but higher than Opus 4.8’s 56.

Artificial Analysis also noted that, compared to Kimi's previous model, K3's answer accuracy improved from 33% to 46%, but its hallucination rate increased from 39% to 51%.

Notably, after the announcement that Kimi K3 topped the front-end development ranking, the U.S. stock market experienced "DeepSeek Moment 2.0"—the AI sector lost approximately $470 billion within 72 hours.

The 21st Century Business Herald pointed out that the core of this Wall Street plunge lies in Kimi K3 targeting the lifeline of U.S. tech giants: code generation is one of the most lucrative pathways for AI commercialization, directly threatening the technological moat that overseas closed-source giants rely on to maintain high pricing.

According to Artificial Analysis, the average cost for Kimi K3 to complete a task is $0.94, placing it in the same tier as GPT-5.6 Sol at $1.04, and just half the cost of Opus 4.8.

Zhang Yutong, CEO of Moonshot, recently revealed that data showed the company recorded its largest single-day increase in annualized revenue (ARR) on the K3 launch node on July 17.

From “no rush to go public” to preparing for an IPO

According to Tianyancha information, Beijing MoonShadow Technology Co., Ltd. (short for "MoonShadow") was officially registered in April 2023, with Yang Zhilin as the actual controller, holding a 51.83% stake.

Public records show that Yang Zhilin earned his bachelor’s degree from Tsinghua University and his Ph.D. in computer science from Carnegie Mellon University. He previously worked at Google Brain and the U.S. startup FAIR, and is also one of the key technical contributors to China’s earliest large models, including PanGu and Wudao.

According to Yang Zhilin’s advisor, Russ Salakhutdinov, Yang could have stayed in the U.S. to join Apple, but he remained determined to return home and start his own company. In his view, he would regret it for the rest of his life if he never even tried to found his own company.

Or, due to confidence in Yang Zhilin, Moonshot AI quickly attracted significant attention from investors after its establishment.

According to Tianyancha, as early as June 2023, Moonshot AI had secured nearly RMB 2 billion in seed funding from Sequoia Capital China, ZhenFund, Jinri Capital, and Lisi Capital.

The following month, Moonshot completed its Series A funding round, with investors including BlueRun Ventures, Xianghe Capital, and Meituan Dragonfly.

At the beginning of 2024, Moonshot AI completed an A+ round of financing exceeding $1 billion, bringing the company’s valuation to $2.5 billion.

Subsequently, Moonshot AI completed additional funding rounds: a Series B round of over $300 million and a Series C round of $500 million.

According to LatePost, Yang Zhilin revealed in a 2025 internal memo that, from September to November last year, the company’s average month-over-month growth in paid users overseas and domestically exceeded 170%. By year-end, the company had over RMB 10 billion in cash reserves.

Yang Zhilin frankly stated that, compared to the secondary market, he can raise significantly more capital from the primary market, so “there’s no rush to go public in the short term.”

Indeed, this year alone, Moonshot has completed three additional funding rounds, pushing its valuation past $20 billion.

However, Yang Zhilin’s attitude of “not rushing to go public” is quietly changing. In May this year, Phoenix Weekly, citing Bloomberg, reported that Moonshot AI has notified its shareholders and officially initiated the dismantling of its VIE and red-chip structures to clear the way for a Hong Kong IPO.

Some analyses suggest that this move indicates the company has entered the IPO preparation phase. Huang Lichong, President of Foresight International Capital, stated that dismantling the VIE structure and red-chip arrangement essentially clarifies the placement of control, responsibility, and regulatory boundaries.

Huang Lichong analyzed that, if things go smoothly, the window could be six months to a year; however, if regulatory, shareholder, tax, and foreign investment arrangements are complex, it could take longer.

Shortly after the release of Kimi K3, Moonshot AI has made new progress in its listing plans. According to Nanfang Daily Business News on July 20, Moonshot AI has sent a listing proposal to investors and expects to complete its Hong Kong stock exchange listing within as little as six months.

The report also noted that in June of this year, Moonshot's annual recurring revenue (ARR) reached $300 million and grew several-fold after the release of the K3 model, leading the company to believe that the timing for an IPO is now right.

However, Moon Shadow has not yet made any public response regarding the latest listing schedule, nor has it disclosed specific progress on filing documents or overseas listing filings.

It is worth noting that on June 29, Moonshot AI posted a statement indicating that recently, multiple false claims regarding financing and equity transactions under the company’s name have emerged, and the company has identified certain institutions and individuals suspected of fraud.

In this regard, Moonshot AI issues a formal statement regarding the relevant matters: All financing activities of the company are solely managed by the company itself; any transfer of existing shares (including common shares, incentive equity, etc.) must be approved internally by the company; and the allocation of new share financing amounts is confirmed solely based on actual fund receipts.

We at Radar Finance will continue to monitor the listing progress of Moon Shadow.