# Moonshot's Bet on a Bigger Kimi K3 Is Paying Off in China's AI Race > Source: > Published: 2026-07-20 15:58:23+00:00 *Moonshot AI's bet on going bigger with Kimi K3 is paying off so well that it had to lock the doors on new users.* Moonshot AI didn't just release another chatbot upgrade on July 16. It released a 2.8 trillion parameter model, Kimi K3, that beat Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on LMArena's Frontend Code Arena benchmark, and three days later the Beijing startup had to shut the door on new subscribers because it didn't have enough chips to serve them. That's the story in one breath. Kimi K3 scored 1,679 Elo across 1,757 votes on the coding leaderboard, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. On July 19, Moonshot said on X it was temporarily pausing new subscriptions and prioritizing compute for current members. Demand in the first 48 hours after launch came close to the ceiling of the company's GPU capacity. You don't pause sign-ups when a product is struggling. You pause them when it's working too well. For most of 2026, the assumption in AI circles was that Chinese labs would keep winning by doing more with less, the DeepSeek playbook of cheaper training and cheaper inference. Moonshot took the opposite bet. Where DeepSeek's breakthrough was about squeezing costs down, Moonshot wagered that a much bigger model, engineered the right way, would pay for itself. According to Bloomberg, that wager is working. Moonshot says Kimi K3 delivers roughly 2.5 times the scaling efficiency of its predecessor, Kimi K2, through two architectural changes: Kimi Delta Attention, a hybrid linear attention scheme, and Attention Residuals, which change how information passes between layers. Together they cut KV-cache memory use by as much as 75% and speed up decoding by up to six times at the model's full one million token context window. Bloomberg's own reporting zeroes in on exactly that point: Kimi K3's real advantage may be memory engineering rather than raw compute, which is what lets Moonshot scale a model up without needing a proportional pile of new GPUs. Investors didn't need a technical paper to get spooked. Nvidia shed roughly $590 billion in market cap in a single session after Kimi K3's release. The VanEck Semiconductor ETF fell more than 20% from its late June record, and the Philadelphia Semiconductor Index posted its worst week since April 2025. AMD and Intel slid alongside Nvidia. The math is blunt: if an open weight model out of Beijing can beat Claude and GPT on a coding benchmark while undercutting their API pricing, the case for endless GPU spending gets harder to make. Frankly, it's the same panic DeepSeek caused in early 2025. It just comes with a bigger price tag this time. None of this is happening on a shoestring. Bloomberg reported in May that Moonshot raised about $2 billion in a round led by Meituan, valuing the company at more than $20 billion. By June, Bloomberg reported Moonshot was already back in the market seeking a new round at a valuation as high as $30 billion. The company's annual recurring revenue hit $300 million that same month, driven largely by API demand rather than consumer subscriptions. ## The GPU Math Still Has to Work Serving Kimi K3 isn't cheap on Moonshot's own infrastructure. One estimate puts the requirement at roughly eight Nvidia H100 or H200 chips just to run a single instance of the model. That's the practical reason behind the July 19 freeze, and it's why Moonshot is splitting future subscriptions into two tiers, including a dedicated plan for coding users, so it can ration compute toward the customers paying the most for it. Relief has a date attached, not just a promise. Moonshot has said it plans to release Kimi K3's full weights as open source by July 27, letting developers and cloud providers run the model on their own hardware instead of waiting on Moonshot's servers. That won't fix Moonshot's own capacity problem. It will spread the compute burden across everyone else's GPUs instead. Moonshot is also pushing toward a Hong Kong initial public offering, according to earlier Bloomberg reporting, and a model that tops coding leaderboards while selling out its own servers makes for an easier IPO pitch than most. Whether Kimi K3 settles the scaling laws debate for good or just wins this quarter is still an open question. What isn't in question is that Moonshot bet big, and right now, the market can't stop watching. **Also read:** [Jeff Bezos and the UK Government Just Bet $450 Million on CuspAI](https://startupfortune.com/jeff-bezos-and-the-uk-government-just-bet-450-million-on-cuspai/) • [Google Shares Rise After Report of a Chip Designed Around Gemini](https://startupfortune.com/google-shares-rise-after-report-of-a-chip-designed-around-gemini/) • [Elon Musk Calls a Reported $52 Billion SpaceX Foxconn Server Deal Fake News](https://startupfortune.com/elon-musk-calls-a-reported-52-billion-spacex-foxconn-server-deal-fake-news/)