Kimi’s AI Leap: Bitcoin Feels the Squeeze as China’s Kimi Tops Claude and GPT in Coding Benchmark
Bitcoin, Ether, and the wider crypto market fell on Friday. The name traders kept bringing up was Kimi K3, a new AI model from a Beijing startup that reportedly outscored Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 in coding benchmarks. Analysts dubbed it the “Kimi moment,” comparing the sell-off with earlier market shocks tied to Chinese AI releases. Crypto looks like a separate story. It isn’t. Lately, Bitcoin has moved alongside semiconductor stocks, while several mining companies have begun turning their facilities into AI data centers. My take: that connection now matters more than the crypto headlines.
Moonshot AI released Kimi K3 on Thursday. By Friday morning, AI and semiconductor stocks were sliding across Asia. Traders quickly reached for the “DeepSeek shock” comparison, which Bloomberg said erased roughly $600 billion from Nvidia’s market value in one session. Most comparisons stop there. That’s only half right. According to Moonshot, K3 has 2.8 trillion parameters and a one-million-token context window, four times the size of its predecessor’s. Huge numbers. Still, they miss the important engineering detail: K3 uses a mixture-of-experts design and calls on only 16 of its 896 specialists for any given task. Moonshot says that makes the model cheaper to run than its total size suggests. The company also says K3 scales 2.5 times more efficiently than the previous version.
K3 scored 1,679 on Arena’s Frontend Code leaderboard. Claude Fable 5 scored 1,631. GPT-5.6 reached 1,618. K3 took first place in six of seven categories and rose 17 spots from Moonshot’s previous model, which ranked 18th. That’s quite a leap. Did K3 beat its rivals everywhere? No. Moonshot admits K3 may still lag behind the strongest Claude and OpenAI setups on wider general-knowledge tests. I’ll be honest: that caveat matters less to me than it might seem. K3 won in coding, and coding is where companies are prepared to spend serious money.
K3’s license could matter more to investors than the benchmark result. Counter to the usual obsession with leaderboard rankings, access may be the real market-moving feature here. Moonshot plans to release the full open-weight model on July 27. People will be able to download it and run it on their own hardware without license fees. Anthropic’s Fable 5, released last month, and OpenAI’s GPT-5.6, released a week ago, remain closed and charge users according to usage, their developers say. Much of the investment in AI infrastructure rests on a blunt assumption: powerful models will remain scarce and expensive. They will also remain largely American. A free Chinese model sitting atop a coding leaderboard makes that assumption harder to trust. Chinese stocks took a beating too. Market data showed Z.ai down about 27% and MiniMax off roughly 16%.
Crypto’s latest headache did not start on-chain. It arrived through semiconductor stocks. Analysts said Bitcoin tracked chip shares all week. Last Friday, BTC rose 4% while South Korea’s Kospi gained 8%; SK Hynix priced $26.5 billion in American depositary shares. This Friday, Bitcoin fell after a model released in Beijing made the same bet look too expensive. Why does this matter? Because large investors increasingly trade Bitcoin like any other risky technology asset. When AI stocks tied to hardware fall, Bitcoin often follows. BTC slipped below $65,000 without any major news about crypto regulation or adoption. Investors got nervous elsewhere and sold risky assets across the board. Nothing more exotic happened. Sometimes it really is that simple.
Crypto infrastructure has a direct stake in this too. For the past two years, Bitcoin miners have tried to recast themselves as landlords for AI data centers. Reports say they have signed long leases with model developers. They have also poured money into sites used for training and inference. The bet depends on computing power staying scarce and expensive. K3 complicates that pitch. If capable models are free to download and cost less to run, demand for expensive specialist data-center space may not grow as quickly as miners expect. Revenue forecasts could suffer. So could valuations. Marathon Digital (MARA) and Riot Platforms (RIOT), for instance, have invested heavily in this shift. If investors stop believing in it, their shares could fall and sour the mood across crypto. My read: that outcome is hardly guaranteed, but it is much tougher to shrug off now.
What this means
K3 suggests capable open-weight models could arrive sooner and cost less than investors expected. For crypto, the immediate concern is the promise miners have made about becoming AI data-center operators. Many of those plans assume demand for computing capacity will keep climbing. They also assume margins will stay high. K3 puts both assumptions in doubt. Is this enough to break the miners’ AI strategy? Not by itself. Investors may, however, start asking whether these hybrid mining and AI businesses can deliver the returns their owners have advertised. Public miners would probably take the first hit. Bitcoin could follow if traders continue to use their shares as a stand-in for the wider crypto market. In my view, that proxy relationship is the part worth watching.
Marathon Digital and Riot Platforms deserve attention over the next few weeks. Another bad market day would be noisy; lower AI revenue forecasts would say far more, particularly if either company delays leases or cuts capital spending. July 27 is the other date to watch. Moonshot plans to release the full K3 model then, allowing developers to test its price and performance claims beyond controlled benchmarks. Yes, that shifts attention away from Friday’s benchmark win. It should. Those tests may change how investors value AI infrastructure and crypto-linked data-center companies. For Bitcoin, $60,000 is the next clear line. If BTC stays below that level, Friday’s sell-off may have been the start of something broader rather than a one-day scare.