Wall Street is slow to learn. When shares of various tech companies plummeted early last year, investors pointed to DeepSeek-R1. Just like then, a Chinese AI company is proving that the best LLMs can be virtually matched with an open-source alternative. Moonshot’s Kimi K3 scores slightly lower than Claude Fable 5 and OpenAI’s GPT-5.6 Sol on most benchmarks, but outperforms them on several benchmarks. Once again, this achievement is no direct cause for concern, despite the turmoil on the stock market.
At the time of writing, Asian companies in the chip industry have been particularly hard hit. Bloomberg reports that this sector has fallen by 6 percent today, while the impact on U.S. stocks remains limited. Nevertheless, commentators are eagerly describing another “DeepSeek moment,” and financial analysts see this “Kimi moment” as a new dimension, raising major questions once again about the massive AI expansion taking place primarily in the U.S.
Not a DeepSeek moment
The shock surrounding DeepSeek-R1 in January 2025 revolved around the fact that OpenAI had launched o1, a groundbreaking “reasoning” model, a few months earlier. Not only did it perform much better on benchmarks than other LLMs, but it also placed a new emphasis on computation time during inference. Following academic research into “test-time compute,” simply put, an AI model taking longer to think through its answers, o1 put this into practice in a way that other LLM players couldn’t easily replicate. The model didn’t have to come up with a final answer immediately but could first weigh multiple possibilities against one another.
The result was panic on the stock market, with estimates putting nearly one trillion dollars in market value lost in a short period, nearly 600 billion dollars of which was at Nvidia. The fear was that the entire AI expansion, as proposed at the time, would ultimately prove unprofitable.
Another factor was that GPT-4, released in March 2023, was clearly outperformed on many benchmarks by Claude 3 Opus just one year later. If o1 had remained dominant for just as long, OpenAI would have been sitting pretty. Nothing could be further from the truth. An obscure Chinese AI lab, which began as a “side project” of the quantitative hedge fund High-Flyer, first built the impressive DeepSeek-V3 and then R1. The result was an LLM that, on many reasoning benchmarks, resembled an o1 twin, albeit with slightly less impressive numbers. Shortly after R1, OpenAI released the previously announced o3-mini, which was many times cheaper than o1 and “only” twice as expensive as DeepSeek-R1. In other words, all the R&D funding and infrastructure expansion couldn’t prevent a sudden challenger from emerging.
Kimi K3 launches into a completely different world
In the year 2026, it is not o1 but the combination of Claude Fable 5 and OpenAI’s GPT-5.6 Sol that dominates. Some assume that open-weight competitors are always about six months behind the American closed labs. GLM-5.2 from the Chinese company Z.ai was released a few weeks ago and demonstrated on benchmarks that it could nearly match the performance of Opus 4.8 at a significantly lower price. Moonshot’s Kimi K3 is even more powerful and much larger. If this LLM is released as an open-weight model at the end of the month as announced, Moonshot claims it will be by far the largest AI model ever available for download. With 2.8 trillion parameters, initial benchmarks place it around the Fable 5 level, though its average score is still slightly lower than that of Fable 5 and GPT-5.6 Sol. Moonshot recommends supernodes with 64 accelerators or more for deployment, meaning that virtually only companies with rack-scale systems will be able to run this LLM on-premises.
The significance of DeepSeek-R1 and Kimi K3 for the market differs fundamentally. Aside from the fact that LLM development from January 2025 to July 2026 has been enormous, the situations were also completely different on their respective release dates. DeepSeek surprised both friends and foes by undercutting the price of OpenAI’s o1 by a factor of approximately 25. This is not the case for Kimi: according to Artificial Analysis, Kimi K3 is more expensive per actual task than GPT-5.6 Sol on Medium. Just as with GLM-5.2, a token-guzzler, the price per token is misleading for K3 when compared to the offerings from OpenAI and Anthropic. Of these models, only Fable 5 is significantly more expensive.
Why the panic?
There are plenty of reasons to be skeptical about the massive AI investments being made by big tech companies. Because new generations of Nvidia GPUs quickly outperform their predecessors and their efficiency gains make follow-on investments attractive, the pressure to depreciate AI hardware is high. And every time a new LLM emerges that’s more cost-effective than the current leader, the price of the AI capabilities that companies purchase drops.
Yet today’s panic isn’t entirely logical. The capabilities of xAI’s Grok 4.5 and Meta’s Muse Spark 1.1 were already approaching those of the top models from OpenAI and Anthropic. Although those models are not open-weight or Chinese, they certainly offer price competition for the still-expensive Opus 4.8. Kimi K3, as good as it is, actually offers less cause for panic than those models. It is currently available only as a hosted service and is not yet open-weight. As a result, users are still dependent on Moonshot’s infrastructure and must take into account the Chinese jurisdiction under which the company operates.
Frontier models have been interchangeable for some time now. Since 2024, there hasn’t been a period longer than a few months in which a single LLM builder has led the pack across the board. For now, Kimi K3 primarily poses a threat to the scarcity premium of proprietary AI models, not to the demand for computing power. The model is large, expensive to run independently, and not necessarily cheaper per completed task than its American competitors. That means the panic on the stock market is not a rational reaction to K3 alone. However, K3 is further proof that investors should not take the technological lead and pricing power of U.S. AI companies for granted.