Who's afraid of Kimi K3?
Davide Sciannimonaco — 21 July 2026
Moonshot's new model matches some of the best US systems and will be free to download. The sharpest sell-off hit Chinese model companies, not American ones. That tells you where the value in AI is going.
Bottom line
- Frontier AI is becoming a commodity. Kimi K3 matches Anthropic's Opus 4.8 and OpenAI's GPT-5.6 Terra on independent benchmarks, undercuts them on price, and is available now. Pricing power at the model layer is eroding on both sides of the Pacific.
- The whole AI complex sold off on the news, but the dispersion is the signal: Chinese model pure-plays fell hard (Z.ai -28%, MiniMax -16%) while semiconductors and platforms lost mid-single digits. The market is repricing the model layer, not the spending that flows through it.
With capability parity now demonstrated and China tech still trading at roughly half the growth-adjusted valuation of US peers, the shift of value toward the monetization layer accelerates. No change to our positioning; conviction increases.
What happened
On 16 July, Moonshot AI, a Beijing startup backed by Alibaba and Tencent, released Kimi K3, a mixture-of-experts model with 2.8tn parameters and a 1mn-token context window. The model is available now through Moonshot's API and apps, with an open-source release to follow at the end of the month. From that point, any company with sufficient computing capacity can download it and run it on its own infrastructure, free of licensing fees. On current benchmarks it sits within the leading group, trailing only Claude Fable 5 and GPT-5.6 Sol, at a fraction of their price:
| Model | Issuer | Artificial Analysis Intelligence Index | Input ($/mn tokens) | Output ($/mn tokens) |
| Fable 5 | Anthropic | 60 | 10 | 50 |
| ChatGPT 5.6 - Sol | Open AI | 59 | 5 | 30 |
| Kimi K3 | Moonshot AI | 57 | 3 | 15 |
| Opus 4.8 | Anthropic | 56 | 5 | 25 |
| ChatGPT 5.6 - Terra | OpenAI | 55 | 2,5 | 15 |
Still, K3 is priced several times above the previous Kimi generation. A Chinese lab is now raising prices, not cutting them.
The release landed days before the World AI Conference in Shanghai. Investors called it a new "Kimi moment". One caveat: part of the benchmark record is self-reported, and independent testing is still in its early days.
Impact on our Investment Case
Commoditization now runs in both directions
Like most recent frontier models, K3 is a mixture-of-experts: on any given request, only a small fraction of the model does the work, 16 of its 896 specialized blocks. Running it therefore costs far less than the 2.8tn headline suggests, which is how Moonshot can price it at a fifth of Fable 5. On the news, Semiconductors fell too, with Asian chip indices down around 6% on the day, as investors revisited AI capex assumptions. But a repricing of capex is not the same as losing your core asset overnight, and that is what happened to Z.ai and MiniMax: listed companies whose main asset is a proprietary Chinese model just saw a domestic rival giving a better one away. The same logic caps the pricing power of US labs over time. Owning a model is a shrinking moat, in Beijing as much as in California.
The money lands one layer down
Moonshot recommends running K3 on clusters of 64 or more accelerators; in its native 4-bit format, the model occupies roughly 1.4TB of memory. This is datacenter equipment, not a desktop, and every deployment adds to inference infrastructure demand. Monetization then happens through the companies that apply the model to specific industries, the pattern we described in April: SOE procurement under the 90% AI adoption target for 2030 buys verticalized solutions, not raw models. The US export controls that took Anthropic's Fable 5 offline on June 12 were lifted after 18 days, and access is restored. The precedent stands regardless: enterprises worldwide now know that access to a closed US model can be revoked overnight, which strengthens the case for open-source alternatives they can host themselves. K3 raises the capability ceiling of that fallback.
One point of caution
Private valuations of Chinese model labs are inflating quickly. Moonshot raised $2bn at a ~$20bn valuation in May and is reportedly in talks for a new round around $30bn; DeepSeek is said to be discussing a round near $70bn pre-money, a month after closing its first external round. Investors chasing the model layer should remain vigilant at these marks. The listed, liquid part of the value chain captures the spending with less froth.
Our Takeaway
Each shock has pushed in the same direction. DeepSeek proved China could build frontier AI cheaply. The June shutdown of Fable 5, reversed after 18 days, proved that access to closed US models is a policy variable, not a given. Kimi K3 proves China can build at frontier scale and charge for it.
Our conviction is on the ecosystem, not on any single layer or name. The China Tech strategy is systematic: it replicates our proprietary index of more than 1,000 Chinese technology companies, from small caps to mega-caps, with a mid-cap tilt and pure A/H-share exposure, rebalanced quarterly. There is no discretionary bet in it beyond the theme itself, and that construction fits the moment. If value keeps migrating from models to the infrastructure and applications that deploy them, the index follows the migration mechanically, wherever it lands on the cap spectrum, backed by the $2tn spending mandate of the 15th Five-Year Plan. Launched at $100 on February 4, 2025, the strategy stands at $162.7 as of July 20, 2026, up more than 60% since inception.
What could weaken the thesis operates at ecosystem level too: policy spending that fails to show up in corporate revenues through 2027, or a sanctions escalation broad enough to disrupt domestic deployment itself. Shifts within the ecosystem, including a re-concentration of value at the model layer, would change which names lead the index, not the case for owning it. The capability gap has closed faster than the valuation gap. We expect the second to follow the first.
Back to all articles
