Kimi K3 is not the threat you think it is

Etwa alle anderthalb Jahre macht an den Märkten dieselbe Sorge die Runde: Die enormen Investitionen in führende KI-Labore könnten sich als Blase erweisen, die von einem chinesischen Wettbewerber zum Platzen gebracht wird. Anfang 2025 war es DeepSeek R1, das diese Befürchtungen auslöste. Nun richtet sich die Aufmerksamkeit auf Kimi K3 von Moonshot AI, ein Mixture-of-Experts-Modell mit 2,8 Billionen Parametern und einem Kontextfenster von einer Million Tokens. Den von Moonshot veröffentlichten Benchmarks zufolge übertrifft es Claude Opus 4.8 in einer Vielzahl von Tests, bleibt jedoch hinter den Spitzenmodellen Claude Fable 5 und GPT-5.6 Sol zurück. Die Reaktion der Märkte ließ nicht lange auf sich warten: KI-bezogene Aktien gerieten unter Druck, begleitet von zahlreichen Kommentaren, die bereits von einem möglichen „DeepSeek 2.0“ sprachen.

Dabei lohnt sich ein Blick darauf, was nach DeepSeek tatsächlich passiert ist. Anfang 2025 herrschte die Befürchtung, ein leistungsfähiges, frei verfügbares und lokal betreibbares Modell könnte die wirtschaftliche Grundlage der westlichen KI-Wertschöpfungskette infrage stellen – von Grafikprozessoren über Rechenzentren und Energieinfrastruktur bis hin zu den KI-Laboren selbst. Rund anderthalb Jahre später zeigt sich jedoch ein anderes Bild: Die Investitionen in KI-Infrastruktur sind deutlich gestiegen, nicht gesunken. Gleichzeitig erwirtschaften die führenden KI-Unternehmen höhere Umsätze denn je. Unseren Schätzungen zufolge erzielt die Branche inzwischen annualisierte Erlöse von rund 175 Milliarden US-Dollar, während OpenAI und Anthropic ihre gemeinsamen Umsätze im selben Zeitraum von etwa 8 Milliarden auf rund 110 Milliarden US-Dollar gesteigert haben. DeepSeek hat die Erwartungen an die Branche also nicht zunichtegemacht. Rückblickend war das Modell vielmehr ein bedeutender technologischer Fortschritt, der die Effizienzgrenzen des KI-Trainings verschoben hat, ohne die Marktführer von ihrer Spitzenposition zu verdrängen.

Kimi K3 is not DeepSeek

This is the part of the analogy that collapses under scrutiny because Kimi K3 is, in almost every structural respect, the opposite of DeepSeek.

DeepSeek’s threat, to the extent it was one, was that it let you bypass the infrastructure buildout entirely. It was small enough and cheap enough to run locally, a model you could put on commodity hardware and walk away from Nvidia, the hyperscalers and the whole apparatus. Kimi K3 does none of that. K3 is so large, at more than 2.8 trillion parameters, that Moonshot itself says you need a cluster of 64 of Nvidia’s latest chips wired tightly together just to run it well. Here’s the napkin maths: at roughly 1.4 TB just for the weights in MXFP4, you need ten-plus H200s before you’ve even touched the KV cache on a million-token context window. This is not a model you self-host on a gaming rig. It requires Nvidia’s latest chips to run.

The efficiency gains in K3’s architecture, with Kimi Delta Attention, which cuts KV-cache transfer requirements by up to 10x, sound at first pass like exactly the kind of thing that should scare Nvidia and some other players in the semiconductor space. The reality is the opposite. Because those large weights still need to be distributed across GPUs via an optimisation called WideEP, spreading K3’s 896 experts so each chip holds only a sliver, the model actually consumes more network bandwidth. It is precisely the copper backplane of an Nvidia Blackwell 300 NVL72 system that makes this workable at all. K3 is a showcase for Nvidia’s highest-end rack-scale product, not a threat to it.

And then there’s price. DeepSeek’s whole pitch was being radically cheaper than the frontier. Kimi K3 is not cheap. Its API is priced at USD 3 per million input tokens and USD 15 per million output tokens, essentially in line with Claude Sonnet’s standard rate. Compare that with GLM 5.2, another well-regarded Chinese open-source model, at roughly USD 1.40/USD 4.40, and K3’s output price alone is more than three times higher. Once you account for the fact that K3 currently only runs in a maximum-reasoning mode, there is no cheap, low-latency variant for simple tasks, the real per-task cost looks even less like the old open-source discount. Artificial Analysis puts K3 at roughly USD 0.94 per task, in the same neighbourhood as GPT-5.6 Sol’s USD 1.04. Moonshot has clearly had to lean on aggressive infrastructure tricks, with a 90% discount on cached input and a system called Mooncake that offloads the KV cache to CPU DRAM and SSD, simply to make the economics work at this scale. That is not the profile of a cut-rate competitor. It is the profile of a lab straining to achieve frontier-adjacent pricing on a model that likely runs at much thinner margins than Anthropic’s or OpenAI’s, given Moonshot’s smaller user base and lower aggregate compute utilisation.

Distillation, and the asymmetry nobody talks about

None of this should take away from what Moonshot has accomplished. Building a model of this scale, with this level of agentic and long-horizon coding capability, is a genuine engineering feat, and the Chinese AI ecosystem deserves real credit for it. But it would be naive to pretend the achievement exists in a vacuum, uninformed by the outputs of frontier US labs.

Distillation, for readers unfamiliar with the term, is the practice of training a smaller or newer model to mimic the outputs of a larger, more capable one, essentially using the frontier model’s own responses as training data so the student model inherits much of the teacher’s capability without having to discover it independently. It is a legitimate and widely used technique inside labs on their own models. It becomes something else when it is done against a competitor’s model, at scale, via its API, which is functionally an attempt to extract the value of billions of dollars of training compute through a chat interface.

American labs and startups do not do this to each other’s frontier models at scale, and the reason is structural. If a well-funded US startup were caught distilling GPT or Claude and reselling the result, it would be buried in litigation and terms-of-service claims by legal teams with essentially unlimited budgets, and no venture capitalist would touch it. Chinese labs face no equivalent exposure. Anthropic said as much earlier this year when it disclosed that it was seeing large volumes of suspicious accounts, many traced to China, systematically querying its API in ways that violated its usage agreements, patterns consistent with exactly this kind of extraction. This is the asymmetry: one side of the ecosystem operates under enforceable rules against copying, while the other largely does not.

Maybe it’s not that they’re catching up

The narrative writing itself right now is that Chinese open models are closing the gap with the American frontier. That may be true. But there is a second explanation that gets much less airtime: American labs may simply be releasing more cautiously than they otherwise would. Anthropic’s own Mythos-tier models were released, then had access suspended within days to comply with export controls before being restored weeks later. That is a fresh, visible example of a frontier release running headlong into government scrutiny, and it would be surprising if it did not make every lab a little more conservative about how aggressively it pushes its most capable systems into public deployment. Some of the “gap closing” we are seeing may be less about the frontier standing still and more about it being walked onto the field more slowly than it could be.

What actually matters

Strip away the panic and there is a real structural question underneath it, and it has nothing to do with whether Kimi K3 is “as good as” Opus or Sonnet. It is about gross margins. If the model layer becomes more competitive, with more open-weight alternatives, more price pressure and thinner margins for everyone, the optimistic reading is Jevons Paradox: cheaper intelligence gets used more, and the increase in volume more than compensates for the lower unit price. There is something to that. But it is not the whole picture, and treating it as such misses the actual risk.

The real danger is the possibility that the model layer becomes commoditised before AI adoption has reached a steady state, leaving frontier labs without the inference gross profit that currently funds the next round of pre-training. That profit is, functionally, a tax, one that enterprises and consumers pay on every token, and one that every part of the rest of the stack quietly depends on because it is the mechanism pushing the frontier forward. Remove that tax too early and you do not get a permanently cheaper equilibrium. You get labs that can no longer afford to train the next model that would have justified all the compute being built for it. Enterprises are asking to spend the same on AI and get more capability. That only happens if someone keeps training smarter models, and smarter models are expensive to build.

Once adoption is closer to a real steady state, a more competitive, lower-margin model layer is unambiguously good for the rest of the stack. Semiconductor companies, hyperscalers and the enterprises deploying this technology all benefit from having one fewer layer extracting a toll. But we are not there yet, and pretending otherwise risks starving the part of the stack that is still doing the genuinely hard, genuinely expensive work of moving the frontier forward.

It is also worth reflecting on the actual track record here because it argues against the panic on its own terms. Open and Chinese models have repeatedly narrowed the gap with the closed American frontier. Not once, across two years of these panics, has one become the leader. DeepSeek did not. No open model since has. Kimi K3, for all its genuine impressiveness, is, by Moonshot’s own framing and every independent benchmark, a very good model that sits roughly one generation behind the frontier, not at it.

The AI investment flywheel

Source: DPAM, 2026

The optimistic case

Here is the point worth coming back to: every one of these panics has, so far, been a false alarm about the wrong risk. The risk was never that a competitor would build something good enough to end the category. It was that the companies doing the hard, expensive work of pushing capability forward would lose the economic ability to keep doing it. So far, they have not. Revenue is up more than tenfold in a year and a half. Capital expenditure is up. And now we have a second data point: a genuinely capable open model that, rather than undercutting the frontier’s margins, has been forced by its own scale to price itself into the frontier’s neighbourhood.

Instead of being a story about the West losing its lead, this is a story about what it actually costs to compete at the frontier, told this time through someone else’s balance sheet. The “model tax” survives because it turns out that nobody, American or Chinese, has yet figured out how to build something this good without also having to charge for it.