Models like Kimi K3, Grok 4.5, and Muse 1.1 may prevent the dominance of 2-3 frontier labs with 90% inference margins from hurting other AI ecosystem layers
> lower modelBirk /@bare_birk:Interesting post from @GavinSBaker! Kimi K3 is bad for Anthropic and OpenAI but good for all other companies. Margins will go from the frontier labs, to all other companies in the sector. Infrastructure will still be very important in both scenarios (Opensource vs closed source)Steven Lubka /@dzambhalahodl:Kimi3 is a good model, but it is not a “cheap” model nor is it a “small” model. Kimi3 showed that China can train a decent competing model less than 6 months behi
@gavinsbakerGavin Baker
Context & Ripple Effects
Related coverage traces Kimi K3 from Moonshot AI’s announced release and planned weight publication to benchmark results that place it near leading proprietary systems in coding-oriented tests. It also includes a counterpoint that the model is neither small nor inexpensive and should not be read as evidence that China has overtaken the frontier.
The significance is therefore less a single benchmark rank than the growing availability of credible alternatives to a tightly concentrated set of frontier-model suppliers. The article frames Kimi K3 alongside Grok and Muse as evidence that model-layer economics may be contested even while infrastructure remains strategically important.
First-order effects
Anthropic and OpenAI face more immediate pressure on model positioning and the inference margins implied by exclusive access to top-tier capability.
Companies building AI products gain another potential supplier—and, if Kimi’s planned weight release materializes, a route to more direct control over deployment—rather than relying solely on closed frontier APIs.
Second-order effects
More credible substitutes strengthen model buyers’ negotiating leverage, pushing frontier labs to compete more on price, reliability, distribution, and product integration rather than benchmark leadership alone.
Application companies and infrastructure providers can capture more of the value if lower model-layer rents leave customers with more budget and flexibility to spend on deployment, tooling, and inference operations.
Third-order effects
If several near-frontier models remain available, the model layer could become less concentrated even without models becoming cheap: differentiation would shift toward distribution, operational execution, and specialized products.
Infrastructure remains central in either a more open or more closed model market, but the durable question becomes who captures inference economics—model owners or the broader stack that serves model users.
The trend: This is a data point in the shift from frontier-model scarcity toward a more competitive AI supply market, where buyer power and value capture increasingly extend beyond a few model labs.
That missing token efficiency is coming. Can't say more now, but efficient frontier long context reasoning/understanding and other TTC breakthroughs are on the horizon. Architectural improvements are often ignored in these benchmarks, but I expect major shifts by EOY.
This post is key. The cheaper AI gets, the more opportunity there is for the entire ecosystem - especially including end-customers - to benefit. Everything is bottlenecked by being able to successfully and cost effectively deploy AI in real workloads. Any time we can lower the
A bullish Kimi K3 argument: “Anything that lowers margins and increases competition at the model layer is good for every other AI layer: power, semiconductors, hyperscalers, neoclouds and yes even software.”
...the hypothesis that they have much more advanced model checkpoints internally that are already being used for RSI. In the latter scenario, reaching RSI even a few months ahead of other labs might be enough to cement a permanent lead.
“Anything that lowers margins and increases competition at the model layer is good for every other AI layer: power, semiconductors, hyperscalers, neoclouds and yes even software.”
genuinely one of the best bull posts I've ever read > China open source ai is better than US? Long all capex beneficiaries > China open source ai is NOT better than US? Long all capex beneficiaries
Great insights. Gavin nails it. If an oligopoloy of labs sustain 90% inference margins, they capture most of the economics and eventually vertically integrate the stack & squeeze chips, power, data centers, cloud and software. More competition at the model layer —> lower model
Interesting post from @GavinSBaker! Kimi K3 is bad for Anthropic and OpenAI but good for all other companies. Margins will go from the frontier labs, to all other companies in the sector. Infrastructure will still be very important in both scenarios (Opensource vs closed source)
Kimi3 is a good model, but it is not a “cheap” model nor is it a “small” model. Kimi3 showed that China can train a decent competing model less than 6 months behind the US. It did not show that they can compete with Frontier at a similar efficiency advantage to Deepseek etc.
oh no what if the big labs can't easily recoup all of their capex on massive data centers and some of the buildout capacity gets offloaded to other providers who deliver it to the long tail of enterprises with a software stack for serving and continually improving open models
Kimi k3 is an incredible model. It is not an incredible value. In most tasks, it comes out to roughly the same cost as GPT-5.6 Sol. K3 is half the price of 5.6 Sol per token. GPT-5.6 uses half as many tokens. Price evens out. GPT-5.6 is 2x faster TPS, so it gets work done ~4x [im…
I've been using Kimi K3 for ~16 hours now. The model is clearly good at a lot of different things (especially frontend), but non obvious reason why people are enjoying it so much is that it clearly does not follow the same rules in terms of safeguards and copyright. Kimi will [im…