Kimi K3 has most of the tech community enthused about the potential of open-source AI, but two companies are likely not amused.
That’s the argument being laid out by Atreides Management CIO Gavin Baker, who has taken to X with a lengthy breakdown of why Moonshot AI’s new open-weight model, which has beaten Fable 5 and GPT 5.6 Sol on the Frontend Code Arena leaderboard and placed third overall on Artificial Analysis’s Intelligence Index, could mark a genuine turning point for the industry. Baker’s framing is blunt: this is good news for almost everyone in AI, and bad news specifically for Anthropic and OpenAI.
The Core Argument: Margins Are A Zero-Sum Game Across Layers
Baker’s thesis rests on a fairly simple idea about how value gets distributed across the AI stack. A world where only two or three frontier labs dominate, each running inference at 90% margins, is a world where those labs eventually become the buyer of last resort for power, data centers, chips and cloud capacity. Baker argues they would inevitably vertically integrate into those layers over time, while also swallowing up the application and software layers built on top of their models.
Kimi K3 disrupts that outcome. According to Baker, anything that compresses margins at the model layer redistributes profit to power providers, semiconductor makers, hyperscalers, neoclouds, and software companies. This is precisely why, in Baker’s view, Nvidia’s Jensen Huang has been such a vocal champion of open-source AI. Huang has made similar arguments before, once calling the open-sourcing of a major Chinese model a win for America precisely because it kept the ecosystem running on Nvidia hardware rather than ceding ground to alternatives. An open-weight model demands exactly the same compute to run as a closed model of comparable size and architecture, so every dollar that isn’t captured as model-layer margin flows downstream instead.
Why K3 Specifically Complicates The Picture
Where Baker’s analysis gets more nuanced is on the question of whether K3 is actually the model that triggers this shift. He points out that K3 is priced roughly in line with GPT 5.6 Terra on a per-token basis, which he reads as a signal that Moonshot’s model is less computationally efficient than OpenAI’s, since GPT 5.6 is almost certainly priced at a fatter margin. K3’s own published pricing puts output at $15 per million tokens, a steep jump from K2.6, and Artificial Analysis has noted the model costs around $0.94 per task on its Intelligence Index, close to GPT-5.6 Sol’s $1.04.
Baker’s issue is token efficiency, not raw benchmark scores. He describes K3 as a “token wastrel,” meaning it needs more tokens to complete comparable tasks than GPT 5.6 or Grok 4.5, which drives up its real cost per task even where headline pricing looks competitive. Intelligence per dollar, not intelligence per token, is what Baker considers the actual metric that determines which company wins over time. That’s a meaningfully different lens than the one most benchmark writeups apply, and it’s why he stops short of calling K3 itself the “Sputnik moment.” The real disruption, in his telling, only arrives with an open-source frontier model that is both highly capable and token-efficient, something K3 hasn’t quite delivered yet.
Two Reasons Anthropic And OpenAI Might Still Be Fine
Baker attaches two caveats to his own thesis, both borrowed from public commentary elsewhere. The first, credited to Benchmark’s Eric Vishria, is that the product and harness built around Claude and ChatGPT may now matter more to users than the underlying model’s raw capability. The second is the possibility that both labs are sitting on internal checkpoints well ahead of what’s publicly available, checkpoints already being used to accelerate recursive self-improvement. If either lab reaches that stage even a few months before its rivals, Baker suggests the lead could become effectively permanent.
He’s careful to note that both of these remain open questions rather than settled conclusions, and expects clarity within a matter of months rather than years.
The Bigger Picture: Vertical Integration Works Just As Well As Open Source
Baker makes a point that’s easy to miss: open-source models aren’t the only path to compressing model-layer margins. A vertically integrated model company such as Meta, SpaceX or Google achieves the same effect, because those companies don’t need to extract profit at the model layer specifically when they can capture it anywhere else in their stack. That’s why, in Baker’s view, it was so damaging for OpenAI and Anthropic when Google briefly matched them on model competitiveness, and why Grok 4.5 and Meta’s Muse 1.1 carry similar weight to K3 in this analysis, even though only one of the three is open-weight.
The reaction across the AI community to K3’s release has largely echoed this framing, with several researchers and industry figures describing the model as a genuine inflection point for how the US and Chinese AI ecosystems compete, and for what that competition means for the labs sitting at the top of the pile today. Whether K3 turns out to be the actual trigger or merely a preview of what’s coming, Baker’s underlying argument is that the days of assuming Anthropic and OpenAI will simply keep compounding their lead are no longer a given.