Kimi K3: The AI Model That Could Break the Oligopoly—Or Just Burn Cash

Gavin Baker didn’t just publish a memo. He fired a shot across the bow of every AI model company, and the echo is rattling crypto’s AI token economy. The CIO of Atreides Management dropped a bombshell: Kimi K3, a new frontier model from China’s Moonshot AI, might mark a turning point for the entire industry. Not because it’s brilliant—but because it’s almost there, and that’s the threat. Baker’s logic is cold, calculated, and deeply familiar to anyone who’s watched DeFi compress margins into dust: when competition heats up, value doesn’t stay with the models. It flows upstream to the picks and shovels—power, chips, data centers, cloud, and software. Hackers don’t hack, they listen. And right now, capital is listening to the sound of model profits vaporizing. Here’s the raw data that sparked the fuss. According to Artificial Analysis, K3 costs about $0.94 per task. Compare that to GPT-5.6 Terra at $0.55 and GPT-5.6 Sol at $1.04. K3 is 71% more expensive than the cheapest GPT variant. That’s not a winning price tag. Baker calls it a ‘token efficiency’ problem—the model burns too many tokens for the value it delivers. But he’s not dismissing K3. He’s using it as a signal: the barrier to entry for frontier AI is no longer insurmountable. A new player can match the big dogs on capability, even if not on cost. That, he argues, is the beginning of commoditization. I’ve seen this movie before. During the Ethereum merge, I hosted watch parties in Mexico City, tracking the shift from PoW to PoS in real-time. The crowd was buzzing about how the merge would change mining profits. Sound familiar? The merge wasn’t just a technical upgrade—it was a redistribution of value. Miners got squeezed, and L2s, staking pools, and liquid staking tokens popped up to capture the overflow. The same thing is happening in AI. Baker explicitly calls out the winners: power companies, chip makers, data center operators, cloud providers, and software firms. He says ‘almost all other companies’ outside the model layer will benefit. That’s the DeFi playbook—when the base layer gets industrialized, the margins move to the infrastructure. But let’s get into the core: why now, and why K3? Baker’s thesis rests on the idea that model-level competition will compress profit margins until they’re razor-thin. Currently, OpenAI and Anthropic sit on a duopoly, charging premium API fees and raking in cash. A single new entrant with similar quality—even at higher cost—shows that the moat is not as wide as people think. K3 may be inefficient now, but efficiency can be optimized. The real turning point, Baker says, will come when an open model (think Llama or Mistral) achieves the same capability at a fraction of the cost. That opens the floodgates. And when that happens, the model layer becomes a commodity. Margins collapse. The only ones who profit are those selling the shovels: NVIDIA’s GPUs, Azure’s cloud, liquid cooling at data centers, even the power plants that keep the lights on. Based on my audit experience with DeFi protocols, I can tell you the same pattern repeats. When Uniswap v3 launched, everyone thought the AMM model was untouchable. Within a year, forks and optimizations ate into its market share. The real winners? The gas-guzzling L1s and the MEV searchers. In AI, the analog is clear: the model is the AMM, and the infrastructure is the blockchain. K3 is just the first fork that shows the original isn’t invincible. Baker even nods to the importance of ‘product, tooling, and in-house models’ as potential moats for OpenAI and Anthropic—but he implies those are weaker than the narrative suggests. He’s betting on commoditization. Now for the contrarian angle—and this is where the story gets spicy. Baker is a professional investor. His fund likely owns positions in power, chip, and software stocks. His ‘value transfer’ argument could be self-serving. If everyone piles into infrastructure, the models might get cheaper faster, but that also means the infrastructure gets overbought. There’s a real risk that K3 is a mediocre model hyped by a fund manager who wants attention for his thesis. Let’s look at the data: K3 costs $0.94 per task, but that’s only 70% more expensive than GPT-5.6 Terra. Maybe Moonshot AI can cut that in half with a better inference engine within six months. If they do, K3 becomes competitive, and the whole commoditization narrative accelerates. But if they don’t, K3 is a footnote. The market already knows this. AI tokens like FET, AGIX, and RNDR have pumped on the infrastructure narrative, but they’re pricing in a future that might not arrive on K3’s timeline. The real turning point, as Baker admits, needs an open model with high token efficiency. That hasn’t come yet. And here’s the cryptonative twist: decentralized compute networks. If model profits are squeezed, the demand for cheap compute will explode. That’s where Render, Akash, and IO.NET come in. They offer decentralized GPU resources at potentially lower prices than AWS or Azure. But they need scale. K3’s inefficiency actually helps them—if training and inference need more tokens, more compute is required. That’s bullish for decentralized compute. But it also means the AI model companies will fight to control that compute, possibly building their own chips or signing exclusive deals. The merge wasn’t a smooth transition, but the real merge is between AI and crypto—and it’s happening on the infrastructure layer, not the model layer. So what’s the takeaway? Watch the cost curves. Over the next 12 months, track K3’s token price per task. If it drops below $0.40, the commoditization thesis gains legs. Also watch for any open model that matches GPT-5 on both quality and cost. That will be the real bellwether. For crypto investors, the play isn’t to bet on AI model tokens—those are the new DeFi forks that will come and go. The play is the infrastructure: decentralized compute, data availability layers (yes, I still think DA is overhyped for most rollups, but for AI data it might matter), and energy assets. Hackers don’t hack, they listen to the flow of capital. Right now, capital is flowing toward picks and shovels. K3 is just the noise that proves the signal. The turning point isn’t here yet. But the warning shot has been fired. Don’t get caught holding the model tokens when the squeeze comes.