The numbers are out. Kimi K3, the latest model from Moonshot AI, costs $0.94 per task. That is 71% more expensive than GPT-5.6 Terra's $0.55. Even against GPT-5.6 Sol's $1.04, it barely competes.
Investor Gavin Baker from Atreides Management calls this a potential turning point for the AI industry. But not for the reasons most retail think. He argues that model-layer profits are about to get squeezed. The real winners? Power, chips, data centers, cloud—and yes, decentralized compute networks built on blockchain.
I have been watching this divergence for months. In my copy trading community, members ask me daily: "Should I buy AI tokens?" The answer requires understanding where the value actually flows.
Context: Kimi K3 as a Catalyst, Not a King
Kimi K3 is a serious challenger. Its benchmark performance nears GPT-4o and Claude 3.5 Opus, but its inference cost is a liability. The model requires more compute to achieve the same output. That inefficiency is baked into its architecture.
Baker's core thesis: "Token efficiency" is the bottleneck. Without it, no model company wins the profit game. He sees a future where 2-3 monopolies collapse under competition. Open models—like Llama or Mistral—will drive token costs down further. That is the real turning point.
But here is what Baker does not say: his view is an investment thesis. His fund likely avoids AI model tokens and piles into infrastructure. The same logic applies to crypto.
Core: Why Crypto Infrastructure Wins
The math is brutal. At $0.94 per task, K3 burns cash. To be viable, it needs that number under $0.40. That demands better hardware, cheaper electricity, or algorithmic leaps. All three rely on physical infrastructure—chips, cooling, power grids.
Crypto enters here. Decentralized compute networks like Render, Akash, and io.net offer GPU time at market rates. They thrive when AI demand grows. They also benefit when model providers race to cut costs—because cheaper compute becomes a competitive weapon.
Look at the Terra/Luna collapse in 2022. I ran a post-mortem study group. We learned one lesson: don't trust the center. Centralized models like K3 rely on single points of failure—closed data, closed hardware, closed governance. Open models and decentralized compute spread risk.
Baker admits the same. He explicitly says open models are the true turning point. That is music to crypto ears. Open-weight models mean anyone can run inference on decentralized infrastructure. No gatekeepers. No API rent-seeking.
Contrarian: Retail Overvalues AI Model Tokens
Most traders chase the shiny model. They buy tokens of AI platforms that host proprietary models. They trust the charts, not the hands.
But the hands that control the hardware hold the real power. NVIDIA's market cap proves it. In crypto, the equivalent is GPU-backed tokens. They provide the raw material for AI—computational power.
Baker's logic flips the narrative. If AI models become cheap commodities, only the network effects and brand of OpenAI/Anthropic persist. Their tokens? Vulnerable. Meanwhile, decentralized compute networks get higher utilization and more demand as models proliferate.
I see it in my own copy trading data. When model competition heats up, infrastructure tokens see correlated volume spikes. Retail chases the app; the smart money buys the pick and shovel.
Takeaway: Watch for Open Model Releases
The next signal is simple: when a token-efficient open model drops below $0.30 per task, the game changes. That will be the inflection point for decentralized compute adoption. Until then, Kimi K3 is a warning—a sign that model profits are finite, but infrastructure need is infinite.
Trust the hands, not just the charts. Community first, coins second. Always.
Follow the people, follow the profit.