The thing that was supposed to decentralize everything is centralizing, while the thing Big Tech runs is starting to scatter.
The Summary
- Bitcoin mining is trending toward centralization while AI computation may actually decentralize, according to new research on infrastructure economics.
- Edge computing and open-source AI models could reduce dependence on corporate data centers, reversing assumptions about who controls AI inference.
- The divergence reveals something uncomfortable: economic incentives matter more than ideology when it comes to how distributed systems actually distribute.
The Signal
Bitcoin mining's drift toward centralization isn't news to anyone watching hash rate distribution. But the fact that AI might be moving the opposite direction is. The research points to edge computing as the driver, the idea that inference workloads can run on devices you already own instead of round-tripping to Google's servers.
This isn't about ideology. It's about economics. Bitcoin mining centralizes because economies of scale matter when you're burning electricity to solve puzzles. Bigger operations get cheaper power, better cooling, first access to new hardware. The small miner gets priced out. The pool gets bigger.
"The thing that was supposed to resist centralization is following the same path as every other industrial process: toward consolidation."
AI is different because edge computing changes the cost structure. You don't need to compete on electricity or chip fabrication to run inference on an open-source model. You need:
- A device with a decent chip (phone, laptop, edge server)
- A model small enough to fit in local memory
- Enough battery or power to run it without melting the case
Open-source models make this viable. When Llama, Mistral, or Qwen release weights anyone can download, the barrier to running AI drops from "build a data center" to "download a file." Training still centralizes around whoever has the compute budget. But inference, the part users actually interact with, can scatter.
This matters because inference is where the control is. If your AI agent has to phone home to OpenAI or Anthropic every time it thinks, those companies see everything. They decide what gets answered and how. They change the terms. Edge inference breaks that dependency. Your agent runs on your hardware, using weights you downloaded, processing data that never leaves your network.
The Implication
Watch where the open-source AI community focuses next. If edge inference becomes genuinely competitive with cloud inference on cost and capability, you'll see a wave of products that don't need permission from hyperscalers to exist. That's the Web4 pattern: agents that build and run independently, not as tenants in someone else's infrastructure.
For Bitcoin, the centralization trend suggests the network's security assumptions need updating. The narrative that thousands of independent miners secure the chain looks weaker every quarter. Either the economics shift or the ideology does.