코인톡

Episode 232: The Energy Layer

We review @mariohsouto's essay on the orchestration layer between energy and AI compute.

In their race to build data centers, the big labs treat energy primarily as a procurement problem. This misses a bigger picture.

(Yeah I'm gonna hand this one to ChatGPT. Human commentary follows)

AI infrastructure is becoming a real-time optimization problem across energy, storage, models, memory, workload routing, agent state, verification, and cost.

It is not just deciding which GPU runs which prompt. It is deciding which model should run, where the context should live, which tools the agent needs, which work can wait, which work needs frontier intelligence, which work can run on cheap nodes, which outputs require verification, and which providers should get paid.

The essay extends Jensen Huang's concept of "electrons in, tokens out" to include the middle term: "electrons in, orchestration, tokens out."

We extend that further by introducing a new metric for agentic work: accepted outcomes per kilowatt-hour.

So our version is: "electrons in, orchestration, accepted outcomes out."

Bitcoin miners are the obvious players to pay attention to here.

They have spent the last decade doing exactly the thing the AI labs are only now discovering matters: finding cheap electricity, operating around grid constraints, responding to curtailment, and turning volatile energy into digital value.

That does not mean Bitcoin miners automatically become AI labs.

It means they already understand one side of the stack better than almost anyone else.

They understand power markets. They understand stranded energy. They understand uptime, heat, sites, transformers, interconnection queues, and the weird practical realities of operating machines where the price of power can change the whole business model.

The missing piece is the workload layer.

Bitcoin mining is simple in one important way: every hash is the same. AI work is not like that. Some jobs need to run immediately. Some can wait. Some need GPUs. Some can run on CPUs. Some need high memory. Some need verification. Some are only valuable if the output is accepted by the user.

That is where agentic inference changes the picture.

Agentic work is not just a chat response. It is tasks. It is code. It is research. It is data cleaning. It is benchmark runs. It is verification. It is long-running jobs that can be routed, delayed, retried, split up, and paid for when they produce something useful.

So the opportunity is not just: miners add GPUs.

The opportunity is: miners become part of an agentic compute market.

They bring cheap electrons and operational discipline. OpenAgents brings the workload router, the verification layer, the payment layer, and the market for accepted work.

This is the bridge from Bitcoin mining to AI.

Not abandoning Bitcoin.

Not pretending every mining site becomes a hyperscaler.

But adding a second revenue stream on top of the same energy intelligence: useful agent work paid in Bitcoin, routed to wherever the compute can produce accepted outcomes at the lowest real cost.

That is the energy layer as a market.

1. Cheap power in.
2. Energy-aware routing.
3. Verified agent work out.

---
"Hey Car do you know any soon to be major AI labs that are based in Texas?"

"I think there's one, it's called OpenAgents."

"OpenAgents?! Hang on, and they keep talking about Texas Texas Texas and ERCOT in here. ... Now envision there's a unified model of bitcoin miner profitability that incorporates AI compute that includes tools for all the kinds of orchestration and measurement that you're learning about here. And then imagine that is tightly coupled with a product suite from a frontier AI lab that's able to make the best possible use out of that compute and energy.

"Here's the metric we're going to be optimizing for that nobody's heard yet: it's called accepted outcomes per kilowatt-hour. Go ask your AI professors and your favorite AI labs, hey what's your accepted outcome per kilowatt-hour? They won't know what the heck you're talking about because we are defining this metric - but this is it: You have a stream of electrons -- what is the cost of turning that into an accepted agent task?

"Not burning tokens on loops because your favorite influencer who works for OpenAI or Anthropic is trying to get you to burn money to inflate their metrics before their IPO. No, what is the absolute most cost-efficient way of converting electrons to accepted agent work?

"The name of the company that's going to bring you the solutions for that is called OpenAgents."