In 2021, a Bitcoin miner in rural Quebec was a punchline — racks of ASICs burning hydroelectric power to produce a speculative token with no intrinsic use. In 2026, that same warehouse, retrofitted with H100s and direct-to-chip cooling, produces inference tokens for a Fortune 500 client at a contracted rate. The power bill looks identical. The physics is identical. Even the business model — convert electricity into a digital unit of account and sell it — is identical. So here's the uncomfortable question: was crypto mining ever actually a different industry, or just an early, badly-branded version of AI infrastructure?
The parallel is structural, not rhetorical
Strip away the ideology and both operations reduce to the same equation: joules in, tokens out, margin per unit of energy. Bitcoin's unit is a hash; AI's unit is a token. Both are produced in facilities whose site selection logic is pure energy arbitrage — cheap stranded power, cold climate, grid capacity. I've sat in site-selection reviews for data center projects in North America and the criteria haven't changed since the mining boom. The people negotiating the energy supplier interconnects just wear different lanyards now.
Nvidia's own framing leans into this openly: revenue equals tokens per watt times gigawatts available. That is, verbatim, how a mining CFO describes hashrate economics. When your hardware vendor starts talking like a mining pool operator, the analogy has stopped being an analogy.
Even the market structure rhymes. Core Scientific, Marathon, Bitfarms — half the listed miners have pivoted capacity toward AI hosting since 2024, and why wouldn't they? A power-connected shell with cooling and land is the scarce asset. What runs inside it is a leasing decision.
But here's where the analogy breaks — and it's instructive
Bitcoin's token is the product. The hash is a means; the coin is the end, and its value is set by a market that has agreed, collectively and somewhat mystically, that it has value. The AI token is the opposite: it's a means, never an end. Nobody hoards inference. Nobody speculates on GPT-5 output tokens appreciating. The token is more like a kilowatt-hour than a coin — consumed the instant it's produced, its value entirely downstream.
That difference matters for the economics. Bitcoin mining difficulty adjusts upward forever; your revenue per joule is designed to decay. AI inference doesn't need a difficulty algorithm — the market built one organically. Price per token at fixed quality has been falling 5–10× per year, faster than any halving schedule Satoshi ever wrote. Both industries, in other words, are on a treadmill. The AI treadmill just runs on Moore's Law and competition instead of protocol.
And this is where I take a position: the AI version is the healthier one. Deflation in Bitcoin mining is artificial scarcity propping up an asset. Deflation in inference is real productivity deflating a cost. One is monetary policy. The other is industrial progress. I've watched both up close — I know which spreadsheet I'd rather defend in a board meeting.
The HPC twist
There's a third actor in this genealogy that both camps prefer to ignore: classical HPC. Weather models, crash simulation, Nuclear reactor neutronics — supercomputing was converting megawatts into valuable numerical output decades before anyone minted a coin. Neither Bitcoin nor AI invented the power-to-computation economy. They financialized it, each in their own way: one by making the output a speculative asset, the other by making it a metered utility.
Arguably the data center industry is converging back to its HPC ancestor — but with a spot market attached. When inference capacity gets traded the way cloud reserved instances already are, with futures curves on token delivery, the transformation will be complete: the token as a commodity, with hedgers, speculators, and a contango curve.
The forward look
Watch the miners. They're the canary. If AI inference margins follow the mining playbook, the next phase is consolidation: whoever controls cheap power and grid interconnects controls production, and the token becomes a wholesale commodity sold by a handful of energy-advantaged producers. The question nobody has answered is whether token prices can fall another 100× without the industry hitting the same wall mining hit — a point where the unit of production is worth less than the electricity it consumes. Bitcoin solved that with price appreciation. What's AI's answer? Genuine demand, one hopes. But I've seen enough optimistic utilization curves in pitch decks to know that hope is not a load forecast.
Share: