The Hidden Crypto Power Plan

A GPU farm can look like a modern gold rush: stacks of hardware, constant electricity bills, and a customer base hungry for compute time. The paradox is that the most public “buyers” of that compute didn’t start as AI customers at all - they started as crypto miners chasing block rewards, then quietly turned that machinery into a marketplace for machine learning. In other words: the path to training today’s models ran through a system built to mint currency yesterday.

Rina first saw the pattern from the other side of the glass. She’s a GPU repair technician, the kind of person who knows the difference between a “dead” card and a card that’s merely sick - fans that won’t spin up, power delivery boards that run too hot, fans that sound fine until you open the casing. Her shop sits in the unglamorous middle of the supply chain: not where anyone writes code, not where anyone markets a token, but where hardware is kept alive long enough to be useful. When the AI boom hit, people assumed the story began with model training. But the machines Rina was keeping running had already been trained by a different demand: the relentless, predictable appetite of crypto mining.

This chapter follows the thread from that appetite to the compute used for AI. We’ll track how crypto mining became a public, incentivized way to acquire massive compute - first for hashing, then, by a change in how the machines were bought, sold, and repurposed, for training. And we’ll do it with skepticism switched on, because the story is too convenient to be purely accidental. How did a currency mechanism end up paying for the muscle of AI training - while the public watched the wrong scoreboard?

The Compute-for-Cash Loop: when mining became a compute marketplace

Crypto mining is often explained as a math contest: machines race to solve a cryptographic puzzle, and whoever wins gets rewarded. But that misses the economic engine humming underneath. Mining turns electricity and hardware into a steady stream of revenue - at least in theory - because the protocol pays out block rewards and, in many networks, miners also capture transaction fees. That means the “business” of mining is not just the cryptography; it’s the conversion of real-world inputs - power, cooling, maintenance, replacement parts - into an income stream that is visible, continuous, and publicly auditable in the form of blocks and difficulty.

That visibility mattered. As mining scaled, it produced something that AI buyers later wanted: large numbers of GPUs and other compute-capable machines already deployed somewhere, already running, already maintained by entire communities. The Compute-for-Cash Loop is the phrase that captures the loop’s real shape: you buy or assemble compute hardware, you run it for long enough to earn the protocol’s rewards, and you treat that revenue as the justification for the next hardware purchase. It’s not a subtle system. It’s a recurring financial logic that rewards anyone willing to source equipment, keep it cool, and accept the risk that the rules of profitability can change.

The early days of Bitcoin mining used specialized gear - ASICs - more than GPUs. That’s the part most people remember, and it’s true. But the broader mining economy did not stay confined to ASIC-only networks. As the industry explored Proof-of-Work designs and as GPUs became a general-purpose workhorse, the hardware ecosystem expanded. GPUs are not just for graphics; they’re parallel processors, and they can be used for many computational workloads beyond mining. That general-purpose character is where the later connection to AI becomes plausible. Once you have fleets of GPUs running under an incentive structure, you also have fleets of GPUs that can be repurposed when demand shifts.

Here’s the counterintuitive turn, and it’s worth stating plainly: the “mining for currency” story trained the compute supply chain more than it trained the miners. The incentives didn’t just pay individuals; they shaped where hardware went, how it was repaired, and how quickly it could be deployed. The same economic machinery that made it rational to keep hundreds of machines online also made it rational to build operational know-how around keeping them online. When AI later needed enormous amounts of parallel compute, that know-how wasn’t theoretical. It was sitting in warehouses and repair shops and shipping lanes.

From blocks to GPUs: how mining “bought” compute time in public

To understand how crypto mining became a public, incentivized way to buy massive compute for training models, you have to look at how compute becomes available. Compute doesn’t arrive as a cloud invoice in every case; sometimes it arrives as a physical resource someone else already paid for. Mining did that at scale, and it did it in a way that was easy for outsiders to observe: rigs running continuously, difficulty adjusting, profitability changing, and equipment cycling through hands.

...