Quick note: We are at Yotta this week from 9/28 to 9/30. Let’s meet up if you are too! Just respond to this email or book time here.

My cryptocurrency passion started in college in 2017. I taught myself Solidity (the Ethereum programming language) outside my computer science classes, built an NFT game tied to renewable energy stocks, and spent nights and weekends deploying dApps (decentralized applications) on Ethereum. I once finished a work week at Apple, flew to Denver for a crypto hackathon, slept in a hostel, and flew back in time for Monday. Almost a decade later, I find myself in another computational wave: AI infrastructure. And many of the names and themes are strangely familiar.

Some of the biggest names in data centers today — CoreWeave, Crusoe, IREN and Hut 8 — all emerged from crypto. Why?

What crypto miners learned before everyone else

What do crypto mining facilities and AI data centers have in common? Power, hardware, grid interconnection, real estate.

Bitcoin miners earn newly minted BTC plus transaction fees for every block they win. They can't set the price, so their main lever is the cost of power.

That made them an unusually qualified operator class: they learned how to secure hundreds of MWs, work with utilities, build substations and data centers, and procure enormous hardware fleets.

How AI changed the economics

TLDR: Crypto winter was the push; AI economics were the pull, and existing power infrastructure was the bridge.

The crypto winter happened around 2022, and I remember experiencing this viscerally as a builder in the space. The UST death spiral (Terra) happened in the spring. And in November, FTX collapsed. At that time, I was working on a crypto protocol for demand response through virtual power plants. But winter was coming, so we put the crypto idea on hold.

And 2022 was also when the data center electricity curve started accelerating.

Let’s take a deeper dive into the incentives:

1. Crypto mining became a much harsher business economically. Core Scientific’s 2022 average Bitcoin price fell 41%, power costs rose $136.5 million, and it filed Chapter 11. Bitcoin also has a structural squeeze built into the protocol. The April 2024 halving cut the block subsidy from 6.25 BTC to 3.125 BTC, while increasing the network hashrate/difficulty means every unit of mining capacity competes for a smaller slice of the reward pool.

2. For Ethereum GPU miners, their original workload disappeared. Ethereum's Merge on September 15, 2022 eliminated proof-of-work mining in favor of proof-of-stake, and cut Ethereum's energy consumption by roughly 99.95%.

But some crypto-era providers were already addressing AI workloads. CoreWeave had launched its cloud platform by 2020, two years before the Merge and three years before the generative-AI frenzy.

Crusoe Cloud began operating in mid-2022, initially with A100 GPUs, and says demand immediately exceeded supply. That was before ChatGPT launched.

So the best operators weren't merely fleeing crypto. They had already discovered that general-purpose compute could be a valuable use of their infrastructure. And AI radically increased the revenue potential of a powered site.

In FY2026, IREN’s Bitcoin mining business still generated $578 million versus only $129 million from AI Cloud. But there’s $4 billion of contracted ARR for 2026 AI capacity, and it’s largely sold out.

Essentially, 2022–2026 created the perfect storm:

Ethereum mining is obsolete > Bitcoin mining becomes structurally more competitive > AI demand explodes, and power becomes the gating constraint.

The standouts

Here are some of the biggest stories in the crypto → AI compute transition:

• CoreWeave: Started in September 2017 as The Atlantic Crypto Corporation, renamed CoreWeave in 2019. In 2016, before the company formally existed, they bought a GPU, put it on a pool table in their Lower Manhattan office, and started mining Ethereum. Because Ethereum ran on GPUs rather than Bitcoin ASICs, their machines had uses well beyond crypto. CoreWeave Revenue went from $229 million in 2023 to $5.1 billion in 2025.

• Crusoe: In 2018, the founders Chase Lochmiller and Cully Cavness saw an ugly oilfield problem: natural gas produced alongside oil was often flared because there was no economical way to transport it. Crusoe built modular data centers next to oil wells and turned stranded gas into power to mine Bitcoin. In 2025, it sold its Bitcoin mining business to NYDIG to go all-in on AI.

• IREN: Founded in 2018 by Australian brothers Daniel and Will Roberts as a low-cost Bitcoin producer running on renewable power. It launched AI Cloud in 2024 and, during FY2026, began decommissioning Bitcoin mining hardware to reallocate its power toward AI.

• Hut 8: Formed in November 2023 by merging Canada’s Hut 8 Mining with US Bitcoin Corp. In 2025, Hut 8 signed a 15-year, $7.0 billion lease with Fluidstack for 245 MW of AI capacity, backstopped by Google, as part of a broader partnership involving Anthropic.

Also see: MARA and TeraWulf.

Where BTC mining and AI infrastructure diverge

Bitcoin uses ASICs: application-specific integrated circuits optimized for hashing. That specialization is the point, and mining is a very specific, unusually forgiving computational workload. If a miner loses power for an hour, it loses an hour of potential BTC production and can often curtail when grid prices spike and resume later.

But AI customers demand much more uptime and sophisticated infrastructure. Crypto mines lack redundant power distribution, backup generation, and sophisticated cooling. AI requires extreme rack densities, liquid cooling, high-performance networking, and greater resiliency.

Networking/fiber is a huge difference because AI workloads require GPUs to communicate with each other. For example, NVIDIA’s rack-scale architecture combines GPUs, NVLink switches, high-speed networking, and liquid cooling as one integrated system.

The pattern: One boom trains the workforce for the next

Crypto created an economic opportunity to solve problems around stranded power, flexible load, specialized hardware fleets, and industrial-scale compute before AI became the next boom.

That said, AI infrastructure comes with new challenges we haven’t seen before:

• Training can tolerate remoteness, but inference workloads care more about latency and connectivity.

• Facilities designed today have to accommodate rapidly increasing rack densities and cooling demands. New chips and systems might even outrun facilities currently under construction.

On the personal front: Since 2017, I've been thinking about smart contracts, tokens, and Proof of Work. Today, with Spark, I spend most of my time thinking about power generation, data centers, and how quickly new infrastructure can actually get approved. My appreciation for crypto hasn’t changed. Which is probably why my 2018 project, CryptoRenewables, feels like a full-circle moment.

But underneath both eras is the same physical truth: leaps in computation demand abundant energy.

The chips changed. The workloads changed. The tokenomics changed. The enduring asset is the megawatt.

Developing AI compute infrastructure (with a crypto past or not)? Let’s trade notes on permitting, development, and navigating community opposition.