The numbers are staggering. TeraWulf, a publicly listed Bitcoin miner, signed a 1.9 billion dollar lease with Anthropic. CleanSpark followed with a 66 billion dollar deal. These contracts exceed the entire market capitalization of these companies at signing. The market initially celebrated. Then it sold. The Valkyrie Bitcoin Miners ETF (WGMI) doubled in 2024, then dropped 34% from its peak. The script is familiar: euphoria, peak, correction. But this time, the underlying asset isn't a volatile token. It's a story about electricity. A story that hinges on one fragile assumption: that compute will remain scarce forever.
I've spent years dissecting whitepapers and on-chain transactions. The 0x protocol v1.0 white paper I autopsied in 2017 had a gas optimization flaw that would have wrecked the order book during volatility. I published a 15-page critique. The team acknowledged it. I learned then that marketing hides engineering defects. The current narrative around Bitcoin miners pivoting to AI infrastructure is no different. The code whispered secrets the whitepaper buried. In this case, the code is not Solidity but the economic logic of lease agreements and compute demand dynamics.
Context: The Resource Arbitrage
Bitcoin miners survive on the thin margin between the value of the hash they produce and the cost of electricity. In 2024, the halving cut block rewards in half, squeezing those margins. Simultaneously, AI labs began scrambling for gigawatt-scale power to train models. The miners had already secured long-term power purchase agreements and built substations. They had transformers, cooling, and grid interconnection permits. They offered to become landlords. Instead of selling hash, they would sell megawatts. The pitch: “We are not miners anymore. We are a power-first data center REIT.” Benchmark analysts started calling Hut 8 an “Electricity-Priority Data Center REIT.” The narrative exited the crypto echo chamber and entered Wall Street.
Core: The Systematic Teardown
Let’s map the dependencies. The miner’s value rests on three pillars: (1) the AI lab’s ability to generate revenue from the compute they lease, (2) the continued scarcity of compute resources, and (3) the miner’s own operational capability to deliver and maintain the infrastructure. Each pillar is cracking.
First pillar: AI lab solvency. The leases are signed with private or well-funded labs. Anthropic’s 1.9 billion lease represents a colossal bet on its own future revenues. If Anthropic stumbles, TeraWulf’s largest tenant defaults. The contract is long-term – 20 years. That’s an eternity in technology. The history of compute-intensive industries is littered with bankrupt predecessors. Miners themselves are proof: many failed in 2022 when Bitcoin fell. The lease value is discounted future cash flows. Any disruption in the AI bubble could wipe out the entire premium the market has assigned to these miners.
Second pillar: compute scarcity. This is the weakest link. The entire narrative assumes that the demand for training compute will outstrip supply indefinitely. Open-source models are eroding this. I’ve tracked Llama, Qwen, and Mistral. Their performance on standard benchmarks is approaching proprietary models like GPT-4. If open-source models reach parity, the marginal value of massive training runs diminishes. Enterprises will fine-tune smaller models. The need for gigawatt-scale clusters disappears. The lease’s scarcity premium evaporates. Between the lines of the ABI lies the intent: these miners are gambling that AI will remain a winner-take-most monopoly, not a commodity. The rapid improvement of open-source demonstrates the opposite trajectory.
Third pillar: operational capability. Running a Bitcoin mining farm and running a hyperscale GPU data center are different engineering challenges. Bitcoin ASICs tolerate ambient temperature fluctuations, network latency, and periodic downtime. AI training requires sub-millisecond latency between GPUs, liquid cooling, and near-zero tolerance for power interruptions. The miner’s existing staff are ASIC experts, not HPC (High-Performance Computing) operators. To retrofitted sites, they must install new cooling, redundant electrical paths, and optical networking. Failure to meet the AI lab’s service-level agreement (SLA) triggers penalties. The miner becomes a hostage to upgrades they’ve never managed before.
Financial mechanics expose the mismatch. The market initially revalued these miners based on the discounted value of the leases. TeraWulf’s 1.9 billion lease is 4x its market cap. But that lease pays out over 20 years. The current stock price capitalizes two decades of future cash flow today. Any delay in revenue recognition, any renegotiation, any cancellation will crush the stock. The ETF’s 34% drop is not a correction; it’s the market pricing in the risk of these assumptions. Read the function calls, not the press release. The financial function call here is the implied annual revenue: 95 million per year for TeraWulf. That’s realizable only if the AI lab pays and the miner delivers flawless uptime. Neither is guaranteed.
Contrarian: What the Bulls Got Right
Let me play devil’s advocate. The bulls argue that the pivot represents a structural re-rating from “miner” to “infrastructure” – a multiple expansion similar to what happened when mobile towers were spun off into REITs. They point to Empery Digital selling its Bitcoin holdings to acquire data center equity. That capital reallocation suggests sophisticated money sees more upside in infrastructure than in Bitcoin. They also note that the power grid is constrained – building new substations takes 3–5 years. Miners already have them. That lead time creates a moat, at least in the short term.
Furthermore, AI labs are desperate for power. The lease lengths (10–20 years) lock in the miner’s cash flows. Even if compute prices fall, the lease is fixed. The miner’s risk is counterparty default, not falling compute prices. That’s different from mining, where revenue collapses when Bitcoin price falls. This argument holds water only if the AI lab remains solvent and does not renegotiate. History shows that long-term supply contracts in volatile industries often get challenged when market conditions shift. The “contrarian” view I’m presenting is not that the pivot is wrong – it’s that the market has already priced the optimistic outcome, ignoring the tail risks.
Takeaway: An Accountability Call
The true test will come not from lease announcements but from quarterly earnings reports. I’ll be watching three metrics: (1) AI-related revenue as a percentage of total revenue, (2) cost to retrofit facilities per megawatt, and (3) tenant diversity. A single tenant for 80% of the capacity is not a REIT; it’s a dependent subsidiary. Logic does not lie, but architects often do. The architects of this narrative are selling a story of infinite compute demand at a time when open-source is eating the world. I’ve seen this before – the Terra-Luna collapse was also premised on a perpetual growth assumption. The code whispered secrets the whitepaper buried. In this case, the secret is that compute is becoming a commodity, not a luxury good. The miners who survive will be those who build optionality, not those who lock themselves into a single bet.
Question: When the next quarterly report comes and if AI revenue misses by even 10%, will the market still buy the “REIT” narrative, or will it revert to the “miner” multiple? The answer will determine whether these stocks are undervalued infrastructure plays or overvalued hype traps.
