Scarlett Davis Crypto Media Editor-in-Chief April 2025
In mid-July, TeraWulf announced a $19 billion, 20-year lease with Anthropic. The market shrugged. Within weeks, the WGMI ETF, which tracks a basket of Bitcoin miners pivoting to AI, had fallen 34% from its peak. The disconnect between the headline numbers and the price action is not a mistake—it’s the market doing what it does best: pricing in the real risks behind the glossy narrative.
I’ve spent the last eight years watching miners fight for survival on thin hashprice margins. They have become expert at selling a commodity—computation—into a market that values it only as a function of Bitcoin price and difficulty. Now, they are trying to sell a different commodity: electricity itself. And the market is beginning to realize that this pivot is not a simple upgrade; it’s a structural gamble on a fragile assumption.
The Energy Landlord Thesis
Let’s start with the mechanics. Bitcoin miners own or control vast amounts of power capacity—hundreds of megawatts to multiple gigawatts—connected to the grid, often with existing substations and cooling infrastructure. For years, they used this power to run ASICs that solve SHA-256 hashes. But as Bitcoin’s block reward halves and hashprice compresses, the economics of operating at scale become brutal. The only way to survive is to either find cheaper power (a race to zero) or find a higher-value use for the power they already hold.
Enter the AI industry. Training large language models requires clusters of GPUs that can consume 100 MW or more per site. The leading AI labs—OpenAI, Anthropic, Google DeepMind—are desperate for power, and they are willing to sign long-term leases at rates that make mining look like a hobby. TeraWulf’s lease alone is worth more than the company’s entire market capitalization at the time of signing. CleanSpark followed with a $6.6 billion lease. Hut 8 was re-rated by Benchmark as a “power-first data center REIT.” The narrative writes itself: miners are becoming landlords of compute real estate.
But here is where my editor’s red pen comes out. The energy landlord thesis sounds clean, but it collapses under the weight of two words: execution risk.
The Hidden Assumptions
All of these lease valuations hinge on a single, unstated assumption: that the demand for AI training compute will remain persistently scarce for the next 10–20 years. Not just high—scarce. Because if compute becomes abundant—if anyone can spin up a cluster cheaply—then the premiums that AI labs are paying for guaranteed power will vanish. The miners’ long-term leases will become liabilities, not assets.
This is not a fringe risk. It is a central, structural fragility that the market is only now beginning to price.
Consider the open-source model ecosystem. In 2024, the Llama 3 model from Meta reached performance within striking distance of GPT-4. By early 2025, the Qwen 2.5 series from Alibaba and the Kimi K3 from Moonshot AI had matched or exceeded closed-source benchmarks on several key tasks. The implication is clear: if frontier-capable models become publicly available and can be run on commodity hardware, the need for exclusive, multi-billion-dollar training runs drops dramatically. The compute scarcity that underpins the miners’ pivot starts to erode.
And yet, the market priced miner stocks as if this risk didn’t exist. WGMI ETF doubled in six months. The narrative was intoxicating: “Miners are becoming the backbone of AI infrastructure.” But that narrative ignored the fundamental truth I’ve learned from watching every hype cycle since the ICO boom of 2017: when everyone agrees on the story, the story is already finished.
Truth over hype. Always.
When I audited the token distribution of EOS and Golem during those early ICO days, I found that the whitepapers promised decentralization but the token unlock schedules were designed to concentrate power. The same pattern repeats here: the miners’ AI pivot whitepaper looks solid, but the small print—the assumption of persistent compute scarcity—is the hidden unlock schedule. If that assumption breaks, the whole model revalues downward by an order of magnitude.
The Market’s Reality Check
The price action of July and August 2025 tells the story. After the TeraWulf and CleanSpark lease announcements, the sector initially rallied. Then, as traders digested the details, the slide began. WGMI dropped 34% from its high. Hut 8, despite the Benchmark upgrade, fell 18% in two weeks. TeraWulf itself gave back nearly half its post-lease gains.
Why? Because the market started doing its homework. It realized that:
- These leases are long-term, but they are not guaranteed revenue. The AI labs can cancel—or more likely, renegotiate—if their own funding or technology changes.
- The miners lack the operational expertise to run AI data centers. Running ASICs is one thing; managing GPU clusters with liquid cooling, high-speed interconnects, and 99.99% uptime SLAs is a completely different engineering discipline.
- The competition is not other miners. It’s established data center operators like Equinix, CoreSite, and Digital Realty, as well as hyperscalers like AWS and Microsoft who are building their own power capacity. Miners are late to the party and undercapitalized for the scale required.
I spoke with a traditional data center operator who declined to be named. He told me, “We have been doing this for 20 years. We have the supply chain, the cooling expertise, and the customer relationships. Miners have a substation and a dream. That’s not a moat.”
Noise filtered. Signal preserved.
The signal in all this noise is that the market is now differentiating between miners with real AI capabilities and those with just press releases. The first wave of euphoria lifted all boats. The second wave is separating the competent from the hopeful.
Take TeraWulf. They have a legitimate lease with a top-tier AI lab. They have hired experienced data center managers. They are spending capital on retrofitting their sites. But even they face execution risks: can they deliver the required power density? Can they maintain the network performance? Can they navigate the local permitting and grid interconnection timelines? The lease is a first step, not a finish line.
Compare that to smaller miners who have announced “AI partnerships” with little detail. One small-cap miner I follow issued a press release about “exploring AI compute opportunities” and saw its stock jump 40% in a day. Within a month, it had given back all those gains and more. Investors are learning to ask: “Where is the signed contract? What is the minimum commitment? Who is paying the upfront costs?”
The Contrarian Angle: Open Source Eats the Compute Premium
Now let’s dive into the contrarian angle that most analysts are missing. The bull case for miners relies on the idea that AI training compute will remain a scarce resource controlled by a handful of players. But open-source AI is a direct counterforce to that scarcity.
When Llama 3 was released, it became immediately available for anyone to fine-tune and deploy on their own hardware. The cost to train a model similar to GPT-4 dropped from an estimated $100 million to under $10 million for a determined team using open-source code and rented GPUs. The “compute barrier” that protected the value of exclusive GPU clusters is dissolving.
What happens if, within two years, a model that matches GPT-5 is released under an open license? The marginal value of a dedicated 100 MW training cluster plummets. AI labs will still need compute for inference and fine-tuning, but those workloads are less power-intensive and can be distributed across many smaller sites. The demand for multi-gigawatt “AI factories” becomes a niche, not the mainstream.
This is exactly the scenario that miners’ business model does not survive. A 20-year lease at premium rates assumes that the tenant will need that exact compute for the entire duration. But if technology shifts—if a model architecture emerges that requires less compute, or if inference becomes more efficient—the tenant may sub-lease or walk away. The miner is left with a custom-built facility that has little alternative use.
Trust is the only currency that matters.
I learned this lesson during the 2022 crash, when I sheltered my junior writers from the volatility by focusing on fundamental resilience rather than speculative hype. The same principle applies to evaluating miner stocks: trust can only be built through consistent execution, not through press releases. The miners that will survive are those that underpromise and overdeliver on their AI transitions. So far, we have seen plenty of promise and very little delivery.
The Execution Gap
Let’s be specific about the gap between promise and reality. A typical Bitcoin mining site operates with a power capacity of 50–200 MW. It uses air cooling for ASICs, which run 24/7 at a consistent temperature and humidity. The network requirements are minimal: the miners just need to connect to a mining pool over the internet. The staff are electricians and hardware technicians.
An AI data center of the same power capacity requires: - Liquid cooling (direct-to-chip or immersion) because GPUs generate 10–20x more heat per square foot than ASICs. - High-bandwidth, low-latency fiber connections to the internet and cloud providers, because AI training shards data across thousands of GPUs. - Specialized power distribution and backup systems to handle variable loads (GPU clusters can spike demand during training). - A team of network engineers, software engineers, and data center operators who understand CUDA, InfiniBand, and HPC cluster management.
Very few mining companies have these capabilities in-house. Those that are serious—like Hut 8, which hired a former Google data center executive—are investing heavily. But the majority are outsourcing the AI operation to partners, which eats into the margins and limits their value capture. The “REIT” analogy is misleading because a REIT owns real estate that has many potential tenants; a miner’s AI facility has only one tenant (the AI lab) and its value is tied to that specific relationship.
The Risk Matrix
Summarizing the risks I see from my vantage point:
- Technical risk (high): Miners lack AI infrastructure expertise. Retrofitting a mining site to meet AI standards is expensive and time-consuming. Any delay or cost overrun can destroy the lease economics.
- Market risk (extreme): The core assumption of persistent compute scarcity is fragile. Open-source AI models could drastically reduce demand for proprietary training compute, rendering the leases uneconomical.
- Counterparty risk (moderate): AI labs are venture-funded and may not have the cash flow to honor 20-year leases if their own funding cycles stall. The recent layoffs at several AI companies are a warning sign.
- Regulatory risk (moderate): Energy regulators may scrutinize the allocation of grid capacity to cryptocurrency-related entities, especially during grid stress. Some states have already imposed moratoriums on new mining operations.
- Valuation risk (high): Current stock prices already price in successful AI transitions for many miners. If execution falls short, the revaluation could be brutal—a classic Davis double-kill as both earnings and multiples contract.
The Narrative Cycle
I’ve been writing about crypto narratives for a decade. The pattern is always the same: a disruptive idea catches fire, capital floods in, a few early successes are celebrated, then the inevitable correction separates the real from the fake. The “miners as AI landlords” narrative is entering the correction phase.
The hook of the story—the $19 billion lease—was real, but it was also a peak. The market had already priced in the narrative before the details were known. Now, we are in the phase of scrutiny. Every quarterly earnings call will be a test. Every announcement of a new lease will be met with skepticism. The companies that can point to actual AI revenue on their income statements will survive; the rest will fade.
My Personal Take
I have been through enough cycles to know that the most dangerous phrase in crypto is “this time is different.” The miners’ AI pivot is real in the sense that power capacity is a real asset. But the valuation of that asset is being inflated by a narrative that may not hold. I have seen the same dynamic play out with DeFi summer yields, NFT floor prices, and Layer 2 token airdrops. The market always overestimates the speed of adoption and underestimates the difficulty of execution.
In my own work, I have always tried to be the voice that asks: “What is the hidden assumption? What happens if it fails?” That is the service I provide to my readers. Trust is built by being right over time, not by being exciting in the moment.
Takeaway: The Next Narrative
So where does this leave us? The miner-to-AI pivot is not dead, but it is entering a new phase. The next six months will reveal which companies have real operational capability and which are simply riding the narrative wave. The contrarian position is not to short the sector outright, but to focus on the miners that have the strongest balance sheets, the most specific contract terms, and the most credible management teams. Avoid the ones that announced a pivot but have no details.
Longer term, the real question is whether AI compute demand will remain centralized enough to sustain these leases. If open-source models continue to advance, the scarcity premium will disappear, and miners will be left with expensive, single-purpose infrastructure. If, instead, AI becomes even more compute-intensive (e.g., video generation, autonomous driving simulations), the demand could explode and justify the leases.
I don’t have the answer. No one does. But I know that the current market prices the most optimistic scenario. As an editor, my job is not to predict the future, but to identify the signals that matter. The signal I see today is that the market is waking up to risk. The next signal to watch will be the first quarterly AI revenue report from a major miner—and whether it meets or misses expectations.
Until then, treat the hype with skepticism. The code is cold, but the community is warm—and the community is starting to ask hard questions. That’s a healthy sign.