WGMI, the Valkyrie Bitcoin Miners ETF, is down 34% from its June peak. Over the same period, TeraWulf announced a $19 billion lease with Anthropic — a number that exceeds its own market cap. The stock didn't rally; it drifted. This is not a sell-the-news reaction. It is something more structural: a market beginning to doubt whether the miner-to-AI-hypothesis is a sustainable re-rating or a leveraged bet on a fragile assumption.
Let me rewind. For years, bitcoin miners lived on the spread between block rewards and electricity costs. Revenue was denominated in bitcoin, costs in fiat. Then, in early 2024, the AI industry’s insatiable demand for compute hit the headlines. Miners realized their existing infrastructure — massive power capacity, already-built data centers, and grid interconnection — was exactly what AI labs needed. The pivot was born: stop mining, start renting.
Today, the sector is crowded with narratives. TeraWulf signed a 20-year lease with Anthropic for 150 megawatts. CleanSpark inked a $6.6 billion deal. Hut 8 was upgraded by Benchmark as a “power-first data center REIT.” Archer analysts began calling them “AI landlords” — a metaphor that sounds convincing until you stress-test the lease economics.
But the market is no longer buying every story. WGMI’s decline, the dispersion among individual miner stocks, and the growing skepticism on earnings calls all point to a shift from narrative-driven euphoria to fundamentals-driven scrutiny. The core question is no longer “Will miners pivot?” but “Can they deliver?
Let me offer a dissection from four angles: the scarcity-of-compute assumption, the valuation gap, the execution risks, and the externality that could invalidate the whole thesis.
The Scarcity-of-Compute Assumption: A House Built on Sand?
The entire miner-to-AI bet rests on one premise: compute for training frontier models will remain scarce for the next decade. If that premise holds, the 20-year leases at $8–10 per kWh make sense. If not, those leases become stranded assets.
Open-source models are the elephant in the room. In the past six months, Llama 3.1, Qwen 2.5, and Kimi K3 all closed the gap with GPT-4. Rudy’s 70B outperformed GPT-4 on several benchmarks. The cost of training a 175B-parameter model fell from roughly $100 million in 2023 to under $5 million today. If open models continue to improve, the value of proprietary training compute declines. Why rent 150 MW to train a model that can be replicated with a fraction of that compute?
This is not a distant risk. It is happening now. DeepMind’s open-sourcing of AlphaFold and Meta’s Llama family are already compressing the demand curve for custom training clusters. Miners are betting that the scaling laws of transformers will hold and that chips will remain the bottleneck. But scaling laws are not laws of nature; they are empirical observations. An algorithmic breakthrough — like liquid networks or sparse attention — could reduce training requirements by an order of magnitude.
I have seen this playbook before. In 2018, during the ICO winter, I audited smart contracts of three failed projects. The common flaw was a vesting schedule that assumed continuous token demand. When demand collapsed, the contracts became insolvent. The parallel is uncomfortable: the miner lease contracts assume continuous compute demand. If the demand curve flattens, the “landlord” is left with empty floors and a power bill.
The Valuation Gap: Hashprice vs. AFFO
Wall Street still values most miners based on hashprice — the revenue per terahash per day. A typical miner trades at 10–12x EV/hashrate. That is a commodity play. But if a miner pivots to AI, it should be valued as a data center REIT, on AFFO (adjusted funds from operations), which implies multiples of 20–25x forward revenue.
This is the “re-rating trade” that every bull pitch outlines. Buy TeraWulf at $3.50 because market is pricing it as a miner, but it will eventually trade as a REIT. The argument is seductive, and in theory, correct.
Empery Digital, a $2 billion macro fund, just sold its entire bitcoin position to buy miner equity — explicitly betting on this re-rating. When a sophisticated capital allocator makes such a switch, it signals conviction in the structural thesis. But it also signals the market has already priced in much of the transition. WGMI’s peak-to-current decline suggests the re-rating has already partially debased.
The risk is twofold: first, the re-rating window may close faster than execution can catch up. If a miner misses a quarterly AI revenue target by 20%, the multiple snaps back to hashprice. Second, the current prices may have embedded the assumption of a full REIT multiple, leaving no margin for error. Hut 8 trades at 18x forward revenue — already near the REIT range. Any disappointment could trigger a reversion.
From my experience modeling institutional flows during the ETF launch in 2024, I know that institutional capital does not buy stories. It buys cash flows that can be discounted. Miners are selling stories. The conversion requires auditable P&L line items under “AI Infrastructure Services.” Until that appears, the valuation gap remains a narrative, not a cash flow.
Execution Risks: From ASICs to GPUs — A Different Beast
Managing a bitcoin mine is about maintaining uptime on ASICs — relatively simple machines with fixed power profiles. Transitioning to a GPU cluster for AI means managing cooling, networking, latency, and software stacks. It is a completely different technical operation.
Most miner CTOs have spent their careers optimizing for energy cost per hash. They have zero experience with Nvidia GPU cluster orchestration, InfiniBand networking, or Kubernetes scheduling. Hiring the right team is possible but takes time — and the AI labor market is tight. This operational gap is not yet priced into the stocks.
During DeFi Summer 2020, I built a Python model for liquidity provision on Uniswap v2. I discovered that impermanent loss models derived from traditional finance did not apply well to volatile crypto assets. The lesson: using the wrong model leads to capital destruction. Miners using a “mining operations” model for AI infrastructure may face similar blind spots — like underestimating cooling costs or overestimating uptime under variable compute loads.
If a miner fails to meet the service level agreement with Anthropic, the penalty can be severe. The lease is not a simple rental; it often includes clauses requiring 99.9% uptime and guaranteed power delivery. A single failure could trigger liquidated damages that wipe out years of profit.
The Blind Spot No One Talks About
Here is the contrarian angle: the real danger is not that the miner fails to execute, but that the entire “AI compute scarcity” thesis is a self-limiting prophecy. If every miner pivots to AI and adds gigawatts of capacity, the market will become oversupplied. The first-mover advantage vanishes. The scarcity that justified the lease premium disappears.
We are already seeing signs: multiple traditional data center operators (Equinix, Digital Realty) are expanding their AI-ready capacity. Energy companies like Constellation Energy are directly negotiating with AI labs. The miner’s only advantage — existing power permits — is eroding as regulators fast-track AI campuses.
Moreover, the AI industry itself is cyclical. Venture capital funding for AI startups peaked in 2024 Q1 and is now declining. If funding dries up, the downstream demand for training compute will collapse. Miners with long-term leases will be locked into selling compute into a glut.
The market is beginning to price this risk. Over the past seven days, the dispersion between top-tier miners (with confirmed leases) and second-tier miners (just promises) has widened by 40%. WGMI’s fall masks a divergence: some stocks are flat, others down 60%. Liquidity is leaving the weak hands.
What to Watch Next
The next two quarters are pivotal. Q3 earnings calls will reveal whether AI revenue appears on miner balance sheets. I expect most will show negligible AI revenue — because the leases start in 2025, not now. The market’s patience is limited.
Three signals will determine the trajectory:
- Open-source model performance: If a fully open model matches GPT-5 within six months, the scarcity thesis breaks. Track benchmark releases.
- Lease commission dates: TeraWulf must bring its 150 MW site online by mid-2025. Any delay will be punished.
- Regulatory greenlights: The FTC and DOE are both reviewing large-scale AI compute facilities. Delays or stricter environmental rules could stall the entire pivot.
Final Takeaway
The bitcoin miner pivot to AI is not a revolution. It is a high-leverage trade on a single assumption: that compute demand will remain structurally scarce. The assumption is plausible but fragile. The market is now in a transitional phase — from narrative-driven buying to fundamentals-driven differentiation. The winners will be those who can demonstrate real AI revenue and operational excellence. The losers will revert to their commodity multiple.
Tracing the fault lines before the quake hits. Code never lies, but it does omit. Liquidity is just patience disguised as capital.
— This analysis is based on public data and my own quantitative modeling background. Positions may change as new information surfaces.
