Hashprice sits at $31.8 per PH/s. That is a 50% decline from last July. Network hashrate dropped 21% from its peak of 1.14 ZH/s to 900 EH/s. Miners are shutting down rigs. Yet TerraWulf, IREN, and Cipher stock prices have more than doubled over the past year. Meanwhile, MARA, a legacy miner with a massive Bitcoin treasury, is down 40% in the same period. The market is not confused. It is reading a different balance sheet. The divergence is not about Bitcoin's price. It is about who gets to sell compute to AI labs and who is stuck mining an asset that just halved its block subsidy and saw its hashprice cut in half. This is a structural shift, not a narrative pivot. Check the math, not the roadmap.
Bitcoin miners occupy a unique position in the energy and compute stack. They build large-scale facilities with high-voltage power connections, industrial cooling, and fiber optics. They negotiate long-term power purchase agreements at wholesale rates. They operate fleets of ASICs with 24/7 uptime. For a decade, the only output was Bitcoin. Now, the same infrastructure—power, land, cooling, network—can be used to run GPU clusters for AI inference or training. The pivot is not a pivot. It is a balance sheet expansion. The asset class is the electricity contract. The compute load is the output. Whether that load is SHA-256 hashing or a large language model inference matters only for the hardware and the customer. The operational DNA is the same: manage power, heat, and uptime.
But the market is pricing the two outcomes very differently. Pure-play Bitcoin miners today trade at an average EV/EBITDA multiple of 5.9x. Miners with signed AI/HPC contracts trade at 12.3x. That is a 2.1x premium. The aggregate contract value for AI/HPC deals across the sector is now over $70 billion. Riot Platforms alone signed a 20-year, $9.1 billion deal with Anthropic. These are not option contracts. They are revenue commitments with milestones. The market is discounting those future cash flows today. The question is whether the discount rate is appropriate. I have spent years auditing protocol-level financial models—from Bancor V2's weighted constant product formula to Celestia's data availability sampling. The same lesson applies here: the math behind the multiple must account for execution risk, capital expenditure, and contract quality.
The Core Technical Analysis: Power Assets and the Cost of Conversion
The primary asset that miners bring to the AI table is not their mining rigs. It is their power access. The industry has collectively secured gigawatts of low-cost electricity, often at $0.03–0.05 per kWh, with the ability to ramp down during grid peaks. This is valuable because AI training is power-hungry and time-sensitive. An AI lab that needs 100 MW of compute capacity cannot wait three years for a new substation. It can lease a miner's existing facility and bring its own GPUs. The miner provides the real estate, power, cooling, and operations. The AI lab provides the hardware and the client. The revenue split is typically 50–60% to the miner, depending on the contract.
But the conversion is not free. The miner must retrofit the facility. ASIC mining farms use air cooling and simple racking. GPU clusters require liquid cooling, high-density power distribution, and low-latency networking. The cost per MW for retrofitting an existing mining facility to AI/HPC standards is estimated at $2–4 million per MW, depending on the existing infrastructure. New builds are even higher. The capital expenditure required to convert a 100 MW mining site to AI compute is between $200 million and $400 million. That is a significant portion of the miner's market cap. For example, Riot's market cap is around $3 billion. Its $9.1 billion contract with Anthropic implies a massive capital outlay over the next few years. The question is how the miner finances this outlay. Equity dilution, debt, or project financing. Each has its own risk profile.
From my own work verifying the circuit constraints for early zk-Rollup proofs, I learned that the gap between a mathematical model and a real-world implementation is where most failures occur. The same applies here. The miner's financial model assumes that the AI customer will pay the contracted rate for the full 20 years, that the GPU hardware will not become obsolete, and that the miner can operate the cluster at the promised uptime. In practice, AI contracts often include clauses that allow the customer to renegotiate or terminate if the service level drops below a threshold. The miner's operational history is in mining, not in AI compute. The first few quarters will reveal whether the miner can maintain a 99.9% uptime SLA. I have seen too many projects where the whitepaper math assumed perfect execution. Reality is messier. Complexity is the enemy of security.
Valuation Divergence: A Deeper Look
The market's current pricing of miners with AI contracts implies a significant premium for the option on future AI revenue. But the premium is not uniform. TerraWulf and IREN have seen their stock prices more than double. They were early movers. They signed contracts in 2023 and 2024, before the current AI hype cycle peaked. Their facilities are already partially retrofitted. They have demonstrated some operational capability. Cipher, with its Bitmain partnership and power resources, is also climbing the curve. MARA, by contrast, is still largely a pure Bitcoin miner. Its stock price decline reflects the market's judgment that its management team was slow to adapt. The company's large Bitcoin treasury is a double-edged sword: it provides a buffer against hashprice declines, but it also exposes the balance sheet to Bitcoin's volatility. The market is now punishing that exposure.
But the divergence is not just about contracts. It is about the quality of the contracts. A 20-year contract with a tier-1 AI lab like Anthropic is different from a 2-year rental agreement with a smaller AI startup. The length of the contract provides revenue visibility. The counterparty risk is lower. The market rightly values that. However, the market may be overvaluing the early-stage contracts. Many of the announced deals are framework agreements or letters of intent. They are not firm commitments with penalties. The actual revenue may take 12–24 months to materialize. During that time, the miner must spend capital on retrofitting, hiring AI operations staff, and purchasing or leasing GPUs. The gap between contract signing and revenue generation is a window of negative cash flow. This is where the risk lies.
Contrarian Angle: The Blind Spots in the Transition Narrative
The market narrative assumes that every miner with a power contract can convert to AI compute. That is not true. AI data centers require power quality that is different from mining. Mining ASICs can tolerate brief power interruptions; a GPU cluster doing four-hour training runs cannot. The power supply must be stable, redundant, and backed by uninterruptible power supplies. Many miners have interruptible power contracts that allow the grid to curtail their load during peak demand. That is fine for mining, but it is a dealbreaker for AI. The miner must either upgrade the power contract to a firm, uninterruptible agreement—which costs more—or build battery storage and backup generators. Both add capital cost.
Furthermore, the operational expertise required for AI compute is different. Mining is a low-touch operation: you install ASICs, monitor temperature, replace failed units. AI compute is a high-touch operation: you manage GPU drivers, job scheduling, network topology, and customer access. The miner must hire data center engineers who understand these systems. This is a talent acquisition challenge. The market is pricing the transition as if the miner can simply flip a switch. Based on my experience auditing the data availability sampling mechanism of Celestia's testnet, where we simulated 10,000 nodes dropping offline, I know that operational complexity can lead to cascading failures. The miner's transition is not a simple switch. It is a multi-year transformation.
Another blind spot is the concentration risk. The $70 billion in AI contracts is concentrated among a handful of large AI labs. Anthropic, the customer for Riot's deal, is a single company. If Anthropic's funding or demand changes, Riot's contract could be renegotiated or terminated. The same applies to other miners. The market is treating these contracts as if they are guaranteed. They are not. Audits are snapshots, not guarantees. The same principle applies to contracts: they are snapshots of a relationship at a point in time. The future is uncertain.
Takeaway: The Winners Will Be Those Who Execute, Not Those Who Announce
The pivot from Bitcoin mining to AI/HPC is a rational response to hashprice compression. The market is correctly rewarding miners who have signed contracts. But the premium may be excessive for those who have not yet demonstrated delivery. The next 12 months will separate the miners who can operate a GPU cluster at scale from those who cannot. The capital expenditure cycle will test balance sheets. The winners will be the miners with the lowest cost of capital, the strongest power contracts, and the best operational teams. The losers will be those who over-leverage and fail to meet SLAs. The market will eventually learn that contracts are not revenue. I will be watching the quarterly earnings and the power purchase agreements. That is where the real math lives. Code does not care about your vision. Neither does the AI compute market.