CoinShares：AI 与 BTC 的机架博弈

Source: CoinShares Q1 2026 Bitcoin Mining Report; Translated by Jinse Finance Claw

I. Executive Summary

The fourth quarter of 2025 was the most challenging quarter for Bitcoin miners since the halving in April 2024. Due to a significant drop in Bitcoin prices (from a historical high of around $124,500 in early October to about $86,000 by the end of December, a decline of approximately 31%), coupled with hash power nearing historical highs, the hash price has compressed to a five-year low.

Among publicly listed mining companies, the weighted average cash cost to produce a single Bitcoin in the fourth quarter of 2025 rose to about $79,995.

Three defining themes emerged this quarter:

• Profit squeeze: The hash price fell to around $36–38 per PH/s/day, close to or at the breakeven point for many miners. Three consecutive downward adjustments in difficulty (the first consecutive decrease since July 2022) marked the beginning of miner capitulation. Into the first quarter of 2026, this figure further plummeted to $29 per PH/s/day, suggesting more pain ahead for miners.

• Acceleration of AI/HPC (Artificial Intelligence/High-Performance Computing) transformation: The divide between pure mining companies and new infrastructure companies transitioning to AI has widened. Currently, the publicly listed mining sector has announced cumulative AI/HPC contracts exceeding $70 billion. WULF, CORZ, CIFR, and HUT are effectively transforming into data center operators that also engage in Bitcoin mining.

• Capital structure transformation: Several miners have taken on substantial debt to fund AI facility construction. IREN currently holds $3.7 billion in convertible notes; WULF has $5.7 billion in total debt; CIFR has issued $1.7 billion in senior secured notes. The overall leverage in the industry has fundamentally altered its risk profile.

II. Competition for Rack Space between AI and Bitcoin Mining

AI continues to compete for rack space in many data centers, which could drive Bitcoin mining to more intermittent and cheaper energy sources in the long term.

The migration of Bitcoin miners to AI and high-performance computing is accelerating rapidly. According to recent company announcements, by the end of this year, listed miners could derive up to 70% of their revenue from AI, compared to around 30% currently. What initially started as an edge diversification strategy is increasingly becoming a core business.

During the latter part of 2025 and early 2026, Bitcoin miners signed several GPU hosting and cloud service agreements with Hyperscalers, with a total value exceeding $70 billion. While most transactions envision building new data centers, there is a strong likelihood of “cannibalization” or closures of existing mining facilities. Therefore, as capacity from these contracts is released, the share of Bitcoin mining revenue for these operators is expected to decline significantly throughout 2026.

This shift is primarily economically driven. The hash price remains near cyclical lows, squeezing mining margins, while AI infrastructure offers more structured and stable returns. In this context, it makes sense to redeploy power and capital towards HPC, especially for operators with scalable energy and existing data center capabilities.

Nevertheless, this transformation is not uniform. Some miners, like IREN and Bitfarms, are actively repositioning themselves as HPC providers, effectively using mining as a bridge into AI infrastructure. Others, like CleanSpark, continue to prioritize mining in the short term, monetizing recently developed capacity while gradually building AI exposure.

A third group remains committed to Bitcoin mining but is evolving operationally. These operators are no longer pursuing super-sized facilities but instead focusing on the lowest cost, often intermittent energy sources, such as stranded renewable energy or flare gas. For example, Marathon has deployed smaller, localized approximately 10 MW container sites at the edge of the energy grid. These setups are well-suited for mining that can tolerate interruptions but are incompatible with AI workloads that demand near-continuous uptime.

Load balancing may still be a lasting niche in the mining industry. By providing demand flexibility to grids like ERCOT, miners can secure more favorable electricity rates. This role may become increasingly important, although over time, it may attract smaller, more specialized operators.

A core unresolved question is the durability of the AI-driven transformation. While current economic conditions strongly favor AI, mining remains highly sensitive to Bitcoin prices. If mining profitability significantly recovers, some operators might reassess capital allocation between the two activities. In this sense, the current trend may not represent a permanent transformation but rather a reflection of relative rates of return.

In the long run, this could lead to a smaller group of pure miners and a broader mix of hybrid infrastructure companies spanning both mining and AI. At the same time, new entrants may emerge to exploit niches left vacant by existing participants, particularly in energy-constrained or highly flexible market segments.

The cost difference between Bitcoin mining infrastructure (approximately $70-100 thousand/MW) and AI infrastructure (approximately $8-10 million/MW) is substantial, and this conversion opportunity is currently being realized on a large scale:

• CORZ: Approximately 350 MW of power for HPC, with about 200 MW already generating revenue. The contract with CoreWeave has expanded to 12 years at $10.2 billion, aiming for full load at 590 MW by early 2027.

• WULF: There are 39 MW of critical IT capacity online at Lake Mariner. Signed HPC total revenue of $12.8 billion. Additional construction plans are ongoing through the fourth quarter of 2026. The platform plans to expand to about 2.9 GW across five sites.

• CIFR: Developed the 300 MW Barber Lake site in partnership with Fortress Credit Advisors. Reached a multi-billion dollar agreement with Fluidstack (supported by Google). Currently has not generated revenue.

• IREN: Expanded to over 10,900 NVIDIA GPUs. The Childress Horizon 1-4 expansion project (up to 200 MW of liquid-cooled GPUs) is under construction. AI cloud service revenue reached $17.3 million in the fourth quarter.

• HUT: Signed a 15-year lease worth $7 billion with Fluidstack, involving its 245 MW capacity at River Bend Park in Louisiana, with the first data hall expected to go live in early 2027.

The failed merger between CORZ and CoreWeave (voted down by shareholders on October 30, 2025) highlights the tension between infrastructure value and equity value. CORZ subsequently restated its financial statements due to improper capitalization of assets slated for decommissioning during the HPC transition, reflecting the complexities of accounting treatment.

Revenue contributions are still in the early stages but are growing: CORZ’s hosted AI/HPC data center accounted for 39% of fourth-quarter revenue; WULF’s HPC accounted for 27%; IREN’s AI cloud accounted for 9%; HIVE’s HPC accounted for 5%. Mining still dominates, but it is evident that the revenue contributions from AI are growing across nearly all companies.

III. Network Hashrate

The Bitcoin network reached a significant milestone at the end of August 2025, breaking through 1 ZH/s (Zettahash) for the first time. The network hashrate peaked at approximately 1,160 EH/s in early October.

However, a significant decline occurred in the fourth quarter. The hashrate fell by about 10% from the October highs, dropping to around 1,045 EH/s by the end of December (later falling to 850 EH/s in early February before rebounding), and three consecutive negative difficulty adjustments were recorded, the first since July 2022. This was primarily driven by the following factors:

1. The Bitcoin price correction pushed outdated hardware from the S19 series below breakeven (the breakeven electricity price for the S19 XP fell from about $0.12/kWh in December 2024 to about $0.077/kWh in December 2025).

2. Increased winter energy costs and ERCOT (Texas grid) reductions led to a sharp increase in unprofitable mining hours from November to December.

3. Regulatory actions resumed in December 2025 in Xinjiang, China (inspections led to reduced mining operations, although hashrate was not permanently lost).

Despite the short-term decline, the network increased by about 300 EH/s throughout 2025. As of the time of writing, the hashrate remains at approximately 1,020 EH/s, consistent with the end of 2025.

While the recent decline in hashrate may seem alarming, from a logarithmic scale, it is far less severe than during the Chinese mining ban in 2021, representing a combination of cyclical and weather factors rather than ominous signs facing the industry. The hashrate has since rebounded significantly, indicating that many miners still view this as a viable economic activity.

Based on our previously detailed segmented forecasting model, we now anticipate the hashrate will reach 1.8 ZH/s by the end of 2026 and 2 ZH/s by the end of March 2027, a month later than our prior forecast.

Geographical Shift: The top three countries (the U.S., China, and Russia) control about 68% of the global hashrate. The U.S. market share has increased by about 2 percentage points quarter-over-quarter. Emerging markets like Paraguay (300 MW) and Ethiopia (40 MW) have entered the global top ten, driven by miners such as HIVE and BTDR.

IV. Hash Price Dynamics

The hash price (a measure of miner revenue per unit of hashrate) peaked at around $63 per PH/s/day in July and steadily declined throughout the fourth quarter of 2025. By November, it had fallen to approximately $35-37 per PH/s/day, marking a five-year low at that time. A brief recovery to $38-40 per PH/s in late December to early January proved to be short-lived, as the hash price further crashed in the first quarter of 2026, reaching about $28-30 per PH/s/day in early March, setting a new historical low post-halving.

This decline was driven by several factors: record difficulty (which peaked at 155.97T after a 6.31% increase on October 29), a depressed Bitcoin price (approximately 31% lower than the historical high in October), and extremely low transaction fee revenues (always below 1% of total block rewards, with an average fee of about 0.018 BTC per block).

This created the tightest profit environment since the April 2024 halving. Miners operating mid-generation hardware (S19j Pro level, with an efficiency of about 29.5 J/TH) at an average industrial electricity price of $0.05/kWh (for S19 XP, $0.077/kWh) were far below breakeven by the end of the year, and the situation worsened further heading into 2026.

Updated forecast: The hash price environment has exceeded our prior expectations, briefly touching around $28 per PH/s/day in late February before recovering to its current range of about $30-35. At these levels, miners operating mid-generation hardware need to secure electricity below $0.05/kWh to remain cash flow positive, while the latest generation units (below 15 J/TH) still retain considerable profit margins at typical industrial electricity prices. For the hash price to sustainably recover above $40, Bitcoin prices would need to initiate a push towards $100,000 by year-end, with increases outpacing the continued growth in hashrate.

Unless Bitcoin prices experience a substantial recovery, we expect more high-cost operators to capitulate in the first half of 2026. The current mining economic conditions are insufficient to incentivize a widespread hardware upgrade cycle. The hash price may need to decline further first, leading to sufficient old-generation capacity and operator exits, reducing network hashrate levels and difficulty, which could provide entry points for new Bitcoin miners or drive upgrades for existing operations. However, despite continued pressure on profit margins, network hashrate has behaved extraordinarily resilient. This may be due to state-backed mining with strategic objectives rather than purely economic-driven tasks, operators with access to extremely cheap or stranded energy, and ASIC manufacturers connecting unsold inventory to their own facilities to maintain order commitments with foundries like TSMC and Samsung.

The pain in the mining sector has led miners to capitulate, with many liquidating their holdings. Publicly listed mining companies have cumulatively reduced their treasury reserves by over 15,000 BTC from peak levels. Among them, Core Scientific sold around 1,900 BTC (approximately $175 million) in January alone and plans to liquidate nearly all remaining positions in the first quarter of 2026; Bitdeer cleared its treasury in February; Riot sold 1,818 BTC (approximately $162 million) in December 2025.

We believe it is not unrealistic for Bitcoin prices to rebound to the $100,000 mark. At that price, the hash price is expected to recover to $37 per PH/day. If Bitcoin prices remain below $80,000 for the rest of the year, assuming difficulty continues to rise, we predict the hash price will continue to decline. In this scenario, hashrate may further decrease as miners shut down unprofitable machines, making a flat hash price more likely. If we see prices start to challenge the historical high of $126,000, we could see the hash price rise to $59 per PH/day.

Current hash prices have fallen far below our projected range, although we consider this a temporary issue caused by the recent price drop and expect it to stabilize in the range of $30-40 per PH/day.

Current hash prices make mining various models of mining machines unviable. At the current hash price of $30 per PH/day, any miner below S19 XP efficiency and electricity prices above $0.06/kWh is operating at a loss—we estimate this accounts for about 15-20% of the global miner population.

V. Mining Cost Analysis

5.1 Overview

The following table presents the unit Bitcoin cost details for all covered miners in the fourth quarter of 2025. All figures are in USD/BTC.

Core observations:

1. AI/HPC construction distorts the unit cost metrics for mixed operators. Debt, SG&A, and D&A driven by AI infrastructure construction have been allocated to the shrinking Bitcoin production base, inflating the nominal unit cost. For WULF, CORZ, and CIFR, all-in costs increasingly reflect the economics of transitioning to data center operators rather than mining economics.

2. Electricity costs have risen significantly. Compared to the second quarter of 2025, all companies have seen significant increases in electricity costs, reflecting the dilution of unit output by rising network difficulty, increasing winter energy costs, and declining Bitcoin prices.

3. Depreciation is the largest non-cash cost, with significant policy differences. MARA’s $136,000/BTC and CIFR’s $88,000/BTC are outliers (MARA due to its large fleet; CIFR due to a three-year depreciation assumption).

4. Stock-based compensation (SBC) remains a key differentiator. HUT’s $48,500/BTC (driven by a one-time authorization from the CEO/CSO) and CORZ’s $35,500/BTC are outliers. BTDR ($3,900) and CLSK ($6,700) demonstrate the most rigorous cost control.

5. Interest costs currently have a significant impact on several miners. WULF ($145,000/BTC), CIFR ($56,000/BTC), and BTDR ($16,000/BTC) are burdened with substantial debt. HIVE ($320/BTC) and CLSK ($830/BTC) have low leverage, providing a structural advantage.

5.2 Company Details

MARA (MARA Holdings)

• BTC mined: 2,011

• All-in cost: $153,040/BTC

• Cash cost (excl. tax): $103,605/BTC MARA remains the largest publicly listed miner by output in the fourth quarter. Its operational hashrate at the end of December was 53.2 EH/s. Electricity costs were $64,703/BTC, reflecting its geographic diversity and heavy reliance on third-party hosting. Depreciation reached $136,166/BTC, the highest among peers, reflecting its large fleet. The all-in cost was significantly distorted by a $183.4 million income tax benefit (driven by fair value adjustments of BTC holdings). Excluding this non-operating gain, all-in costs rise to $240,407. Notably, MARA formally authorized the sale of its entire reserve of 53,822 BTC in its March 2026 earnings report, marking a significant shift from its “HODL” strategy since July 2024.

IREN (IREN Limited)

• BTC mined: 1,664

• All-in cost: $140,441/BTC

• Cash cost: $58,462/BTC IREN has the lowest unit electricity cost ($34,325), benefiting from favorable agreements and demand response income from its Texas Childress facilities. SBC of $31,717/BTC is the second highest among peers. IREN holds $3.7 billion in convertible debt, the heaviest nominal debt burden among peers, but low coupon rates keep interest expenses manageable.

CLSK (CleanSpark)

• BTC mined: 1,821

• All-in cost: $118,932/BTC

• Cash cost (excl. tax): $71,188/BTC CleanSpark demonstrates excellent operational discipline, with management and stock-based incentive costs being the lowest among peers. Because it is purely a mining business (no hosting/HPC revenue), the analysis is straightforward. Interest costs are extremely low ($830/BTC), reflecting a low-leverage balance sheet.

RIOT (Riot Platforms)

• BTC mined: 1,324

• All-in cost: $170,366/BTC

• Cash cost (excl. tax): $102,538/BTC Electricity costs were $49,196/BTC, benefiting from $9.9 million in grid demand response incentives. Its strategic core is the Corsicana site, where 600 MW has been allocated to AI workloads, making it one of the largest single-site operators in North America.

CORZ (Core Scientific)

• BTC mined: 421

• All-in cost: $168,693/BTC

• Cash cost: $110,282/BTC The fourth quarter marked a milestone for CORZ’s transition to AI, with hosting revenue accounting for 39% of total revenue. Due to the shift of capacity to HPC, BTC output was extremely low, inflating the per-unit metrics. The company restated its financial statements and changed auditors to KPMG.

WULF (TeraWulf)

• BTC mined: 262

• All-in cost: $471,841/BTC

• Cash cost: $384,517/BTC Note: Its per-unit cost is not comparable with pure mining peers. The company has fundamentally transformed into an AI infrastructure company, and its mining business is continuously shrinking. The extremely high all-in cost reflects massive interest expenses ($5.7 billion total debt) and administrative costs for scaling AI.

CIFR (Cipher Digital)

• BTC mined: 591

• All-in cost: $231,980/BTC

• Cash cost: $103,516/BTC Due to a three-year depreciation policy, depreciation costs are high at $87,000. The issuance of $1.7 billion in senior secured notes in November led to a surge in interest expenses. The company was renamed Cipher Digital Inc. in February 2026.

HUT (Hut 8 Corp.)

• BTC mined: 719

• All-in cost: $160,402/BTC

• Cash cost: $50,332/BTC The all-in cost was influenced by a high executive stock authorization in November, with SBC reaching a peer-high of $48,500/BTC. Excluding these one-time items, its cash costs are highly competitive. The company holds 15,679 BTC, with a relatively complex structure.

BTDR (Bitdeer Technologies Group)

• BTC mined: 1,673

• All-in cost: $118,188/BTC

• Cash cost: $87,144/BTC Bitdeer’s costs are highly competitive, reflecting synergistic effects from its multiple business lines. In the fourth quarter, management shortened the depreciation period for mining machines, causing a doubling of book depreciation costs, but this pertains to accounting treatment rather than operational deterioration. Its proprietary SEALMINER chip is its core competitive advantage.

HIVE (HIVE Digital Technologies)

• BTC mined: 884

• All-in cost: $144,321/BTC

• Cash cost: $75,274/BTC Benefiting from expansion in Paraguay, output has significantly increased. Electricity costs are elevated due to VAT accounting in Paraguay. Interest costs are only $320/BTC, the lowest among peers, providing a significant structural advantage.

VI. Mining Company Stock Performance and Valuation

The valuation premium for AI/HPC continued to expand in the fourth quarter. Miners with signed HPC contracts are currently trading at 12.3 times EV/next year’s expected sales (NTM Sales), while pure mining companies only command a multiple of 5.9 times.

The industry has fundamentally split into:

• “Infrastructure companies”: WULF, CORZ, CIFR, HUT

• “Mining companies”: MARA, CLSK, RIOT, HIVE Whether AI-driven high valuations are justified depends on execution: not all announced agreements can be converted into operational infrastructure, and capital needs remain enormous.

VII. Q1 2026 and Future Outlook

1. Hash price recovery depends on Bitcoin price: At a $70,000 Bitcoin price and $30 hash price, many mid-generation fleets are on the brink of loss. A drop below $70,000 could trigger large-scale capitulation, which in turn would benefit survivors.

2. Next-generation hardware deployment: Bitmain’s S23 series and SEALMINER A3 (with efficiencies below 10 J/TH) will scale up in the first half of 2026, accelerating the phasing out of old models.

3. AI/HPC revenue inflection point: The market will closely observe whether signed revenues can convert into billed revenues and whether gross margins can achieve the target of over 85%.

4. Leverage differentiation gives rise to M&A catalysts: Miners with clean balance sheets and strong liquidity (like HIVE and CLSK) may become acquirers.

5. Geographical and regulatory shifts: The U.S. continues to gain market share. Paraguay and Ethiopia are emerging players. Texas SB 6 bill (signed in June 2025) introduces mandatory remote disconnection requirements for large loads connected to the grid.

6. Industry consolidation: We expect more mergers and acquisitions in 2026. The efficiency gap between top-tier fleets (15 W/T) and lagging fleets (25+ W/T) is substantial, making direct acquisitions of efficient capacity economically more viable than upgrading old facilities.