In Issue 8, I showed that Bitcoin's difficulty adjustment makes its scarcity invariant to energy conditions — whether a civilization burns coal or captures stars, exactly the same number of coins come out. In Issue 9, I argued that the real vulnerability isn't energy at all, but revenue: the declining block subsidy and the uncertain transition to a fee-funded security model.
This issue connects those threads to something happening right now. AI data centers are outbidding Bitcoin miners for electricity, and the displacement is reshaping mining economics in real time. The Dyson sphere thought experiment from Issue 8 was about unlimited energy. This is about what happens when a higher-value use for the same energy shows up.
The Revenue Gap Is Already Enormous
The economic displacement is stark. The revenue differential between AI hosting and Bitcoin mining has widened significantly since the April 2024 halving.
Galaxy Digital pivoted from Bitcoin mining to lease 133 MW at its Helios campus to CoreWeave for approximately $4.5 billion over 15 years — roughly $300 million per year. Galaxy's previous Bitcoin mining revenue from similar capacity was approximately $22 million per year (per Galaxy Digital earnings disclosures). That's more than a 10× revenue improvement per MW. By October 2025, the deal had expanded to 800 MW, and Galaxy's CEO Michael Novogratz told investors the company is now "half a data-center company and half a digital-assets company."
Galaxy isn't alone. By October 2025, Bitcoin miners had announced approximately $65 billion worth of AI/HPC hosting contracts — effectively converting mining infrastructure to AI compute. Core Scientific signed a deal with CoreWeave valued at over $8 billion across its announced tranches. Hut 8 signed a $7 billion, 15-year AI data center lease at its River Bend campus in Louisiana with Fluidstack, with Google as financial backstop (Hut 8 press release, 17 December 2025). Crusoe Energy divested its Bitcoin mining assets entirely to focus on AI GPU clusters. The Crusoe exit is worth flagging specifically: Crusoe was built around flared gas mining — the canonical stranded energy use case the industry holds up as proof of mining's environmental fit. If even that model loses to AI economics, the stranded energy floor is narrower than the bull case assumes.
Global data center electricity consumption reached 415 TWh in 2024, projected to hit 945 TWh by 2030, and AI data centers surpassed Bitcoin mining in total energy consumption in 2025 (IEA, Energy and AI, April 2025). Gartner projects AI-optimized servers will grow from 21% to 44% of data center power (Gartner, November 2025).
When a higher-value use for energy exists, Bitcoin mining gets displaced — not destroyed, but pushed to the margins.
The Buyer of Last Resort
Troy Cross of Reed College and the Bitcoin Policy Institute has written extensively on Bitcoin mining's natural niche as a "buyer of last resort" for energy — absorbing electricity too remote, intermittent, or low-quality for other uses. Cross has articulated this in BPI publications and CoinDesk, though this framework exists in advocacy and opinion writing rather than peer-reviewed literature.
Its empirical grounding is real. Riot Platforms earned $31.7 million in curtailment credits from ERCOT during a Texas heat wave in August 2023 — far exceeding the $8.9 million from the 333 BTC it mined that month (per its August 2023 production report). ERCOT data shows miners curtailed 888 GWh across 2023, and ancillary services costs to ratepayers dropped 74% from 2023 to 2024 (ERCOT 2024 Annual Report).
AI data centers, by contrast, demand high-quality, uninterrupted baseload power. They cannot flex. This creates a potential complementarity — mining as flexible interruptible load paired with AI as baseload — that several miners are actively pursuing. IREN (Iris Energy) is the clearest example: its Horizon 1 facility in Childress, Texas is a liquid-cooled 50 MW site designed to shift dynamically between Bitcoin mining and AI workloads depending on market conditions. In November 2025, Microsoft signed a $9.7 billion, five-year agreement with IREN covering 200 MW of NVIDIA GB300 GPU capacity (IREN press release, 3 November 2025) — while IREN simultaneously expanded its mining operation to 50 EH/s (IREN investor disclosures, 2025). The dual-use model works when the underlying infrastructure can genuinely flex; the question is how many facilities are actually built that way versus simply repurposed from one use to the other.
This connects to the energy economics I explored in Issue 8. Bitcoin's long-term positioning isn't about competing for premium electricity. It's about being the economic floor — the buyer who takes what nobody else will. The AI displacement accelerates this: as AI claims the premium power, mining naturally migrates toward stranded and curtailable energy that AI can't use.
The geographic data supports this migration story at the margins, though the US still dominates at roughly 38–44% of global hashrate — a share that has held steady despite the AI pivot. The migration story is real but incremental: the 2024 halving acted as a centrifugal force, spinning marginal hardware out of high-cost US grid operations and into the Global South. Ethiopia now accounts for roughly 2.5% of global hashrate — more than double the entire continent's share in 2023 — entirely via surplus hydropower from the Grand Ethiopian Renaissance Dam, a textbook stranded energy case where Ethiopian Electric Power faced a crisis of electrons with no buyers. In 2024, Ethiopia earned over $55 million from electricity sales to miners (per Ethiopian Electric Power revenue reporting). Paraguay, running on Itaipu surplus, holds approximately 4% of global hashrate at electricity costs as low as $2.8–$4.6 per MWh. Africa overall has grown to roughly 3% of global hashrate, almost entirely via renewables. These are meaningful footholds, but the stranded energy thesis — mining migrates decisively to the margins — remains a directional trend rather than a completed transition.
The question is whether that floor is large enough and profitable enough to sustain the security budget discussed in Issue 9.
The Clean Energy Story Is More Complicated Than the Industry Claims
I want to flag two caveats to the narrative that mining displacement toward stranded energy is inherently good.
First, the renewable share figures tell a complicated story. De Vries, Gallersdörfer, Klaaßen, and Stoll (2022, Joule) found that after China's mining ban, Bitcoin's renewable share dropped from 41.6% to 25.1% as miners migrated to coal-heavy Kazakhstan and natural gas-dependent US operations. The subsequent Cambridge Centre for Alternative Finance Digital Mining Industry Report (2025) shows recovery to 52.4% sustainable (including nuclear), but independent analysis of publicly listed US miners found their actual grid-average renewable share significantly below industry self-reported claims (Stoll et al., "Climate Impacts of Bitcoin Mining in the U.S.," MIT CEEPR Working Paper 2023-11). Mining migrates toward cost, and cost correlates with clean energy only under specific market conditions. The stranded-energy story is real in some regions; it's not automatically the industry average.
Second, and more significant: the carbon externality question. Papp, Almond, and Zhang (2023, Journal of Public Economics 227:105003) used daily Bitcoin price variation as a natural experiment at a coal-powered mining facility, finding that "$1 increase in Bitcoin price generates $3.11–$6.79 in external carbon damages" — exceeding mining's value added at that facility. Jones, Goodkind, and Berrens (2022, Scientific Reports) estimated each $1 in Bitcoin market value created $0.35 in global climate damages across 2016–2021. These findings are specific to fossil-fuel operations and don't generalize to hydro or geothermal mining. But they establish that for the marginal kWh of coal-powered mining, the external cost is substantial and quantifiable.
An article about Bitcoin and energy that doesn't acknowledge this literature has a conspicuous blind spot. Stranded and clean are not synonyms. Whether mining's displacement toward the margins means displacement toward renewables or toward unregulated fossil fuels depends on economic geography, not protocol design.
Is the AI Displacement Permanent?
Before asking whether the pivot is durable, it's worth confronting a fact that complicates the displacement narrative directly: Bitcoin's hashrate — which first crossed 1 ZH/s instantaneously in April 2025 and on a sustained 7-day basis by September 2025, as covered in Issue 8 — continued setting new records through late 2025, during the same period institutional miners were pivoting to AI en masse. The network didn't shrink. It grew.
The explanation is that displacement and growth are happening simultaneously — and in this cycle, replacement outpaced exit. Institutional miners pivoting to AI are being replaced by a new generation of highly efficient ASICs and by operators in lower-cost jurisdictions, faster than they left. The difficulty adjustment does exactly what Issue 8 described, but here in the bullish direction: incoming miners overwhelmed the exits, pushing mining difficulty to an all-time high of 156 trillion in November 2025. The supply of coins is unaffected. So, broadly, is hashrate — and both rose.
But the economics beneath that headline figure are under severe pressure. All-in mining costs — a comprehensive measure covering hardware depreciation, financing, and overhead, and broader than the operational-only estimates JPMorgan reported in Issue 8 — reached approximately $137,800 per BTC in early 2026, well above market price for much of the period (BTC was trading in approximately the $90,000–$110,000 range during early 2026). Hardware payback periods reportedly exceeded 1,200 days, compared to roughly 300–500 days in prior cycles (ApexToMining, January 2026). Only the most capital-efficient operators — those with access to ultra-cheap power — remain viable. This consolidation is the real story: hashrate is at all-time highs, but the industry producing it is narrower, more concentrated, and operating on thinner margins than at any prior point.
Whether the mining-to-AI pivot is durable is worth flagging as uncertain. Goldman Sachs projects a potential 2027 inflection where AI data center supply catches demand — occupancy peaks at ~93–95% in late 2026, then moderates as overcapacity builds. The current mining-to-AI revenue differential may compress if AI infrastructure demand slows.
There's also a question of reversibility. Mining infrastructure that's been physically retrofitted for AI hosting — GPU racks, cooling systems optimized for different thermal profiles, network architectures designed for low-latency inference rather than high-throughput hashing — doesn't trivially convert back. The $65 billion in announced contracts are long-term commitments, typically 10–15 years. If AI demand softens and Bitcoin's price appreciates, some of that capacity is locked out of mining.
On the other hand, the displaced miners who survive will be the ones who've found genuinely stranded energy — the kind that has no higher-value buyer at any price. Those operations may be smaller, more distributed, and more resilient than the current generation of institutional miners competing for premium grid power. That's a different mining industry than the one that exists today, but it might be a more durable one.
Where I Might Be Wrong
First, I may be underestimating the scale of the dual-use model. IREN's early results are promising, but whether the flexible infrastructure approach generalizes beyond a handful of well-capitalised operators remains unproven. If it does, the displacement story becomes a coexistence story at scale.
Second, the carbon externality data I cited is real but concentrated. Papp et al.'s $3.11–$6.79 in damages per dollar of Bitcoin price increase was measured at a coal-powered facility. Extrapolating that to the entire network, which is 52.4% sustainable by the Cambridge survey, overstates the aggregate damage. The right framing is that Bitcoin mining's carbon footprint varies enormously by operator and geography, and aggregate figures conceal more than they reveal.
Third, nuclear energy could change the picture substantially. Several mining operations are exploring co-location with nuclear plants, which provide cheap baseload power with near-zero carbon emissions. If nuclear becomes the marginal energy source for mining — rather than natural gas or coal — both the cost and environmental arguments shift.
Fourth, the pool centralization threat may be less acute in practice than the structural numbers suggest. Under the FPPS model, pool operators earn a small margin on aggregated hash power from thousands of independent miners — their revenue depends on Bitcoin functioning, not on any single profitable attack. Attacking the chain would destroy that franchise value immediately, creating a strong economic disincentive against the very action that pool concentration makes technically feasible. The geopolitical case for AntPool/Bitmain coordination is also contested: Bitmain operates profitable commercial hardware and pool infrastructure precisely because Bitcoin works. The structural concentration is real and the threat model is valid; whether it translates into adversarial action is a genuinely separate and uncertain question.
What the Displacement Reveals
The AI energy boom is the first real-world test of what I explored theoretically in Issue 8. When a higher-value use for energy appears, Bitcoin's difficulty adjustment handles it exactly as predicted: mining becomes less profitable at the margin, some miners exit, difficulty adjusts to a new equilibrium, and remaining miners find stable margins. Block production continues at roughly 144 per day. Supply is unaffected.
What's affected is security — and the mechanism is more direct than the hashrate headline suggests. Issue 9 cited Professor Harvey's estimate that a 51% attack costs roughly $6 billion in total — hardware, data center construction, and electricity combined. That calculation assumes an attacker must acquire new mining infrastructure. It doesn't account for pool-level concentration, which has quietly crossed a threshold that changes the threat model entirely.
As of 2025, Foundry USA (~34%) and AntPool (~18%) jointly control over 51% of global Bitcoin hashrate. Just six pools mine more than 95% of all blocks (b10c.me, 2025). These pools don't own the underlying hardware — they aggregate hash power from thousands of independent miners — but their operators control the block template, transaction selection, and chain extension decisions for a majority of the network. A coordinated action, compromise, or regulatory seizure of those two pool operators would cross the 51% threshold without purchasing a single additional ASIC. The $6 billion attack cost assumes a hardware acquisition scenario; it dramatically overstates the cost of a pool-level attack. It also ignores the selfish mining threat that pool concentration creates at lower thresholds: Eyal and Sirer (2014, FC) demonstrated that selfish mining becomes profitable once a pool controls more than 33% of hashrate — a threshold Foundry USA already crosses on its own, before any coordination with AntPool is considered.
The geopolitical dimension compounds this. AntPool is operated by Bitmain, formally headquartered in Singapore but Chinese in origin, and researchers tracking pool-level data have noted that VPN usage and cloud-based mining services likely cause Chinese-linked entities to control more hashrate than IP-based geolocation surveys capture (Cambridge DMIR 2025; b10c.me, 2025). The US holds approximately 38–44% of global hashrate via Foundry and independent operators, but the second-largest pool by hashrate is governed by an entity with complex ties to a jurisdiction that has banned Bitcoin mining and has clear strategic incentives around financial infrastructure. Bitmain's exposure runs deeper than pool operation: the same company manufactures an estimated 82% of the ASICs running in mining facilities globally (Cambridge Digital Mining Industry Report, 2025) — including those contributing hashrate to Foundry. The vertical integration of chip fabrication, pool operation, and potential geopolitical leverage into a single corporate entity represents a threat model that hardware cost calculations don't capture.
The AI displacement doesn't cause pool centralization — that's a separate structural trend driven by the Full Pay-Per-Share (FPPS) payout model that incentivizes miners to join the largest pools regardless of security implications. But the displacement intensifies it: as marginal miners exit and the industry consolidates around fewer, larger operators with stranded energy access, the distribution of hash power narrows further. The ones who remain are on cheaper energy, which helps their margins — but the industry producing Bitcoin's security is becoming structurally more concentrated, not less.
The Dyson sphere thought experiment showed that Bitcoin's scarcity survives unlimited energy abundance. The AI displacement shows the real-world version: Bitcoin's scarcity survives just fine, but the security model depends on economics that energy competition can erode.
Bitcoin doesn't need energy to be scarce. But it does need mining to be worth doing.
Sources & Further Reading
Hashrate & mining economics
- Bitcoin network hashrate data via Blockchain.com/CoinWarz (Q4 2025 hashrate ATH; first 1 ZH/s instantaneous crossing April 2025, sustained 7-day average September 2025)
- ApexToMining, "Bitcoin Hashrate Hits 1 ZH/s: Mining Costs Surge to $137K" (January 2026)
- b10c.me, "Bitcoin Mining Centralization in 2025" (pool concentration data, Foundry/AntPool)
- CryptoSlate, "Foundry USA and AntPool Command Almost 60% of Bitcoin Mining Pool Market"
- Eyal & Sirer, "Majority Is Not Enough: Bitcoin Mining Is Vulnerable," Financial Cryptography and Data Security (2014) — selfish mining profitability threshold (33%)
- Hashrate Index, "Global Hashrate Heatmap Update: Q4 2025" (US share ~38%)
AI & energy competition
- IEA, Energy and AI (April 2025)
- Gartner press release on data center power forecasts (November 2025)
- Galaxy Digital press releases and Q4 2025 earnings (Helios/CoreWeave deal details)
- Core Scientific, Hut 8, and Crusoe Energy press releases and investor announcements
- CoinDesk/CoinShares, revenue-per-kWh comparisons and $65B AI contract aggregation
- Goldman Sachs, "Generational Growth: AI, Data Centers and the Coming US Power Demand Surge" (2025)
- IREN investor disclosures and Horizon 1 facility announcements (2025)
- Microsoft/IREN GPU capacity agreement reporting (2025)
Geographic redistribution
- Hashrate Index, "Top 10 Bitcoin Mining Countries" (2025, 2026 editions)
- Forbes/DNS Africa, "Africa Produces 3% of Global Bitcoin Mining Hashrate Via Renewables" (December 2024)
- Afriwise, "Cryptomining in Ethiopia — Legal Outlook 2025"
- Ethiopian Electric Power revenue reporting (2024)
- Hashrate Index / CryptoSlate reporting on Paraguay/Itaipu surplus mining operations (electricity cost $2.8–$4.6/MWh, ~4% of global hashrate, 2024–2025)
Mining and grid flexibility
- Troy Cross, Bitcoin Policy Institute (BPI publications and CoinDesk, 2023 — advocacy/opinion, not peer-reviewed)
- Riot Platforms August 2023 production report (curtailment credits)
- ERCOT 2023 and 2024 Annual Reports (curtailment data; ancillary services cost drop 74% year-on-year)
Environmental impact
- De Vries, Gallersdörfer, Klaaßen & Stoll, "Revisiting Bitcoin's Carbon Footprint," Joule 6(3) (2022)
- Cambridge Centre for Alternative Finance, Digital Mining Industry Report (2025)
- MIT CEEPR (2023), working paper on US miner grid-average carbon intensity vs. self-reported renewable claims
- Papp, Almond & Zhang, "Bitcoin and Carbon Dioxide Emissions: Evidence from Daily Production Decisions," Journal of Public Economics 227:105003 (2023)
- Jones, Goodkind & Berrens, "Economic Estimation of Bitcoin Mining's Climate Damages," Scientific Reports 12:14512 (2022)
Geo Nicolaidis
Builder, TrailBit.io
If you found this useful, subscribe to get the next issue in your inbox. Each issue breaks down a different heuristic used in Bitcoin forensics — what it assumes, where it breaks, and why it matters.