Please see relevant disclosures here.
For reasons that only the algorithmic gods know, I had a few conversations over the past week with people who seemed to be under the impression that converting bitcoin mining operations to AI data centers is a straightforward and routine thing. It is not; a lot of “it depends” goes into the decision. What follows explains the complexity latent in the decision of whether to convert a bitcoin mining operation to an AI data center. The bottom line is that converting a bitcoin mining operation to an AI data center requires that a significant number of hurdles be cleared.
Bitcoin mining and AI do not compete for the same asset; they merely share an input commodity. A Bitcoin mine monetizes cheap, flexible, often interruptible electricity: it is a controllable sink for surplus powe. An AI data center monetizes firm, high-availability, networked, thermally engineered critical load: it converts firm power into contracted compute availability. Those are opposite grid products that happen to both draw megawatts.
So when the market talks about “converting a Bitcoin mine into an AI data center,” the thing actually being monetized is neither mining expertise nor mining infrastructure. It is a scarce, partially de-risked power position: land, firm power (whether a grid interconnect or dedicated on-site generation), permits, water and cooling feasibility, fiber, and, decisively, a credit structure that can support a mission-critical tenant. The sheds, the ASICs, and the air-cooling are usually torn down. Core Scientific even restated financials over how it had capitalized assets committed to demolition during its HPC build-out.
What this means in practice: sites that happen to sit on a scarce firm-power position near adequate connectivity can re-rate; sites optimized for cheap, interruptible, remote power, which is most of the installed hashrate, cannot. The overhyped version treats a gigawatt of mining capacity as a gigawatt of latent AI capacity. It isn’t close.
A Bitcoin mine is engineered to be interrupted. A paused ASIC simply stops earning for a while; nothing is lost, nothing corrupts. That flexibility is the miner’s core economic edge. It lets the operator buy the cheapest, most curtailable power on the grid and get paid to shut down during peaks. An AI training run is the opposite: interrupting it forces a restart from the last checkpoint, burning hours or days of compute.
So, when a miner signs a long-term AI lease, it is selling away the very flexibility that made the mining economics attractive. The site becomes a load the grid must deliver firm power to. That is why a curtailment-optimized mine and a firm-load AI campus are not substitutes even when they draw identical megawatts. And regulators are starting to realize the difference: Texas SB 6 (June 2025) mandates remote-disconnection capability for large ERCOT loads, and large-load interconnection is increasingly scrutinized for financial, operational, and reliability impact rather than waved through on headline megawatts.
The lazy version of the screen asks whether a site has a grid connection. The real question is whether the site can deliver power that is there when the training run needs it. This screen has to address several things at once:
Deliverable and firm: enough contiguous megawatts, actually deliverable, not curtailment-contingent.
High-availability and concurrently maintainable: Uptime Institute Tier III equivalence: redundant components and distribution paths, so equipment can be serviced with no shutdown (~99.98% uptime). Mining runs at effectively Tier 0.
Thermally engineered: NVIDIA’s GB200 NVL72 draws ~120 kW per rack and the GB300 NVL72 reference design supports up to 142 kW, against ~40 kW for an H100 air-cooled rack and ~12 kW for a general-purpose rack. Above ~40 kW, air cooling stops working; direct-to-chip liquid cooling and water (or an engineered closed loop) become mandatory.
Networked: dark fiber and low-latency interconnect. Sites chosen for remote, stranded, cheap power frequently sit nowhere near the fiber routes AI wants.
Financeable: a counterparty and capital structure that can carry a mission-critical tenant.
Crucially, dedicated behind-the-meter generation, such as gas turbines, reciprocating engines, or fuel cells, increasingly paired with storage, passes this test. It is firm power, and it is one of the premier AI strategies precisely because it bypasses the multi-year interconnection queue rather than waiting in it. What fails is non-firm power: interruptible or curtailable grid arrangements, and load parked behind intermittent generation with no firming. The dividing line is dispatchability, not where the meter sits. This is also the point Galaxy Research makes: even a site with the right power may lack the acreage, water, dark fiber, skilled labor, permits, and long-lead electrical gear, and existing mining infrastructure is not directly transferable.
Firm, dispatchable power at scale: contiguous, 24/7, ideally 100 MW+, via an energized grid interconnect or dedicated on-site generation. The interconnect queue position (3–7 years to build new) is the scarce, re-ratable asset; behind-the-meter generation is the alternative route to firmness.
A path to Tier III redundancy: concurrently maintainable. Mining’s Tier 0 shells get upgraded or replaced.
Liquid-cooling feasibility + water: for 120–142 kW racks. Air-cooled sheds don’t retrofit gracefully.
Structural and site adequacy: slab loading, expansion land for a cooling plant and added substation capacity, security.
Dark fiber / low-latency connectivity: the gate that quietly kills remote flare-gas and remote-hydro sites.
Capital: the cost gap is the whole story. Mining infrastructure runs ~$700K–1M per MW; AI infrastructure ~$8–15M per MW. The GPUs alone dwarf an ASIC fleet.
A creditworthy counterparty and operating expertise: the hardest non-physical gate.
A site that clears all seven is rare. A site optimized for mining economics systematically fails 1, 5, and 7 because cheap-interruptible-remote is the mining playbook.
The reward in every case is a re-rate from mining multiples (~6–12x earnings) toward data-center multiples (~20–25x). Across the listed sector, miners had signed over $70 billion of GPU colocation and cloud deals by early 2026, and most of these deals envisage new data centers being built, not mines repurposed.
Galaxy Digital / Helios (Dickens County, TX). The archetype. Galaxy bought Helios from a distressed Argo Blockchain in 2022 for ~$65M to mine; ChatGPT changed the math. Galaxy cleared the mining infrastructure, exited Bitcoin mining entirely, and leased the site to CoreWeave, which committed to the full ~800 MW of approved ERCOT capacity. The asset that mattered was the energized power, not a single mining rack.
Core Scientific (multi-site). The CoreWeave colocation contract was expanded to $10.2B over 12 years, converting ~590 MW to HPC (about 350 MW energized, ~200 MW billing, full site targeted by early 2027). Note the twist most coverage garbles: CoreWeave tried to buy Core Scientific outright and shareholders rejected the ~$9B all-stock deal on October 30, 2025. The market judged the power-and-lease position worth more un-acquired.
These are not mine conversions. They are operators that controlled scalable power and built new AI capacity on it. This is evidence that power-first operators can re-rate into AI infrastructure, not that mining sheds convert.
Applied Digital / Polaris Forge (Ellendale, ND). APLD (formerly Applied Blockchain) was a hosting company, not a self-miner. This is a purpose-built campus, engineered for 400 MW of critical IT load with 1+ GW under load study. Total contracted CoreWeave/hyperscaler revenue runs to ~$16B, and after CoreWeave’s Ellendale debt was refinanced to A3 investment grade (up from BB), APLD restructured the leases around the stronger credit.
IREN / Childress (TX). The cleanest “build new on your power” case: a $9.7B, 5-year Microsoft contract for NVIDIA GB300, but the GPUs go into newly built liquid-cooled data centers engineered for 130–200 kW racks at Tier III-equivalent standards, targeting ~$3.1B of annualized recurring revenue. IREN self-operates GPUs and holds zero BTC treasury by choice.
TeraWulf / Lake Mariner (NY) and beyond. ~$12.8B of contracted, credit-enhanced HPC revenue with Fluidstack across purpose-built new buildings, with the platform expanding toward ~2.9 GW across five sites. HPC leasing overtook mining as the primary revenue line in Q1 2026.
Cipher Mining / Barber Lake (TX). Fluidstack leases the full 300 MW at ~85–90% site NOI margins and $9–10M/MW project cost, on 587 energized acres, with AWS attached at two further sites. Hut 8 belongs here too: a $7B, 15-year Fluidstack lease for 245 MW at its River Bend, Louisiana campus, first data hall targeted for early 2027.
Crusoe. Crusoe began as flare-gas Bitcoin mining, then went all-in on AI, and it did not convert its mining fleet. It sold the entire mining business to NYDIG in March 2025 (425+ modular units, 250+ MW) because those wellhead sites were unconvertible, and built greenfield instead. Its flagship 1.2 GW Abilene campus sits on Lancium’s already-permitted crypto site (reusing the substation and interconnect path, not mining hardware), and it later anchored a separate 900 MW behind-the-meter campus for Microsoft, with dedicated on-site generation plus battery storage, bringing the site to ~2.1 GW. Crusoe’s model is the thesis distilled: control the power (its “bring your own capacity” approach), build the compute plant new.
The Crusoe story also carries a narrow operating-expertise lesson. A West Texas winter event in early 2026 knocked liquid-cooling machinery offline for days and strained the Oracle relationship; Oracle and OpenAI dropped a planned expansion of that specific campus over financing and demand-forecast issues. But the existing campus kept building and Microsoft stepped in on the adjacent expansion. So read this as “liquid-cooled AI creates operational complexity mining never taught,” not as “Crusoe failed.” The physics of 142 kW racks are unforgiving, and mining did not prepare anyone for them.
The following are good mining sites and non-viable AI sites, and they are the majority of installed hashrate.
1. Distributed flare-gas / stranded-gas wellhead mining. Crusoe’s old fleet and Marathon’s ~10 MW containerized edge sites are, per CoinShares, well suited to mining’s interruptible load but incompatible with AI’s continuous-uptime requirement. They fail on scale, firmness (gas supply varies), fiber, water, and remoteness at once. These get sold or run until the ASICs die.
2. Load parked behind intermittent generation (wind/solar-only, no firming). Not to be confused with dedicated behind-the-meter generation, which is firm and AI-friendly. Here the load is co-located behind a wind or solar farm to soak curtailed or negative-priced energy; the entire logic is intermittency. AI-viable only by bolting on the grid tie, storage, or firm generation the site was built to avoid.
3. Interruptible ERCOT demand-response mines. Built to curtail. Curtailment is the revenue line, not a cost (Riot booked ~$56.7M in demand-response credits across the year). A load contracted to be shed on command is the antithesis of an AI baseload tenant.
4. Subscale and non-contiguous sites. Below ~50–100 MW of firm, contiguous capacity you cannot host a modern training cluster. Stitched-together parcels don’t assemble into a campus.
5. Remote hydro and international sites. Remote hydro (small upstate NY, rural Québec) and international geographies (Paraguay — e.g., HIVE’s Valenzuela; Ethiopia, Kazakhstan) offer cheap power and lousy everything-else-AI-needs. Excellent for hashing, orphaned for AI.
The tell: these operators are staying in mining, winding down, or selling, not converting. Even the aggressive pivoters keep a hybrid: AI in front, interruptible mining in back as a flexible balancing load on the same interconnect, precisely because much of the power isn’t AI-gradeable.
The sites most worth converting are the least Bitcoin-native: large, grid-connected, contiguous, fiber-reachable, financeable campuses with firm power. The most Bitcoin-native sites, which are remote, interruptible, modular, flare-gas, stranded-hydro, demand-response-optimized, are exactly the ones that fail the AI screen.
So the correlation runs backwards from the hype. The better a site was for pure-play mining economics, the less likely it is to be AI-grade; the better a site is for AI, the more likely the “mine” was only ever a temporary monetization layer sitting on top of a power asset that was always worth more than the hashing. “Miners pivot to AI” is therefore the wrong mental model. The right one: power-first operators were always going to win, and mining was how some of them parked capital on a power position until the higher-value tenant showed up.
For an institutional reader, “conversion” is really a credit-and-collateral question, and this is where “gigawatt = gigawatt” does the most damage.
The counterparty problem. A trillion-dollar hyperscaler will not backstop mission-critical capacity with a sub-investment-grade miner’s balance sheet. Cipher’s CEO put the objection in the tenants’ own mouths. This, not physics, killed many “planned” conversions.
The fix. The sector solved it with the hyperscaler backstop: when a miner-landlord leases to a mid-tier AI cloud (Fluidstack), a hyperscaler (Google) guarantees a slice of the obligations, up to assuming the lease or paying termination amounts under defined default scenarios, in exchange for warrants. Google backstops ~$1.73B at Cipher (~5.4% stake) and, at TeraWulf, raised its support from ~$1.8B to ~$3.2B for ~14% pro forma ownership. Functionally, this converts miner-lease risk into Big Tech credit risk, which is what let Cipher issue senior secured notes around 7.125%. That is collateral bifurcation in the wild: contracted, IG-backstopped cash flow finances cheaply; naked merchant GPU- or mining-backed exposure does not. The same split shows up at the tenant level, such as with CoreWeave’s Ellendale debt going BB→A3.
The froth. Equity has priced the pivot as a near-certain re-rate rewarding optionality (undeveloped power) over execution (contracted, financed, built MW). That is beginning to invert. Two risks sit underneath: oversupply (everyone pivoting into the same window; if AI capacity overbuilds, the margins justifying the re-rate compress) and inflexibility (unlike mining, AI can’t be curtailed on command, which creates grid and regulatory friction mining’s flexibility used to defuse). Treat “repurposing our shells” claims skeptically; even the touted “power shells” are Tier I being rebuilt to Tier III.
Is the power firm or interruptible? Curtailment-credit / demand-response / behind-intermittent-generation sites are not AI candidates regardless of megawatts. (Dedicated behind-the-meter generation is fine.)
One contiguous 100 MW+ campus with expansion land? Sub-scale, stitched-together capacity doesn’t host a cluster.
Where’s the fiber? Remote flare/hydro fails here silently.
Retrofit or teardown/greenfield? Nearly always the latter — the value is the pad, not the building.
Who’s the counterparty, and who backstops it? No creditworthy tenant (or hyperscaler backstop) means no financeable project.
Is the power leaving mining permanently? Genuinely financed AI capacity isn’t coming back to hashrate in the next bull market — that’s the real signal, versus ASICs merely idled on weak hashprice.
The one-liner: a Bitcoin mine is not a latent AI data center; its power position might be, and for most of the installed base, even that is the wrong kind of power. Everything else is capital, time, expertise, and a creditworthy tenant.
If you enjoy this newsletter, consider sharing it with a colleague.
Share Buy the Rumor; Sell the News
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.