From Bitcoin Mine to AI Data Center: The Conversion Math, Line by Line

A bitcoin mine and an AI data center are the same thing to a power company: a large industrial customer with a signed interconnection agreement and megawatts already flowing. That is why 2025 and 2026 turned mine conversions into a sector playbook — over $70 billion in AI and HPC contracts have been signed by converting miners [7], and the two largest, Core Scientific's 590 MW CoreWeave build-out and TeraWulf's Fluidstack leases at Lake Mariner, both started life as mining campuses.

But the conversion decision is not a slogan, it is an arithmetic problem with a brutal asymmetry: mining is a volatile, decaying revenue stream on hardware that depreciates in three years, while AI hosting is a 12-to-15-year contracted lease with investment-grade counterparties — paid for with a capital budget three to ten times the original mining fit-out. This article computes both paths line by line for the same 100 MW Texas site, using only published deal terms and current market prices, and shows exactly where the crossover happens.

Why miners hold the asset AI cannot wait for

The US interconnection queue held roughly 2,600 GW of active requests in 2026 — more than the entire installed generating fleet — with median wait times above five years [4]. A greenfield data center that needs a new substation, a new transmission path and a new large-load tariff is a 4-to-7-year project before the first rack powers on. A converted mine skips most of that: the substation exists, the interconnection agreement exists, the land is zoned and permitted for industrial load, and the operations team already knows how to run a 24/7 high-power facility. That is the whole trade — brownfield conversions reach power in 12–24 months versus 4–7 years greenfield [7][4].

At the same time, the mining half of the business crossed into its worst stretch since the 2021 China ban. Hashprice — daily revenue per petahash, the metric every operator prices against [6] — fell from a $63 peak in July 2025 to a monthly-average record low of $27.7 in June 2026, recovering to about $38 as bitcoin rebounded near $77,000 [4][5]. The network's hashrate dropped 17.5% in the first half of 2026 as roughly 252 EH/s of machines went offline [4], and the average cost to mine one bitcoin reached about $75,500 in Q2 2026 against a quarter-end price of $58,400 — below cash cost for the average listed operator [4]. TeraWulf is deliberately running its mining fleet down as buildings convert [4], and IREN told investors its mining exit would be substantially complete by the end of 2026 [4].

Two curves crossed: the mining revenue curve rolled over, and the AI demand curve outran the grid. The rest of this article is the arithmetic of acting on that crossover.

The per-megawatt revenue ladder

Start with the unit economics. One megawatt runs a different amount of hashpower depending on fleet efficiency: at 25.5 J/TH (a mixed S19-class fleet), 1 MW sustains about 39.2 PH/s; at 13.5 J/TH (current-generation S21-class machines), about 74.1 PH/s. Multiply by hashprice and subtract the power bill (breakeven power = hashprice ÷ (J/TH × 24) [10]):

1 MW × 8,760 h × $0.05/kWh = $438,000/yr power cost — that number is fixed

Use of 1 MW	Gross revenue / yr	Power + opex	Net per MW-yr	Contract?
Mining — S19-class fleet (25.5 J/TH) at hashprice $38	$544k	$438k power	$106k	No
Mining — S21-class fleet (13.5 J/TH) at hashprice $38	$1,027k	$438k power	$589k	No
Mining — S19-class at forward hashprice $30.67	$439k	$438k power	≈ $0 — breakeven	No
AI colocation lease (published deals: $1.44–1.85M/MW-yr)	$1,600k	~20% of revenue	$1,280k NOI	10–15 yr
Operating GPUs yourself	$12,000–16,000k	~$25M+/MW GPU capex first	high, volatile	Spot/short

Table 1 — Mining per-MW arithmetic: 39.2 PH or 74.1 PH per MW at 365 days × hashprice; power at $0.05/kWh all-in. AI lease revenue from signed-deal implied rates (Core Scientific ~$1.44M, TeraWulf ~$1.85M per MW-year); NOI margin 75–85% per company guidance [1][3]. CoinShares independently estimates AI at ~$1.5M profit per MW versus ~$0.5M for mining [4].

Three things the ladder makes obvious. First, an older fleet is already finished: at the six-month forward hashprice of $30.67 [5], an S19-class site nets essentially zero. Second, even a current-generation fleet's $589k per MW is less than half the NOI of a mediocre AI lease — and the lease does not halve every four years. Third, self-operating GPUs looks like the big prize, but it moves you from being a landlord to being a cloud operator with $25M+ per MW of GPU capex and merchant price risk; our earlier per-GPU analysis shows how thin 2-year GPU rental returns already are at market rates (GPU compute rental 2-year ROI). The conversion deals getting signed in 2026 are overwhelmingly the middle row: lease the powered shell, let the tenant bring the silicon.

What the signed deals actually pay

Two transactions define the conversion benchmark, and both disclose enough to check the math.

Core Scientific ↔ CoreWeave: the 590 MW template

Core Scientific is converting its mining campuses (Denton TX, Dalton GA, Muskogee OK, Marble NC, Austin TX) into CoreWeave colocation sites — about 590 MW of HPC infrastructure on roughly 800 MW gross, under 12-year contracts with two 5-year extension options [1][2]. The disclosed terms [1]:

• $10B+ in potential revenue over the contract terms, ~$850M average annual run-rate revenue — an implied $1.44M per MW-year on billable capacity
• 75–80% anticipated profit margin after facilities operations, R&M, security, FTEs, insurance and property taxes
• Client (CoreWeave) pays for capex, power and utilities: up to $1.5M per MW of build-out cost funded by CoreWeave, credited against hosting payments at no more than 50% of monthly fees until repaid
• For a 70 MW expansion, Core Scientific funds $104M ($1.5M per MW) for the powered core and shell with no capex credit — the cleanest public per-MW conversion cost print available

TeraWulf ↔ Fluidstack: the Google-backed lease

TeraWulf signed two 10-year colocation agreements with Fluidstack for 200+ MW of critical IT load (~250 MW gross) at Lake Mariner, its former mining campus in Western New York [3]:

• $3.7B contracted revenue over the initial 10-year terms — implied $1.85M per MW-year — rising to ~$8.7B if both 5-year extensions are exercised
• Expected site NOI margin of 85%, implying ~$315M annually
• Total project cost: $8–10M per MW of critical IT load — the most direct public guidance for conversion capex
• Google backstopped $1.8B of Fluidstack's lease obligations to unlock project debt, taking warrants equal to ~8% of TeraWulf — a stake that has since grown to roughly 14% as Google added support for the CB-5 expansion [3][9]

Both deals price the converted megawatt at $1.4–1.9M per year with 75–85% margins and 12-to-15-year terms. Neither would exist without the mining-era substation — that is the asset being monetized. (CoreWeave's attempt to simply buy Core Scientific outright for ~$9B was voted down by shareholders in late 2025, which tells you how management teams now value their own converted capacity.)

The conversion budget, line by line

A mining hall cost $0.7–1M per MW to fit out: air-cooled containers, interruptible power, minimal redundancy. Converting to AI data center grade means rebuilding almost everything downstream of the substation [3][7]:

Work package	Indicative cost, $M per MW	Why the mining site doesn't already have it
Power redundancy — 2N paths, UPS, standby generators	2.0–3.0	Mining runs on interruptible, curtailment-friendly power with no UPS; AI tenants contract 99.99% uptime
Liquid cooling — DLC ready rooms, CDUs, pumping, dry coolers or towers	2.0–2.5	ASIC halls are air-cooled at 10–20 kW per rack; AI racks want 40–130+ kW with direct-to-chip loops
Structural + shell rebuild — floor loading, clear height, aisle geometry	1.0–1.5	Mining sheds are tents and light steel; AI halls need raised floors or slab ducts, fire separations, containment
Fiber, InfiniBand, network Meet-Me rooms	0.8–1.0	Fiber kills more conversions than anything else — a site 200 miles from long-haul routes stays a mine [7]
Soft costs — design, permits, AHJ approvals, ESG studies	0.8–1.0	New occupancy classification, fire code review, sometimes full re-permitting
Electrical topology — 480V distribution, high-density PDUs, busway	0.5	Mining distribution ends at low-density containers, not 120 kW racks
Total	8–15 (TeraWulf guided $8–10)

Table 2 — Conversion capex per MW of critical IT load. Ranges from TeraWulf guidance ($8–10M/MW) [3] and sector surveys ($8–15M/MW) [7]; the split above is indicative, not a vendor quote. Note the second-order effect: adding UPS, generators and 2N paths consumes floor space, which is why conversions also lose capacity.

That last line deserves its own paragraph, because it is the number that embarrasses uninformed buyers: the megawatts shrink. A "100 MW" mining site carries roughly 45–55% of its nameplate through conversion — gross electrical capacity minus cooling overhead, UPS losses, redundancy (N+1 or 2N), and the power density mismatch between uniform ASIC halls and GPU halls [4]. Our model site yields 50 MW of critical IT from 100 MW gross. Anyone underwriting a conversion on nameplate megawatts overstates revenue by a factor of two.

Timeline: 12–24 months, and the schedule is set by transformer and switchgear lead times, not construction speed [7]. That is the same bottleneck we documented for the broader grid squeeze (US power squeeze and data center power strategy) — a converted site inherits its queue position precisely because the long-lead equipment is already energized.

The 10-year model: same site, two futures

Now run both paths for the 100 MW Texas site. Assumptions, all from published sources:

Parameter	Mining path	Conversion path
Capacity	100 MW IT-equivalent (hashpower)	50 MW critical IT after 50% shrink
Revenue driver	Hashprice path: $38 (2026) → $34 (2027) → halving April 2028 cuts ~50% → cycle repeats [4][5]	$1.6M/MW-yr lease (between Core Scientific's $1.44M and TeraWulf's $1.85M implied rates) [1][3]
Cost	$0.05/kWh all-in = $43.8M/yr; opex beyond power in the margin already reflected	NOI margin 80% (75–85% band) [1][3]
Capex	None in the model — see the warning below	$9M/MW × 50 MW = $450M, spent 60/40 over years 1–2
Revenue start	Immediate	Month 19 (18-month build); mining continues in unconverted halls during the build

Table 3 — Model assumptions. Hashprice path is deliberately conservative-moderate: it uses the current spot ($38), the six-month forward ($30.67) as the 2027 anchor, and a halving haircut in April 2028 with a post-halving recovery cycle modeled on 2024–2026 behavior [4].

Cumulative cash, $M	Yr 1	Yr 2	Yr 3	Yr 5	Yr 8	Yr 10
Mining — S19-class fleet	11	15	−1	−22	−79	−136
Mining — S21-class fleet	59	107	115	154 (peak)	164	133
Convert to AI colocation	−211	−346 (trough)	−282	−154	+39 (crossover)	+167

Table 4 — Cumulative positions from Fig. 3, rounded. Conversion-year cash includes mining income from halls still hashing during the phased build ($59M yr 1, half in yr 2) and a partial year of lease revenue in year 2.

The terminal value line is the one accountants argue about and buyers should not ignore. Recent trades value leased AI facilities near $27M per MW, against under $3M per MW for uncontracted mining capacity; companies with signed AI/HPC leases trade at 12.9× EV/NTM revenue versus 3.7× for contract-less miners [4]. Fifty leased megawatts is therefore a ~$1.35B asset at the end of the model window — while a 100 MW mining site at the bottom of its third post-halving cycle is worth roughly its rig scrap value plus the substation.

What happens to the rigs

Conversion does not delete the ASICs; it re-prices them. Three exits, in declining order of value:

• Sell into the secondary market. The 2026 index prices used 14–19 J/TH machines at ~$4.95/TH and 19–25 J/TH machines at ~$2.11/TH [8]. Our model site's 7.4 EH/s S21 fleet recovers ~$37M — 8% of conversion capex, worth booking but not decision-changing. Used-ASIC prices have compressed precisely because so many megawatts are converting at once [7].
• Keep hashing on the leftover power. The 50% of the site that shrinks out of AI capacity is still a working mine — and every converted megawatt lowers network difficulty for the rigs that remain. 2026 delivered the first six-month hashrate decline since the China ban, with three consecutive negative difficulty adjustments [4]: survivors inherit better per-machine economics.
• Run the hybrid. The most grid-friendly configuration keeps a mining block alive on the substation: ASICs curtail in milliseconds to monetize power the AI tenant has contracted but is not drawing, earn demand-response revenue, and keep cash flowing through the rebuild [7]. The cost is operational complexity — two cooling regimes on one yard.

Core Scientific, for its part, paid $41.9M in Q2 2026 to cancel a 15 EH/s hardware order outright [4] — the cleanest evidence that for converting miners, future hashpower is worth less than the floor space it would occupy.

Six gates decide whether your site qualifies

Not every mine converts. The screening questions, in the order deals actually fail [7]:

1. Contracted power. Interruptible or curtailable supply must become firm supply. What does the interconnection agreement and tariff actually say?
2. Fiber distance. AI training clusters need long-haul or dense metro fiber. This kills more conversions than any electrical issue.
3. Structural fit. Floor loading for liquid-cooled racks, clear height, and room geometry that survives fire code review.
4. A cooling path. Water or dry-bulb headroom for heat rejection at the new density — a 100 MW ASIC hall rejects heat one way; a 50 MW AI campus rejects it another.
5. An inheritable compliance file. Permits, environmental reviews and utility agreements that a hyperscaler-grade tenant's counsel can accept.
6. Capital anchored to a creditworthy tenant. Project debt is sized to the lease, not to the miner's balance sheet — the Google backstop exists because lenders do not trust former crypto miners at $450M a site [3].

Where the math breaks

• Tenant concentration. One tenant on a 12-year lease is the margin story and the risk story. Fluidstack is young; CoreWeave is the counterparty on both sides of half the sector's backlog. If AI demand cracks, conversion-stage sites feel it first [7].
• Overpaying for the shortcut. With leased AI sites trading at $27M/MW and raw mining capacity under $3M/MW [4], powered shells have become their own asset class — and the premium is already pricing in conversions that will never meet the six gates.
• Hashprice regret. The same cycle that makes mining miserable also produces the post-halving recovery; bitcoin at a new high in 2027 would make the un-converted orange line in Fig. 3 look underwritten. This is the real opportunity cost of concrete, and it is why hybrid sites exist.
• Regulatory whiplash. Thirty US states have logged at least 225 data-center moratoria or restrictions, 151 of them still active, and New York imposed the first statewide pause in July 2026 — projects already under construction are generally grandfathered, future phases may not be [4].
• The 2028 halving is a forcing function, not a risk. At the April 2028 halving the block subsidy halves again; on a flat bitcoin price, hashprice follows. The conversion decision gets more expensive in exactly the year the funding gets harder. Sites that convert on 2026–2027 economics are selling into the strong half of the cycle.

The decision, compressed

For a 100 MW site with firm power, fiber and a credible tenant pipeline, the ledger reads: mining nets $0.1–0.6M per MW-year, peaks mid-decade, and requires $90M+ of rig refresh capex nobody has budgeted; conversion nets $1.28M per critical-IT-MW-year on contract, costs $450M, pays back in year 7–8 of a 12–15-year term, and leaves a $1.35B asset behind. The S19 fleet is already at breakeven at forward hashprice — for those sites the question is not whether to convert, it is who funds it. For a current-fleet site on sub-$0.04 power with a bullish bitcoin view, mining still clears the hurdle through 2028 — but the exit window for selling that capacity as "future AI" is the same window. Sell before the halving reprices the option.

If you are evaluating a conversion — as the site owner, the tenant, or the EPC bidder — the hardware side of the equation is the part we work daily: liquid cooling readiness, power density, and the containerized compute modules that drop into converted halls. Send us your site parameters and we will price the compute side of the retrofit. For the module approach in particular, see why modular containers are winning AI buildouts, and for the tenant side of the ledger, the 2-year GPU rental ROI model.

References

1. Core Scientific, Inc., SEC Form 8-K investor presentation — CoreWeave contract summary (590 MW, $10B+ potential revenue, ~$850M avg annual run-rate, 75–80% margin, $1.5M/MW capex funding). investors.corescientific.com
2. Core Scientific, Second Quarter Fiscal Year 2026 Earnings Presentation — CoreWeave buildout status (~590 MW leased, 437 MW billing as of mid-July 2026). investors.corescientific.com
3. TeraWulf Inc., SEC Form 8-K Exhibit 99.1 — "TeraWulf Signs 200+ MW, 10-Year AI Hosting Agreements with Fluidstack" ($3.7B/$8.7B, 85% NOI, $8–10M/MW, Google $1.8B backstop, Aug 14, 2025). investors.terawulf.com
4. CoinShares, "Bitcoin Mining Report — Q2 2026" (hashprice history, hashrate capitulation, AI ~$1.5M vs mining ~$0.5M per MW, $75,500 cash cost vs $58,400 quarter close, $27M/MW leased vs <$3M/MW unleased print, 12.9× vs 3.7× EV/NTM, Core Scientific $41.9M cancellation, IREN exit, 225 state moratoria, 2,600 GW queue). coinshares.com
5. BlockOps Mining, "Bitcoin Hashprice Hits $38 as BTC Tops $77,000" (spot $38.33, forward curve $30.67, per-machine net at $0.08/kWh, Aug 21, 2026). blockopsmining.com
6. Luxor, Hashrate Index — real-time hashprice, difficulty and forward curve reference. hashrateindex.com
7. MillionMiner, "Bitcoin Mining to AI: How to Convert a Mining Farm Into an AI Data Center" (2026) ($8–15M/MW, 12–24 months, six gates, four paths, hybrid operations). millionminer.com
8. Bull Miners, "Why ASIC miner prices fell in 2026" (used-market index pricing by efficiency band, forward-curve payback math). bullminers.eu
9. Startup Fortune, "Google's $3.2 Billion TeraWulf Backstop Shows How AI Data Centers Get Built" (backstop growth to $3.2B, ~14% stake, CB-5 expansion context). startupfortune.com
10. Rax Mining, "Hashprice Explained" (hashprice definition, breakeven formula $/kWh = hashprice ÷ (J/TH × 24)). raxmining.com