Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
On 10 August, Riot Platforms announced a twenty year data centre lease at its Rockdale, Texas campus and declined to name the tenant, describing it only as one of the world’s leading frontier AI labs. The shares fell more than five percent after hours. Then CNBC reported the tenant was Anthropic, and the stock climbed as much as twenty five percent.
Nothing about the contract changed in those few hours. The megawatts were the same, the term was the same, the revenue was the same. What changed was the market’s ability to price the counterparty.
That single session is the most useful thing to happen in this sector all year, because it tells you exactly what investors now think they are buying. Not hashrate. Not Bitcoin exposure. A landlord’s claim on a tenant’s credit.
The Migration Is Not a Rumour Anymore
Cumulative AI and high performance computing contracts announced across the public mining sector have passed seventy billion dollars. That is not a projection, it is signed paper.
The individual deals are large enough to have rerated entire companies. TeraWulf has locked in around $12.8 billion of contracted HPC revenue, anchored by Google backed Fluidstack. Hut 8 signed a fifteen year, $9.8 billion lease covering a 352 megawatt Texas facility built to NVIDIA’s reference architecture. Galaxy Digital committed 800 megawatts to CoreWeave over fifteen years for roughly $4.5 billion. IREN secured a $9.7 billion agreement with Microsoft covering 76,000 NVIDIA GB300 GPUs across 200 megawatts at its Childress campus, and holds no Bitcoin in treasury at all, by choice rather than necessity.
Core Scientific offers the clearest picture of how fast the composition changes once it starts. Colocation revenue reached $136.7 million in the second quarter of 2026, against $10.6 million a year earlier, and now represents roughly 83 percent of total revenue. In July the company added AMD as a second anchor tenant on a fifteen year agreement covering about 530 megawatts and more than $14 billion of base contracted revenue, with options on up to 2.5 gigawatts beyond that. BBAE’s survey of the year puts total leased customer power across its sites near 1.1 gigawatts.
At the smaller end, the same logic applies. Hyperscale Data ceased Bitcoin mining at its Michigan facility on 1 September to free 20 megawatts for an unnamed California AI cloud provider, on an agreement it values above $1.2 billion across twenty years, and plans to sell the mining servers.
The Revenue Mix Is Inverting Inside a Single Year
The scale of the change is easier to see in projections than in headlines.
CoinShares expects mining to fall from roughly 85 percent of total revenue in early 2025 to under 20 percent by the end of 2026 for the companies that have secured AI contracts. S&P Global Market Intelligence models HPC reaching 71 percent of revenue at IREN, up from three percent in 2024, and 71 percent at Core Scientific, up from five percent. TeraWulf lands near 70 percent from a standing start. Cipher is modelled at 34 percent, HIVE around 15 percent, Riot around 13 percent.
Blockchain Council’s summary of those projections draws the right conclusion: these are no longer mining firms with a side business. They are power backed data centre developers that happen to retain some Bitcoin exposure.
The economics explain the speed. CoinShares’ own analysis puts AI contract revenue at roughly three times mining revenue on a per megawatt basis, at operating margins of 80 to 90 percent. Set that against a business where margin is a function of a volatile commodity price and a difficulty adjustment nobody controls, and the capital allocation decision makes itself.
Why It Happened This Year Specifically
The pull from AI is only half the story. The push from mining economics is the other half, and it has been brutal.
Network hashrate briefly crossed one zettahash per second in January 2026. On 19 February the network posted a 14.73 percent difficulty increase to 144.4 trillion, the largest absolute jump in its history. Higher difficulty met a block subsidy already halved to 3.125 BTC in April 2024, and margins compressed across the board.
Then the hashrate itself started falling. By late June, CoinMarketCap reported total network hashrate at 886 EH/s, down 12 percent in a single month and 23 percent below the October 2025 peak, triggering the second largest downward difficulty adjustment of the year.
Hashprice tells the same story from the revenue side. It bottomed near $27.66 per PH per day in late June, recovered to $31.89 by mid August, and reached roughly $38 by late August as Bitcoin traded near $77,200. Shattered’s analysis of the year quotes Twenty One Capital’s Rapha Zagury describing the period as “the first bear market in hash rate” the network has seen.
The balance sheet evidence is starker still. CoinShares’ Q1 report documents listed miners collectively reducing Bitcoin treasuries by more than 15,000 BTC from peak, with Core Scientific selling around 1,900 BTC in January alone, Bitdeer taking its treasury to zero in February, and Riot selling 1,818 BTC the previous December. Across the first quarter, public miners disposed of more than 32,000 BTC, the largest institutional sell off on record.
Companies do not liquidate the asset their entire thesis rests on because a better opportunity appeared. They do it because they need the cash.
What Investors Are Actually Underwriting Now
Here is where the trend gets genuinely interesting, because the risks have not disappeared. They have changed category entirely.
Counterparty credit replaced commodity price. A miner’s old risk was Bitcoin falling. A converted data centre’s risk is a single tenant failing to pay across a fifteen or twenty year term. That Riot share price swing was the market repricing exactly this, in real time, on nothing more than a name.
Leverage replaced volatility. Cipher issued $1.7 billion in senior secured notes, and its quarterly interest expense jumped from $3.2 million across nine months to $33.4 million in a single quarter. TeraWulf carries around $5.7 billion of total debt. As Crypto News observed, these are not mining company balance sheets. They are bets that contracted revenue arrives fast enough and reliably enough to service obligations that dwarf anything the mining business ever carried.
Construction execution replaced uptime. Core Scientific has been flagged for slipping on construction milestones, which matters enormously when revenue recognition depends on energising megawatts on a schedule a hyperscaler agreed to years earlier.
And the transition itself carries a cost that shows up in earnings. IREN’s AI cloud revenue surpassed mining revenue for the first time last quarter, and the company simultaneously wrote down $450.4 million in asset values, largely retired mining equipment. The shares fell. That is the pivot working precisely as designed and still destroying book value on the way through.
The Distinction That Matters Most
Contracted revenue and optioned revenue are being reported in the same sentences, and they are not the same thing.
Hyperscale’s $1.2 billion figure requires the customer to exercise both five year extensions on a ten year base term. The path above $3 billion additionally requires a 32 megawatt expansion option. Core Scientific’s $14 billion AMD base is real, and the 2.5 gigawatts of options beyond it are not yet.
None of this is deceptive. It is disclosed plainly in every announcement. But headline aggregation flattens the distinction, and the aggregate figure is what moves the share price.
The discipline worth applying is simple. Separate base term from extensions. Separate committed megawatts from optioned megawatts. Separate energised capacity from announced capacity. What remains after that subtraction is the number the business can actually be valued on today.
What Would Break the Thesis
The bear case is not that the deals are fake. It is that the conversion is one directional.
A miner that has sold its ASICs, written down its fleet and committed its power to a single tenant has no hashrate to fall back on. If AI infrastructure demand cools before the buildouts complete, these companies hold a building, a substation and a contract, and the optionality that once let them switch machines on and off with the Bitcoin price is gone.
Watch for the tell. ChainUp’s analysis makes the point that AI data centres are now bidding for the same grid capacity, the same substations and often the same real estate that miners historically claimed with cheap or stranded power. That competition is what makes the megawatts valuable. It also means the sector’s fortunes are now correlated to a capex cycle it has no influence over.
The counter argument deserves equal weight. Goldman Sachs and JPMorgan both project hyperscaler capital expenditure rising for the next decade. If that holds, the miners that moved early will have converted a commodity business into an infrastructure annuity, and the sceptics will have missed one of the cleanest rerating trades available.
The Question Nobody Is Pricing
There is a longer arc here that sits outside any individual company’s income statement.
Bitcoin’s security depends on people finding it worthwhile to point electricity at SHA-256. If the marginal megawatt in every jurisdiction with cheap power now has a bidder willing to pay three times more for it at ten times the margin, the network’s long run security budget is competing against the entire AI industry for the same input.
This publication has already traced that rotation at the portfolio level, where the argument was that Bitcoin spent eighteen months competing with semiconductors rather than with the dollar, and lost badly enough that the marginal speculative dollar stopped showing up. What 2026 added is the industrial version of the same trade. The competition is no longer only about where discretionary capital chooses to sit. It is about where physical power gets delivered, and megawatts committed under fifteen year leases do not rotate back the way a position does.
Difficulty adjusts, so the chain keeps producing blocks regardless. But it produces them behind progressively less energy, and it becomes progressively more concentrated among whoever retains the cheapest power. That is a structural question about Bitcoin itself, arriving disguised as a story about data centre leases, and almost nobody covering the sector is treating it as one.
For now the trade is legible enough. These companies should be read as power developers with tenant concentration risk and rising leverage, valued on energised megawatts and contracted base revenue rather than on hashrate or Bitcoin beta. The market figured that out in a single after hours session in August. The rest of the sector’s coverage is still catching up.
This is analysis, not investment advice.
