Bitcoin Miners: $9.7 Million/MW in AI vs. 0.5 Mining

The contract IREN signed with Microsoft yields approximately $9.7 million per megawatt per year; that same megawatt, mining Bitcoin at mid-2026 compute prices, generates around half a million — a nearly twenty-fold difference that explains why miners are selling the Bitcoin they once claimed to hoard. IREN —formerly Iris Energy, an Australian listed miner— closed a five-year, $9.7 billion deal with Microsoft in November 2025 to host NVIDIA GB300 GPUs (Nvidia's most advanced AI chips) across 200 megawatts of liquid-cooled data centers, projected to generate $1.94 billion annually with a project-level EBITDA margin of 85%. The entire public mining sector has already announced over $70 billion in AI compute contracts. The breakdown is simple, yet rarely published in the same unit: revenue per megawatt in AI versus revenue per megawatt mining, with the arithmetic in plain sight. This report does the math, measures what portion of the revenue for five listed miners already comes from AI, and identifies the risk that the growth narrative often ignores: five-year contracts and brutal client concentration against a compute demand that no one can guarantee.

Why are Bitcoin miners becoming AI landlords?

The answer lies in a comparison of margins. The hashprice —the daily revenue generated by a unit of computing power mining bitcoin, measured in dollars per PH/s per day— spent much of 2026 hovering around 30 dollars, with readings fluctuating between approximately 29 and 35 depending on the window and the aggregator (32.34 dollars on July 20, according to the Luxor index). At these levels, bitcoin mining is a thin-margin business: revenue barely covers the electricity bill for the least efficient equipment, and the 10.09% drop in difficulty in June 2026 was precisely the on-chain record of many marginal miners shutting down machines, as we reconstructed in the analysis of the July 11 difficulty adjustment.

While that margin was compressing, a different client came knocking: AI hyperscalers —Microsoft, Google, Amazon, and GPU cloud providers like CoreWeave— need what miners already have and what is scarcer than chips in 2026: contracted megawatts, substations, and land with grid access. Swapping ASIC (specialized circuits that only know how to calculate hashes) for GPU racks turns that same power connection into an AI data center that bills for hosted capacity.

The test case is IREN. On November 3, 2025, it announced the Microsoft deal; on May 14, 2026, it closed a $3 billion convertible bond offering (1.00% coupon, maturing in 2033) to finance the conversion, and to purchase the GPUs, it signed a $5.8 billion order with Dell for over 50,000 B300 GPUs, aiming for a fleet of 140,000 to 150,000 GPUs by the end of 2026 and a total projected annualized revenue of around $3.4 billion once fully deployed. Mining, which used to be the core business, is starting to become a footnote.

How much does a megawatt yield mining versus leasing it to AI?

It is worth doing the math transparently so the reader can replicate it. The Microsoft contract projects $1.94 billion annually over 200 megawatts: 1,940 ÷ 200 = $9.7 million per megawatt per year in the AI business, with an 85% EBITDA margin according to IREN's guidance.

Now, the same megawatt mining Bitcoin. A modern, efficient fleet operates at around 21 joules per terahash (J/TH: the energy the equipment consumes for every trillion hash operations). With that efficiency, one megawatt —one million watts, divided by 21— powers about 47.6 PH/s of compute power. At the mid-2026 hashprice of roughly $30 per PH/s per day, that amounts to about $1,430 daily, or approximately half a million dollars per year per megawatt, and that figure is gross revenue: before paying for electricity, which in many operations consumes half or more. The numerical conclusion is impossible for any miner's CFO to ignore: a megawatt dedicated to AI generates sixteen to nineteen times more revenue than one dedicated to mining, and it does so with an operating margin that mining hasn't seen even in its best week.

This comparison has limits that must be addressed, otherwise it is mere propaganda. The first is capital: $5.8 billion in GPUs for 200 megawatts equals about $29 million in capex per megawatt just for chips, not counting data center construction. Converting a mining warehouse into an AI center is neither free nor instantaneous, whereas an ASIC costs a fraction of that per megawatt. The second is the nature of each revenue stream: hashprice is a volatile number recalculated daily that can bounce upward when the Bitcoin price rises, while an AI contract fixes revenue for five years —more predictable, but also more rigid and tied to a single payer. Even with these caveats, the distribution over the life of the contract favors AI: $9.7 million a year for five years is $48.5 million in revenue per megawatt against $29 million in GPU capex, and the 85% EBITDA does the rest of the math.

What portion of each miner's revenue already comes from AI?

The sector is not moving at a uniform speed. Some have already converted the majority of their balance sheets; others maintain mining as a cash cow while keeping the option to pivot toward AI when the right contract appears. The projected snapshot for 2026, based on the latest press releases and earnings calls available as of this writing, is as follows —it should be read as a trend indicator, not as audited year-end figures.

The extremes define the range. TeraWulf already derives more than 60% of its revenue from high-performance computing leasing, with a contracted backlog backed by Alphabet's credit through Fluidstack; on July 6, 2026, it also added a 19,000 million dollar, twenty-year lease with Anthropic—the largest individual contract in the sector to date—bringing its cumulative backlog to around 31,800 million. The term is striking: twenty years, when the standard for these agreements is five, is a signal that a top-tier AI lab is willing to commit to computing capacity over a long decade. A company that only mined Bitcoin two years ago is now an infrastructure lessor for two of the world's largest GPU buyers. On the other side, CleanSpark signed a twenty-year lease on July 14, 2026, for 175 megawatts of its Sandersville (Georgia) campus with a global tech giant it declined to name—the press had linked previous negotiations to Meta—for approximately 6,600 million dollars in contracted revenue (up to 11,600 million if the two five-year extensions are exercised), with deliveries starting in the fourth quarter of 2027; meanwhile, its revenue remains almost entirely Bitcoin. MARA, led by Fred Thiel, presented itself in its first-quarter call as a "digital energy" company rather than a miner, and acquired control of Exaion (64%), the computing subsidiary of the French utility EDF, as well as the Long Ridge power plant (505 megawatts) to have its own megawatts to repurpose. Collectively, listed miners could derive up to 70% of their revenue from AI by the end of 2026, compared to around 30% at the beginning of the year.

Why are they selling the Bitcoin they claimed to hoard?

The optimistic version of the pivot usually avoids this point. For years, the thesis for listed miners included a second engine: accumulating the Bitcoin produced on the balance sheet as a reserve, mimicking the MicroStrategy treasury model, to give shareholders leveraged exposure to the price without buying the coin.

The AI conversion breaks that argument because of cash requirements. GPU data centers cost billions upfront —recall the $5.8 billion order to Dell for chips alone— and there are two ways to pay for them: issuing debt or equity, or selling Bitcoin from the balance sheet. Core Scientific explicitly chose the latter: it sold 2,385 Bitcoins for $208.3 million in the first quarter of 2026 and announced it would liquidate most of its reserves during the year to finance its expansion with CoreWeave, with which it has approximately $10 billion in contracted revenue. The company that claimed to hoard Bitcoin is selling it to stop being dependent on it.

The contrast with pure treasuries is instructive. Companies like MicroStrategy sell Bitcoin —when they do— due to liquidity pressures or capital structure issues, a dilemma we analyzed in the Saylor trilemma and the broken mNAV discount loop. The pivoting miner sells for a different, colder reason: it has a business in front of it that yields sixteen to nineteen times more per megawatt, and accumulated Bitcoin is the cheapest fuel on hand to start the engine.

Where is the risk that the growth narrative doesn't mention?

The story is almost always told as an upward curve —85% margins, five-figure contracts, a sector beating Bitcoin in the stock market— and that version omits three tensions that should be named out loud.

• Client Concentration. The AI revenue for these companies is not diversified: it hangs on a handful of names. IREN depends on Microsoft; Core Scientific on CoreWeave; TeraWulf on Google and now Anthropic via Fluidstack. At Core Scientific, the dependency is already quantifiable: nearly 67% of Q1 2026 revenue came from a single client, compared to 11% a year earlier. If one of those clients renegotiates, delays deployments, or switches providers, it's not just a business line that suffers: nearly the entire new business does.
• Fixed Terms vs. Uncertain Demand. Contracts are for five years; the AI compute demand that justifies them is not guaranteed for even a quarter. The 85% margin assumes that the appetite for GPUs remains as tight in 2028 as it is in 2026. If the AI investment cycle cools —as other overbuilt infrastructure cycles have before— those who committed to $29 million per megawatt in capex will be left with racks to amortize and a shorter line of clients to fill them.
• Balance Sheet Leverage. The conversion is paid for with convertible debt, reserve sales, and equity raises. The $3 billion convertible bond issuance at 1% that IREN closed in May 2026 covers only a portion of the planned capex; the rest depends on sold Bitcoin and future rounds arriving on time, and the company itself estimated a funding gap of billions to complete its plan. A stumble in deployments or the stock price suddenly makes the next round much more expensive.

None of these tensions invalidate the pivot; they condition it. A capex of $29 million per megawatt just for chips, financed with debt and the sale of Bitcoin reserves, marks the true scale of the bet that the upward curve doesn't show.

For a CFO, with 2026 numbers in hand, the decision to convert megawatts to AI is hard to argue against: sixteen to nineteen times more revenue per unit of power, a margin mining cannot reach, and clients with top-tier credit willing to prepay parts of the contract. What isn't obvious is that the bet will pay off for everyone: converting requires capital that many miners don't have without going into debt or selling their Bitcoin, expertise in operating GPU data centers that mining doesn't provide, and AI demand that must hold for the five-year contract term. The miner who executes the transition well bills like AI infrastructure; the one who arrives late is left with data center debt and thin-margin mining.

What to watch in the next quarterly results?

The Q2 2026 earnings season —which several of these companies publish between late July and August— is the first opportunity to contrast guidance with reported figures, and there are three numbers that separate the signal from the narrative. The first is the actual weight of AI in revenue: not future contracts, but what was billed in the quarter. The second is the Bitcoin sell rate: how much each miner liquidated and at what price, as it reveals how much their AI plan depends on their own reserves. The third is deployed vs. promised capex: GPUs installed and in production versus those announced, which is where delays first appear.

Read with the megawatt breakdown in mind —$9.7 million in AI versus half a million mining— these three numbers tell whether the sector is executing the transition or just announcing it. The Bitcoin these companies are selling today to finance it is the most honest signal that they have already chosen which of the two businesses they prefer.

Sources and links: IREN — $9.7B Contract with Microsoft (Press Release) · DataCenterDynamics — Microsoft-IREN $9.7B · The Block — Core Scientific to sell Bitcoin in 2026 · CoinDesk — Core Scientific sold $208M of BTC in Q1 · News.Bitcoin.com — TeraWulf and AI Contracts · CoinInsider — $70B in Sector AI Contracts