The number that matters in Microsoft's capacity leak is not 38 gigawatts. It is the footnote.
On September 11, sources surfaced an internal roadmap: Microsoft intends to move from roughly 12 gigawatts of global data center capacity today to more than 38 gigawatts by 2032 — a tripling in seven years. Capex for the latest fiscal year reached $145 billion, and analysts expect the line to keep climbing. The company has already declined AI and cloud subscriptions in key US and European regions because it could not deliver power to the racks. The plan, as reported, counts self-built and leased facilities. It explicitly does not count compute rented from the neoclouds — CoreWeave and its cohort.
That exclusion is the signal. When a hyperscaler's internal capacity model stops at its own fence line, the binding constraint has moved outside the balance sheet. Signal extraction from the noise floor. The market read this as an AI capex headline. It is a power headline. And power is the substrate Bitcoin mining sits on.
Context: the queue is the scarce good
Data centers do not fail on silicon. They fail on interconnection. Transformer lead times in the United States have stretched to three to five years for large power apparatus. The PJM and ERCOT large-load queues now hold tens of gigawatts of projects that will never energize, because the queue is a first-come instrument and the substation headroom was consumed years ago.
Between 2020 and 2023, the only buyer willing to sign long-dated power purchase agreements at stranded generation sites was the Bitcoin miner. That is the historical fact the AI narrative keeps eliding. Miners went to West Texas, to upstate New York, to the Alberta gas fields, to the Nordic hydro pools, and they signed ten-year contracts at prices no industrial offtaker would touch. They absorbed the curtailment risk. They built the substations. They secured the interconnect.
Mapping the invisible currents of liquidity taught me one thing in 2020 that transfers directly here: depth and resilience are not the same variable. DeFi pool depth looked like safety until it wasn't. The same distinction applies to electricity. A gigawatt of signed interconnect is not a gigawatt of delivered capacity. But it is the only gigawatt available for purchase, and it is already spoken for.
The neocloud layer deserves a precise definition, because the reporting treats it as a footnote when it is the actual mechanism. CoreWeave and its peers do not build. They lease. Their product is not compute — it is speed of delivery. They will sign for capacity at a miner's existing site within a quarter, because the site already has the interconnect, the transformers, and the county's blessing. Microsoft cannot do that. Microsoft's procurement runs through a construction schedule measured in years and a permitting calendar measured in political cycles. The exclusion in the roadmap is therefore not an omission. It is a statement of dependency.
Core: the conversion is a one-way door
Here is the arithmetic, stripped of narrative.
A 200-megawatt site contracted to a neocloud under a twelve-year HPC hosting agreement generates roughly $1.4 to $1.5 million per megawatt per year in contracted revenue — that is the shape of the Core Scientific–CoreWeave structure, scaled. Bitcoin mining at a comparable site, at a hashprice in the $40 to $55 per PH/s per day band, produces something closer to $0.8 to $1.0 million per megawatt per year, entirely unhedged, exposed to difficulty adjustment and to the next halving.
The revenue gap is roughly 1.5x. The margin gap is wider, because mining carries the full volatility of the hashprice while hosting carries a contracted escalator. This is not a close decision. It is a structural arbitrage, and every publicly listed miner with an interconnect already runs it in a board deck.
Now the part the hashrate charts have not absorbed.
When a mining site converts to AI hosting, the ASICs come off the wall. But not all sites are equal. The ones AI wants first are the ones with the cheapest power, the deepest substation headroom, and the cleanest permitting — which is to say, the lowest-cost quartile of the mining fleet. That is the selection effect. The conversion process removes the cheapest power from the mining cost curve first. What survives is a fleet with a higher average cost basis, a higher breakeven hashprice, and a thinner cushion against the next difficulty step.
The ledger remembers what the market forgets: hashrate is not a stock, it is a flow, and it has been repricing continuously toward the surviving sites. A network drawing in the mid-teens of gigawatts can lose three to five gigawatts of its best-sited capacity over four years and still print a flat chart — because new builds elsewhere fill the arithmetic. The headline number will look stable. The cost structure underneath it will not. This is the divergence the difficulty ribbon conceals.
The difficulty adjustment makes the process self-correcting over a decade and violent in the interim. When low-cost capacity exits, survivors earn more per terahash until the next retarget absorbs the windfall — and the retarget arrives every 2,016 blocks. That is a two-week feedback loop operating against a five-year infrastructure cycle. The mismatch is the trade.
In early 2024, I modeled how passive ETF accumulation would strip available circulating supply out of exchange reserves — a 15% reduction that most desks missed because they were watching price, not microstructure. The same mechanism is running here on a different ledger. Passive institutional demand for AI compute is quietly converting the mining fleet's best assets out of the mining pool, and no prospectus will disclose it.
Contrarian: AI does not kill mining — it reprices the grid
The consensus framing is that AI absorbs capital, power, and political oxygen, and that Bitcoin mining is the loser. The consensus is often the contrarian trap.
The two industries are not competitors for capital. They are competitors for interconnect, and the miners already won that race, years ago, at prices that will not be offered again. What AI demand actually does is mark those interconnect positions to market. A miner sitting on 300 megawatts of pre-permitted load in 2022 was carrying an asset the equity market valued at roughly nothing. In 2026, that same asset is a call option on the most inelastic demand curve in the industrial economy.
Patterns repeat, but the participants change. In 2017 the mispriced asset was code integrity. In 2022 it was counterparty opacity. Today it is electrical interconnect under a fixed-price legacy contract. Architecture reveals the true intent — and the architecture of the Microsoft roadmap, with its deliberate exclusion of third-party capacity, tells you that even the largest buyer in the world cannot see its own supply five years out.
The blind spot is on the other side. Every analyst modeling miners as a levered beta on Bitcoin is using the wrong denominator. The balance sheet is increasingly a real estate and power book with a hashprice hedge attached, and the two legs decay at completely different rates.
Takeaway
The 38-gigawatt figure will be revised. The exclusion footnote will not. Watch the interconnection queue, not the hashprice, and watch which operators convert their sites versus which ones sell them outright — that split will define the 2028 fleet. Survival is a function of position sizing, and in this cycle position is measured in megawatts under a signed PPA, not in tokens on a screen.