The number that circulated on September 11 was 38 gigawatts. Microsoft, per sourced reporting, intends to operate or lease that much data center capacity by 2032 — up from roughly 12 GW energized today, and carrying a $145 billion capex line from the most recent fiscal year. For scale: the entire Bitcoin network, every ASIC in every hosted container from Rockdale to Keflavík, draws an estimated 15 to 20 GW depending on which model you trust. One hyperscaler's roadmap now exceeds the security budget of a two-trillion-dollar monetary network.
That is the trade. Not "AI is a bubble." The trade is that hyperscaler capex and Bitcoin hashrate are bidding for the same three inputs — interruptible power, high-voltage interconnection rights, and liquid cooling at scale — and only one of the two bidders has a governance process for when the auction clears above the marginal producer's cost of production.
What Microsoft Actually Committed To
Read the disclosure like a term sheet, not a press release. Three clauses carry the information.
Clause one: the 38 GW roadmap mixes self-built and leased facilities. Self-build is a capital commitment with a depreciation schedule. Lease is an operating commitment with a counterparty. Blending them into a single gigawatt figure produces a number that no auditor would sign.
Clause two: the 38 GW excludes computing resources rented from "new cloud service providers" — CoreWeave and the rest of the neocloud cohort. This is the carve-out that matters. Microsoft is stating, in plain language, that its reported footprint understates its effective compute exposure by an undisclosed multiple. The number that circulates is not the number that binds.
Clause three: the plan is explicitly adjustable against customer demand and technology shifts. Microsoft has already suspended certain data center constructions and restricted new cloud subscription sign-ups in key U.S. and European regions when power — not silicon — became the binding constraint. Customers routed to competitors. That is a demand signal, not a supply signal.
So: capacity announced and capacity energized are different ledgers. Every analyst model I have reviewed this cycle treats them as the same line item.
The Power Bid Reprices Every Merchant Miner
Start with the actual clearing mechanism. In PJM, the Base Residual Auction sets the capacity price that every load-serving entity pays. Constellation's restart of Three Mile Island for Microsoft — roughly 835 MW on a 20-year power purchase agreement, reportedly north of $100/MWh — resets the anchor. A hyperscaler willing to sign a two-decade fixed-price PPA at that level does not care about spot merchant power. It cares about firm, carbon-attributed, dispatchable capacity adjacent to fiber.
A Bitcoin miner in the same RTO paying $35/MWh under a legacy contract is not competing on ASIC efficiency against that bidder. It is competing against a willingness to pay three times merchant power and absorb the delta as an operating expense line. When the miner's contract renews, the $/PH collapses regardless of what the hardware does.
Run the arithmetic on hashprice — revenue per petahash per day. Post-halving, the block subsidy is 3.125 BTC. At a $100,000 BTC price and a network near 800 EH/s, hashprice sits in the mid-$40s per PH/day before fees. Breakeven for a fleet running at 20 J/TH with power at $0.05/kWh lands around $38 to $42/PH/day. The entire industry is operating on a 10 to 20 percent margin above shutdown. Apply a 20 to 30 percent power repricing at renewal and you have forced capitulation on schedule, not on sentiment.
The Miners Already Read the Tape
Through 2024 and 2025, listed miners signed high-performance computing hosting agreements — Core Scientific to CoreWeave, TeraWulf to Fluidstack, plus Hut 8, IREN, Galaxy Digital. They are converting energized megawatts into AI inference capacity because revenue per megawatt in that business runs three to five times mining revenue and contracts for ten to twelve years.
Call it what it is. That is a hashrate exodus, executed quietly because the narrative is friendlier than the mechanics. The hardware securing the network is being repurposed to serve inference. The security budget does not shrink immediately — it thins at the margin, in the exact cohort that used to absorb difficulty spikes.
I ran this decomposition in 2025, structuring a delta-neutral hedge for a $5 million institutional client using Ethereum call spreads. The reporting template I standardized strips directional bias and shows only Vega and Theta. The same discipline applies here. When a client asks whether AI capex is bearish for crypto, the correct answer is not a view. It is a partition. Which crypto cash flows are power-short, and which are power-long?
Asset owners with owned interconnection rights are long the constraint. Operators on merchant power are short. GPU-aggregation tokens are structurally short, because their unit economics assume a slope of the supply curve that the hyperscalers have already purchased.
The DePIN Blind Spot
Render, Akash, io.net and their cohort aggregate idle GPUs. The pitch assumes the constraint is GPU supply. It is not. The constraint is energized power and interconnection queue position.
PJM's interconnection queue runs four to seven years deep. ERCOT clears faster but has seen capacity prices print above $270/MW-day. An idle H100 sitting in a colocation rack behind a 100-amp residential panel is not firm capacity. Aggregating a million of them yields a theoretical, non-dispatchable, high-latency pool that no enterprise buyer will underwrite for a training run measured in weeks.
Audit the code, then audit the intent. The DePIN compute thesis is a supply-side story priced on demand-side assumptions that were never verified against the interconnection queue.
There is a second structural defect: liquidity fragmentation. Every new chain that hosts a compute marketplace splits the order book further. Ten venues with $2 million of depth each are not equivalent to one venue with $20 million. Routing latency compounds, spreads widen, and the market that was supposed to price compute efficiently ends up pricing nine inefficiently. More interoperability protocols have not solved this. They have industrialized it.
Where the Consensus Has the Sign Wrong
The prevailing read is straightforward: AI capex is frothy, risk assets are correlated, therefore de-risk crypto. That is a lazy beta, and in two places it has the wrong sign.
Hyperscaler capex is a bid for a constrained input that crypto producers also consume. A capex slowdown is therefore not uniformly bearish. It is bullish for asset holders with locked interconnection rights, because removing a marginal buyer at the auction lets the incumbent clear cheaper. It is bearish precisely for operators whose underwriting assumed cheap power renewal — the same cohort that drew the highest multiples on 2021-era fleet expansion.
Second blind spot: the 12 GW to 38 GW comparison treats a gigawatt as a fungible unit. It is not. A gigawatt of continuously energized IT load at 1.15 PUE with N+1 redundancy in a Tier III facility bears no resemblance to a gigawatt of nameplate, and neither resembles the size of an interconnection request. Press figures blend critical IT load, gross facility load, and queue position — three different denominators. The same blending error destroyed models of Bitcoin mining capacity in 2021, when "EH/s" referred to hashboards purchased rather than hashrate energized. Ledger books, not feelings, settle the debt. 38 GW is a purchase order. Only the energized substation is a balance sheet item.
Third: the neocloud carve-out is the actual information content of the story. Microsoft is disclosing that its reported footprint understates its effective compute exposure. Nobody modeled the default-search payments between Apple and Google either, until the number surfaced in litigation. Same shape, different sector.
Three Prints to Watch
Ignore the headline. Watch three numbers.
One: the PJM Base Residual Auction clearing price for the 2027/2028 delivery year. A print above $300/MW-day reprices every merchant miner in the RTO and pulls forward the next wave of hosting conversions.
Two: hashprice in $/PH/day. Sustained prints below $40 force capitulation in the 20 J/TH cohort, and difficulty adjusts down behind it. That is the mechanical floor, not the sentiment floor.
Three: the spread between signed hyperscaler PPA prices and merchant forward power. That spread is the price of the constraint. It is currently widening, and it is the cleanest read available on who actually controls the next decade of compute.
Liquidity dries up when confidence breaks — and the confidence priced into this roadmap rests on a planning document, not a meter reading.
