Microsoft's 38-Gigawatt Roadmap Is Bigger Than Bitcoin's Entire Network

The number that landed on September 11 was 38. Not 38 megawatts, not 38 racks — thirty-eight gigawatts of data center capacity by 2032, against roughly 12 today.

Here is the translation most coverage skipped. Fill one gigawatt with current-generation SHA-256 hardware and you sustain about 57 exahashes per second. Thirty-eight gigawatts is therefore around 2,170 EH/s of equivalent compute. Bitcoin's entire global network has run between 800 and 1,000 EH/s for most of this year. Microsoft's roadmap, denominated in mining units, is more than twice the electrical appetite of every Bitcoin miner on Earth combined.

That is the anomaly. And like most anomalies, it is not really about the headline number.

Where does a figure like 38 GW come from, and what is it worth? On its own, very little. It is a roadmap — a planning document, explicitly subject to revision based on customer demand and technology. Microsoft's capital expenditures for the latest fiscal year reached $145 billion. The plan covers self-built and leased facilities and excludes capacity rented from “new cloud service providers” such as CoreWeave.

That exclusion is the detail I keep returning to. CoreWeave was founded in 2017 as an Ethereum mining company. Before it was an AI cloud, it was a GPU rig pointed at Ethash. It now sits close enough to the center of the compute crunch that Microsoft names it as an outside variable.

This is my beat. In 2017, as a junior analyst at an LA crypto hedge fund, I spent three weeks cross-referencing Ethereum mainnet transaction logs against a token whitepaper and found that 40% of the reported whale movements were internal swaps designed to inflate volume. The lesson was not that the project lied. The lesson was that a claim and a meter reading are two different data types, and markets routinely price the first as if it were the second. Truth is found in the hash, not the headline.

So the useful question is not how large 38 GW is. It is what data actually moves when a number like this lands.

Start with the binding constraint. Data centers no longer fail for lack of GPUs; they fail for lack of interconnection. Average US grid queue waits have stretched past four years in the largest markets, and the entities holding those positions are frequently not utilities. They are miners. Between 2022 and 2025, Bitcoin miners converted themselves into power shells — long-duration energy contracts, substations, permitting already done. Core Scientific, Hut 8, IREN and Galaxy Digital all signed hosting or HPC conversion deals. The hash never stopped; the asset underneath it changed.

That is why this roadmap matters to the sector even though Microsoft touches almost no crypto rails directly. Every gigawatt built or leased must come from somewhere — a queue position, a PPA, a curtailment arrangement. If the roadmap is real, scarcity reprices.

The arithmetic is unforgiving. Twelve gigawatts to 38 means roughly 26 GW of new interconnection across seven years — call it two to four gigawatts annually, against a queue process that in most ISO territories takes longer than that to clear a single large load. Either the roadmap assumes a policy regime that does not yet exist, or it assumes a large share of that capacity is already quietly locked up through existing miner contracts and behind-the-meter arrangements. The second explanation is the one on-chain data can partially test.

Start with hashprice and fleet efficiency. Hashprice — dollars per petahash per day — has compressed through this cycle, and the marginal producer now runs hardware older than 30 J/TH. When hashprice sits below the all-in cost of a 30 J/TH machine, operators have a rational incentive to sell hashrate contracts rather than chase blocks. Watch miner reserve balances: declining reserves against flat hashrate is the signature of fleets being repurposed, not retired.

The cleanest signal is contracted revenue mix. For listed miners, the ratio of contracted HPC/AI revenue to self-mined revenue is now the most predictive line in the filings. An operator at 70% contracted AI revenue has effectively issued a bond against its power shell. An operator at 10% is still a commodity producer. These are not the same asset class, and the market prices them as if they were.

The most contested dataset is tokenized compute. Render, Akash, io.net and a dozen others present themselves as the decentralized answer to precisely the shortage Microsoft describes. Here the on-chain data is unambiguous and unflattering. Network revenue — actual lease payments settled on-chain — is queryable. Silence is just data waiting for the right query. Pull weekly settled compute revenue for each network against circulating market cap, and the ratio tells you what the token is pricing.

I ran a version of this in 2020, mapping impermanent-loss adjustments across 500+ Curve wallets, and found that 15% of yield was being extracted by front-running bots. The protocol was real. The yield was real. The distribution was not what the dashboard said. Tokenized compute has the same structural property: the demand is real, but value capture sits with whoever owns the silicon and the megawatt, not whoever holds the governance token.

Last year I spent six months mapping 50,000+ wallet addresses to entity labels for a large asset manager, reducing ambiguity by 90% so the dataset could survive SEC reporting scrutiny. That work taught me where these dashboards break: attribution. A revenue figure is only as good as the wallet clustering underneath it.

There is a Layer 2 parallel worth flagging for anyone treating decentralized compute as solved architecture. Sequencers on the major rollups remain single-operator systems. A compute marketplace with one matching engine and one settlement layer is the same shape: a centralized chokepoint wearing a decentralized label. I have watched that architecture stay on PowerPoint for two years. The AI compute crunch does not change the topology.

The $145 billion capex figure deserves its own sanity check. At that scale, financing is the constraint, not ambition. Depreciation schedules on AI accelerators are aggressive, and the contracts underpinning them are shorter than the hardware's useful life. That mismatch is a pre-mortem waiting to happen — the pattern my team tracked in 2022, when we flagged $30 million of undercollateralized positions at a lending protocol that took an oracle price at face value.

Here is where I part with the prevailing read. The instinct right now is to treat Microsoft's roadmap as bullish confirmation for anything labeled compute — tokens, miners, DePIN, L2s, all of it, as one trade.

That is a correlation sold as causation, and it fails a basic test. Microsoft building 26 additional gigawatts does not create a single unit of token demand unless a token sits in the payment path — and in almost every case, it does not. The payment path runs through utilities, EPC contractors and hyperscaler procurement. The token is an adjacent claim on the same narrative.

The second error is treating the roadmap as a commitment. The document states the plans may still be adjusted based on customer demand and technological change. That is not hedge language; it is how every capacity roadmap is written. Microsoft has already demonstrated it will pause construction when the demand signal softens — it did exactly that earlier this cycle, which is why some customers migrated to competitors and why new cloud subscriptions were restricted in key US and European regions. A roadmap is a claim. A meter reading is a fact. I know which one I reconcile against.

Next week, I would watch three signals: hashprice's direction against the 30 J/TH breakeven, any new miner-to-HPC conversion filings citing behind-the-meter capacity, and the gap between tokenized compute network revenue and market cap. If that gap widens while the 38 GW headline keeps circulating, you have your answer about what the market is actually buying. The question worth sitting with: when the roadmap is revised — and it will be — which of these assets was ever exposed to the gigawatts, and which was only ever exposed to the story?