Microsoft cancelled a data centre in Essex. The headline blamed power. The headline was correct and roughly three years late.
Here is the part that should bother you. A hyperscaler walking away from a site is not a company losing a queue position. It is a company that never held one worth keeping. In the UK the scarce asset is not land, not capital, not GPUs. It is a signature on a grid connection agreement, and those signatures were allocated years ago to a crowd nobody was watching.
Crypto miners.
I have been on the wrong side of a cost I failed to model. January 2020, a network spike, a $12,000-a-month arbitrage book that printed $3,500 of losses in under an hour. The strategy was fine. I had priced the trade and forgotten to price the pipe it ran through. Gas was the pipe then.
Power is the pipe now. Latency is just a tax on hesitation, and interconnection queues are latency measured in years.
The UK's connection queue is the structural fact under this story. Queue applications have run into the hundreds of gigawatts against a national peak demand of roughly fifty. Most of that is speculative. Zombie projects holding capacity they will never build, blocking projects that would. Ofgem and NESO have been running a reform programme through 2024 and 2025 to purge the queue and re-rank by readiness, which tells you everything about how bad the backlog got. You do not reform a queue that is working.
Essex sits in the southeast load pocket, the most contested interconnection territory in the country.
The policy layer has not closed the gap. The UK's AI growth zones framework, announced at the start of 2025, was explicitly designed to shortcut planning and prioritise grid connections for AI campuses. Read that again. The intervention exists because the default process cannot deliver. Policy is a signal of where the friction is, not evidence that the friction is gone. A zone designation does not manufacture a transformer, and transformer lead times are running into multiple years across Europe.
Microsoft has said the quiet part publicly for a while. Management has described the binding constraint as power and shell space, not silicon. Capex guidance sits near the $80 billion mark for the fiscal year. That number did not go down when Essex came off the board. The money moved.
Which is where the crypto side enters, and where most of the coverage stopped reading.
Between 2019 and 2021, mining operators signed power purchase agreements and interconnection agreements across Texas, upstate New York, the Pacific Northwest, and yes, the UK and Scandinavia. They were not signing them because they were visionaries. They were signing them because mining was the only business that could monetise a stranded megawatt immediately, at any scale, with a container and a shipping date. Then 2022 arrived, hashprice compressed, credit dried up, and the market marked the entire cohort as broken hardware.
The agreements did not expire. They sat there. Some of them sat inside bankruptcy estates. Core Scientific filed in December 2022 with energised capacity and long-dated power contracts on the books, and emerged holding the same capacity and the same contracts. The asset survived the equity. That is the tell. When a business's contracts outlive its balance sheet, the contracts were the business.
The repricing started in 2024 and it is not finished. Listed miners with energised capacity and expandable interconnect are now signing hosting contracts with AI compute buyers at revenue per megawatt-year that runs several multiples of what the same megawatt earns hashing at current difficulty. My own back-of-envelope: a modern 1 MW footprint at roughly 20 joules per terahash produces about 50 PH/s; at a hashprice near $50 per PH/day that is something like $900k of gross revenue a year per megawatt, before power cost. Contracted AI hosting on the same megawatt, on the deals that have been disclosed, lands several times higher. The delta is the entire thesis.
The miners are not pivoting to AI. They are monetising the only asset that was ever scarce, and it was never hashrate. It was the interconnect.
That reframes the Microsoft decision. Essex was not a referendum on UK AI demand. It was a portfolio sort. When you hold a fixed capital allocation and a harder constraint on megawatts, you rank projects by power availability, time-to-energise, and return per megawatt, and you kill the bottom of the list. The bottom of the list in the southeast of England is anything waiting on a queue that NESO itself admits is broken.
Watch what happens to the word "cancel." In practice these decisions are usually deferrals, relocations, or a swap of one substation for another. Media framing collapses all three into a single verb, and the collapse is where the mispricing lives. I trust the log, not the hype. The log here says capex unchanged, geography changed.
The financial structure of these hosting deals is where the risk actually sits, and almost nobody is underwriting it properly. The typical shape is a ten to fifteen year take-or-pay contract with a single counterparty, often with prepayments that the landlord uses to build the shell. That is a beautiful trade when the counterparty is good for it and a catastrophic one when it is not. The miner has converted commodity exposure into credit exposure. It has swapped hashprice volatility for counterparty concentration, and it has taken on construction risk it has never managed at this density. Liquid cooling retrofits on a building designed for air-cooled ASICs are not a line item. They are a rebuild.
I ran a version of this trade without knowing it. In April 2024, around the spot bitcoin ETF launch, I backtested first-hour arbitrage against the underlying and found a 0.3% dislocation that held. We put $2 million through it and took $6,000. The lesson was not that the edge was large. The lesson was that the edge existed because the plumbing between two markets had not been stress-tested yet. The same plumbing is now being built between the power market and the compute market, and it is being built by companies that were, three years ago, in chapter 11.
Now the engineering layer, because that is where the money gets made and lost.
Three supply paths exist for a high-density AI campus today. Grid interconnection, which is cheap per megawatt-hour and catastrophically slow. Behind-the-meter gas generation, which is fast, expensive, and increasingly constrained by permitting and emissions accounting. And long-duration contracted nuclear, which is the right answer on paper and arrives, at best, in the early 2030s. Small modular reactors are a PowerPoint with a construction schedule attached. They do not solve a 2027 energisation date. Anyone modelling SMR capacity into a 2026-2028 revenue ramp is modelling a press release.
The middle path is the one being walked. Fuel cells, gas turbines, hybrid arrangements that bypass the queue entirely. The UK has been slower to this than Texas, where behind-the-meter generation and demand response are a functioning market rather than a regulatory argument. In ERCOT a mining site is not a consumer of power. It is a call option on the spread between what it earns and what it is paid to stop earning. Curtailment revenue is a real line item. Four-coincident-peak management is a real line item. Operators who modelled it survived 2022. Operators who treated power as a fixed input did not.
I learned that once and only once. The $3,500 hour taught me that the unmodelled variable is always the one that settles the account. Alpha decays faster than the code that finds it, and the decay curve is set by infrastructure you do not control. For a mining operator, that infrastructure is a substation. For a hyperscaler, it is the same substation, plus a planning committee, plus a queue position they do not hold.
The blind spot is where the money hides, and the blind spot here is that these companies are being repriced as power infrastructure while their equities still trade as beta on bitcoin.
I have watched this exact error before. In 2022 I held $15,000 of UST and watched the supply mechanics decouple from the peg weeks before the terminal print. The information was public. The framing was wrong. Everyone was reading the price and nobody was reading the mechanism. Same structure here. The market reads "Microsoft cancels data centre" and prices AI demand. It should read "interconnection capacity is now a tradeable asset" and price the balance sheets that already own it.
Which brings the on-chain layer, and I will be blunt about it. Every cycle produces a wave of tokenised energy and DePIN metering projects claiming to turn a megawatt into a yield-bearing instrument. Most of them are measuring something they do not control. A token that tracks the output of a solar array is a data product. A token that tracks a firm interconnection right is a security. The distinction is not semantic. It is the difference between an oracle reading a sensor and an oracle reading a contract, and the settlement layer is only as honest as the meter behind it. I have audited enough of these to know that the interesting part is never the token. It is who holds the signature with the grid operator.
The consensus interpretation of Essex is bearish for AI infrastructure and bearish for the crypto-AI trade. Both readings are lazy.
If power were abundant, Microsoft would have built. It walked because the queue could not deliver a date it could underwrite. That is not weak demand. That is demand so strong it has hit a physical wall, and the wall is now the moat. Scarcity of megawatts is bullish for anyone who already has them energised.
The second-order effect is uglier and more interesting. AI compute is consolidating onto single-operator sites with a single grid connection and a single counterparty. Every one of these campuses is a centralised node dressed in distributed language. We have spent four years being sold decentralised infrastructure while the actual compute buildout went the other direction, into a handful of substations in a handful of counties. The people who understand that asymmetry are the ones who were running nodes when nobody wanted the power.
Watch the queue reform documents out of NESO, because they set the UK's energisation calendar for the next five years. Watch Microsoft's capex guidance on the next call, because if it holds near $80 billion while a site disappears, the money went somewhere with better electrons. Watch the hashrate-to-megawatt disclosures from listed miners, because that conversion ratio is now a valuation input rather than a footnote. And watch the Essex site itself, because a cancelled project with a live connection application is a saleable asset, not a hole in the ground.
The question is not whether AI needs power. It is whether the market has finished repricing the companies that already own the connection. My read is that it has not, and that the gap between those two prices is where this cycle pays.