The Grid Said No
Microsoft walked away from Britain's flagship AI data centre. The chips were ready. The power wasn't. Get used to that sentence.
What happened
On Sunday the Telegraph first reported that Microsoft had pulled out as the expected anchor tenant of Nscale’s planned AI data centre in Loughton, Essex (Telegraph; Guardian).
Starmer’s government had billed it as the “largest UK sovereign AI datacentre” (Guardian). Nscale pledged £2bn to it (BBC). Microsoft was its only publicly announced customer (Guardian).
The problem was the grid. UK Power Networks told Nscale it couldn’t supply enough power until the early-to-mid 2030s (Guardian). Nscale told the BBC it is “exploring alternative power sources” and that the “London-wide grid issue has no impact on our other contracts.” Microsoft declined to comment (BBC).
Why it matters
For years, AI was a story about chips. Lately it’s a story about wires. Wires take time, and time is the one input nobody can buy more of.
Loughton isn’t alone. The FT counts at least 47 large US data-centre projects delayed, withdrawn or blocked this year, and Bernstein thinks only about 35% of the US pipeline is credible (Hindustan Times, summarizing the FT). Data-centre executives say some regulators now refuse grid connections unless a project brings its own power (same).
Credit is tightening, too. CoreWeave, one of the leading neoclouds, carries about $50bn of debt, and its 2031 bonds yield about 16% (Reuters).
Now look at Nova Scotia. On Friday, Anthropic leased the first 95 MW phase of an EverWind campus at Point Tupper for at least 20 years, due late 2027 (CBC). It is Anthropic’s first in Canada (Reuters). EverWind plans four wind farms, though the provincial grid will be the main supply until they’re built (CBC). The partners have pledged to pay all connection and upgrade costs so existing customers don’t see rate hikes (Globe and Mail).
One site waited on the grid. The other showed up with a plan and the checkbook to pay for it.
The investor angle
Who benefits. Whoever controls power that’s ready to use, or can pay to make it ready. Hyperscalers with big cash flows (Microsoft, Alphabet, Meta, Amazon) are far less exposed to the credit squeeze (investingLive, via FT).
Who carries the risk. Builders that need outside capital: CoreWeave (CRWV), Nebius (NBIS), IREN, Cipher (CIFR), TeraWulf (WULF) (Reuters; investingLive; tickers: Portfolio Terminal). Second-order: Nvidia (NVDA), SK Hynix, Samsung and TSMC if orders slip (investingLive).
Private. Nscale is aiming for a New York IPO at about $35bn, which bankers call hard near-term (HT/FT). Also EverWind and Anthropic.
Timeline. Loughton: the early-to-mid 2030s (Guardian). Point Tupper: late 2027, then 325 MW more in 2028, subject to conditions (Globe and Mail). Dates slip. Patient money plans for that.
SI 101 — Neocloud. A newer cloud company that rents out Nvidia computing power to AI developers. It spends heavily up front on chips and buildings, often with debt (Reuters). A landlord for intelligence, with a mortgage.
What this looks like by 2036
Our judgment, not fact. The next decade of AI may be decided as much by utilities and project finance as by model labs. By 2036, the scarcest asset in AI could be a signed grid connection with clean power behind it. Grids can speed up and efficiency can cut demand, so hold this view loosely. Still, this week the market got a good, clear look at what scarce means. It’s worth remembering.
Briefs
Anthropic sets an IPO clock
Anthropic plans to file publicly with the SEC on Oct 26, start its roadshow Nov 9 and price the week after. It’s aiming to raise up to $100bn at a $2trn valuation. All of that comes from one person familiar with the process, so treat the dates as a plan (IFR). Third-quarter revenue was $16.8bn, up 45% on the quarter, after more than 140% the quarter before. New York City’s comptroller has asked the company to pause (same).
China’s rare-earth truce has an end date
The rare-earth export controls China announced last October are suspended until Nov 10, 2026 (CIRS). That’s three weeks from today. On Oct 15, China’s commerce ministry said it would keep approving rare-earth export licenses for the EU, and asked Brussels to return the favor on dual-use cases (China Daily). The suspended package includes controls on magnet-making technology, which motors and robots need. What happens on that date matters well beyond China.
Quantum computing’s costliest step gets cheaper
Useful quantum computers need “magic states,” and making them has been enormously expensive. Some estimates put it at up to 95% of a machine’s capacity. Google says it has now shown a cheaper method, “cultivation,” working on its Willow chip, with errors about 40× lower (Google; Science). The catch: only about 8% of attempts pass every check (Phys.org). It’s a proof of principle, not a product.
Starship asks to fly a lot more
SpaceX is currently allowed 25 Starship launches a year from Starbase, Texas. The FAA has opened an environmental review of a plan for up to 100 a year, and eventually up to 456 a year on weekdays only, “within several years” (Federal Register). Public meetings are Nov 17–19. If rockets become routine, getting things into orbit could become something that’s just there, like the grid.
Primer: Day 1 · A Model Is a File
SI Primer · Monday, October 19, 2026
Let’s start simple, with the thing itself.
An AI model is a file.
That’s it. A big file, sure. But a file. You can copy it. You can download it. You can put it on a drive and hand it to a friend, if your friend has a lot of room.
What’s inside is numbers. Billions of them. Each one is called a parameter, a little dial that got nudged, a tiny bit at a time, while the model read an enormous pile of text. For Beam, below, that was 23.8 trillion tokens, roughly word-pieces. Reflection Nobody set the dials by hand. They got set by learning. That’s the strange part, and it’s worth sitting with for a minute. Nobody programmed the intelligence. It grew into the numbers.
Here’s this week’s example. On October 16, a US lab called Reflection released a model called Beam on Hugging Face. It has 501 billion parameters. At full precision the file is about one terabyte. Squeezed down to 4 bits per number, it’s about 280 gigabytes. Hugging Face
A day earlier, Meta put out Muse Spark Open. 816 billion parameters. About 1.65 terabytes. Meta; Meta docs
So a frontier-grade brain now fits on a stack of hard drives that’d sit easy on a bookshelf.
Two things follow, and they pull in opposite directions.
One. The file is cheap to copy. Once it’s out, it’s out. Anybody with the hardware can run it. That’s why prices fall when good open models show up. Bank of America says average token prices are down about half since summer. Investing.com
Two. The file is expensive to make. Reflection says it ran 10,500 Nvidia GB300 GPUs for four weeks for just one stage of training, the reinforcement-learning run. Reflection The copy costs pennies. The original costs a fortune.
And a file doesn’t do anything sitting on a shelf. To answer a question, it has to be loaded onto chips, fed electricity and cooled. Every time. That’s where the money shows up next. It’s tomorrow’s lesson.
Why it matters for investors
When someone says an AI company is worth a trillion dollars, ask the simple question. Is the value in the file, or around it?
The file can leak, get matched or get given away. What’s harder to copy is what surrounds it: the compute to make the next one, the power to run it, the customers who use it by default, and the trust that keeps them there.
Anthropic is expected to file publicly for its IPO next week, according to IFR. IFR Keep that question in your pocket when the prospectus lands. Check whether the moat is the model, or everything around the model.
Copies are cheap. Making the next one isn’t.
Tomorrow: training vs. inference, the difference between building a brain and renting one out.
Signal to watch
Oct 26: Anthropic’s public filing, if it lands on the date IFR reports (IFR). Skip the headline valuation at first. Go to the parts about compute and power: how much capacity is signed, where it is, and when it arrives. After this week, those pages may say more about the next decade than the revenue line does.