Microsoft’s 23,040-Chip Data Center Stalled on a Power Problem
A grid delay stretching into the 2030s shows why electricity is becoming the scarce resource in the AI buildout
Listen to the audio version of this article (generated by AI).
Money can buy almost anything in this AI boom.
It bought Nscale a plot of land in Loughton, a quiet commuter town on the eastern edge of London. It bought a lease commitment from Microsoft (MSFT), the deepest-pocketed anchor tenant on Earth. It reserved 23,040 of Nvidia‘s (NVDA) top-shelf GB300 chips – silicon that entire nations are lobbying to get their hands on.
The one thing it couldn’t buy was electricity.
Over the weekend, news broke that Microsoft has walked away from its deal to lease Nscale’s planned 50-megawatt data center at its Loughton site. The facility was supposed to go live around the end of this year. The timeline slipped. Then, according to reporting from the Telegraph, grid officials delivered the news that no amount of capital could fix: sufficient power might not reach the site until the 2030s.
The chips and the customer are ready and waiting. The demand, as we’ll see in a moment, is so intense that Microsoft is signing multi-billion-dollar compute deals as fast as it can. And yet a marquee AI project on the outskirts of one of the world’s great financial capitals just stalled for want of electrons – with a wait measured not in quarters but in presidential terms.
We think this story is one of the most instructive AI headlines of the year.
Why Microsoft Walked Away From Nscale’s Loughton Data Center
Nobody canceled the AI boom this past weekend.
Nscale says it remains committed to the Loughton project, is exploring alternative power sources for the site, and that its other contracts with Microsoft are unaffected. Microsoft, for its part, says the decision concerns this single site and that its broader partnership with Nscale is strong.
What happened in Essex is narrower and more interesting: a sophisticated buyer looked at a facility with land, chips, and capital all secured – and deemed that without a reasonable path to power, none of the rest counted.
The episode stings extra for Nscale because the company is preparing for a U.S. IPO, with Microsoft and Anthropic making up the bulk of its contracted business. Bloomberg has reported that its flagship facility in Narvik, Norway, is running about six months behind schedule, with just 6.4 megawatts of a planned 30 live so far.
That makes Nscale a great case study in the defining gap of this phase of the AI trade: between announced megawatts and powered ones.
The AI Data Center Bottleneck Is Moving to the Grid
Delivering them can take most of a decade. And it’s inside that mismatch that the real AI infrastructure story is playing out.
Across the industry, hyperscalers plan to pour roughly $400 billion into terrestrial data centers in 2026 alone. New announcements arrive weekly – gigawatt campuses, sovereign AI zones, “Stargates” of every nationality.
Meanwhile, as far as substations and transmission lines go, interconnection queues in Northern Virginia, California, and Germany now stretch seven to 12 years. U.S. data centers are on track to add 240 terawatt-hours of electricity demand by 2030, and utilities are fielding connection requests that, together, exceed anything the grid has absorbed since the introduction of air conditioning.
And before a single AI workload can come online, every planned campus has to survive the same gauntlet: find the site, win the permits, lock in a grid-connection date that doesn’t drift into fantasy, arrange backup or onsite generation, and build out the cooling. Only then does ‘capacity’ mean anything at all.
Capital clears the first hurdle easily. Chips clear the second.
But electricity doesn’t care how much of either you have. The queue is the queue. Loughton just showed us what happens when the line is a decade long.
Microsoft Is Still Buying AI Capacity – Just Somewhere Else
While backing out of the powerless site, Microsoft has committed to a deal with the very same company worth up to $14 billion for roughly 200,000 Nvidia GB300 chips across other locations.
The centerpiece: about 104,000 GPUs headed to a 240-megawatt campus in Texas – a site leased from a Bitcoin-miner-turned-data-center operator, designed to scale toward 1.2 gigawatts, with an option for 700 more megawatts by late 2027. Another tranche lands at the Start Campus in Sines, Portugal, a site built alongside dedicated power and subsea cables.
So, what did those sites have that Essex didn’t? Texas, with its fast-moving, generation-rich grid. A Portuguese campus engineered around energy access from the jump. Sites with existing industrial power infrastructure, inherited interconnections, or their own generation.
When Loughton stalled, demand simply relocated – flowing, like water, toward wherever the electrons already were. That is what will separate the AI infrastructure players into two groups: those holding press releases, and those holding power.
4 Data Center Power Stocks Positioned Around the Bottleneck
If powered capacity is the AI boom’s scarce commodity, then the companies that create it collect the toll – regardless of which hyperscaler or GPU cloud wins any individual lease.
These are the names we keep coming back to:
- Eaton (ETN) supplies the electrical backbone of the modern data hall – switchgear, distribution, power management – and has been working directly with Nvidia on reference designs for AI-era facilities. When a campus does secure its megawatts, Eaton gear is frequently what carries them to the racks.
- GE Vernova (GEV) sits even further upstream, selling the turbines and grid equipment that create and move power in the first place. In a world where developers increasingly bring their own generation rather than wait a decade in a queue, the makers of that generation hold rare pricing power.
- Vertiv (VRT) owns the inside of the building: power conditioning, backup systems, and the liquid cooling that dense GB300-class deployments demand. Its order book has become a proxy for real – not merely announced – buildout: the company exited 2025 with a record $15 billion backlog, more than double the year before, booking nearly three dollars of new orders for every dollar it shipped.
- Quanta Services (PWR) is the one actually stringing the wire. Transmission lines, substations, electrical construction at scale – the unglamorous work every stalled project on Earth is waiting for. Its acquisition of Cupertino Electric plugged it directly into the technology sector’s buildout.
Each of these, of course, are thematic exposures rather than automatic buys. Several have run hard as the market wakes up to the power trade, and each carries its own project-timing and valuation risks.
The more important point of this watchlist is the pattern behind it: whoever controls the narrowest point in a supply chain ends up pricing everything that flows through it.
The Bottom Line: Powered Megawatts Are Becoming AI’s Scarce Asset
In its first act, the AI trade rewarded whoever had the chips. Its second act will reward whoever can plug them in.
What died in Loughton wasn’t a data center – the site may well get built eventually. What died was an assumption, one still embedded in hundreds of glossy announcements worldwide: that the electricity will simply show up when the servers do. Microsoft’s decision shows that even the deepest-pocketed AI buyers may walk away when a credible path to power disappears.
It won’t be the last. The investors who internalize that earliest will own the choke point while everyone else is still reading press releases.
Where AI Goes When the Grid Can’t Follow
Which brings us to the question we’ve spent much of this year chasing: if the grid can’t carry where AI is headed… what can?
That question sits at the center of a project I call “XPANSE.” And unlike the stalled announcement we just walked through, this one is being built with real money, right now.
Consider what SpaceX (SPCX) disclosed in its first earnings report as a public company: $15.8 billion poured into its AI buildout in a single quarter, on a capacity schedule – roughly two gigawatts of compute this year, approaching ten by the end of 2027 – that no developer stuck in an interconnection queue could dream of matching.
While the rest of the industry waits a decade for a substation, Musk isn’t asking the grid for permission. He’s routing around it entirely.
The modern economy currently balances on two single points of failure: advanced chips concentrated on one island in the Taiwan Strait, and electricity trapped in decade-long queues. XPANSE, as I see it, is Musk’s answer to both.
So, my team and I spent months tracing where those billions actually land – the suppliers, the enablers, the publicly traded companies catching money that’s already in motion. In my new briefing, I walk through the three moves that follow from that map, starting with one company sitting squarely at the center of it all, whose name and ticker symbol I’ll give you for free, just for watching.
The grid queue may be a decade long; but the window to get positioned isn’t.
Watch the full XPANSE briefing now.
