Mourad Elsherei
AI's Power Problem
A sector map, and the three companies I'd start looking for on Monday
London · Prepared for Concept Ventures · July 2026
Five years learning how companies work; the next decade choosing which ones to back. I've built software inside one of the world's largest banks, run product in a private markets firm as it deployed $30m, and now advise businesses adopting new technology through my own consultancy.
Looking back, each role moved me one step earlier in a company's life: from scaling inside an incumbent, to working alongside private capital, to sitting with founders and operators as they decide what to build and buy.
The question that kept getting more interesting wasn't how to build one company well. It was how some founders see the opportunity years before everyone else does. That's an investing question, so when Concept's job ad said interns build sector maps around emerging technologies, I built the first one before applying.
01 · The crunch
AI has made electricity the binding constraint on compute. Data centre power demand is set to more than double by 2030, roughly Japan's entire consumption1, while grid connections in Europe's main markets queue for up to seven to ten years and Dublin has paused new connections until 20282.
When demand is desperate and supply is stuck, value migrates to whatever sits between them.
Billions are flowing into the physical layer. The software layer that finds, connects and squeezes every watt is still cheap, and it's being founded right now by people leaving energy scaleups, grid operators and hyperscalers. That's pre-seed's home turf.
02 · The map
Make more power · new and stranded electrons for compute
Connect faster · siting, queues and grid access
And Europe's first fully islanded microgrid data centre just went live in Dublin4.
Run every watt better · inside the facility
Make load grid interactive · facility ↔ market
One live check on this read: a €235m European energy fund published its own mapping of 100+ companies this month and reached the same conclusion, that software is the most capital efficient answer to the AI power crisis and the most underfunded layer in the stack2.
03 · The white space
№1The Gate 2 company.
Software that gets projects through Britain's rebuilt grid queue. NESO tore up first come first served in December: connection now depends on evidence (land rights, planning, delivery milestones), with termination for slippage and application windows twice a year; the next opens late 2026.
Contracted demand jumped from 41GW to 125GW in seven months3, with data centres driving it. Thousands of projects now live or die on paperwork, and that's a workflow product a $1m first cheque can own.
Founders to watch for: connections team leavers from NESO and National Grid, and energy consultants going product.
№2Grid interactive AI, Europe edition.
Software that shapes training and inference to grid signals so operators win faster, cheaper, flexible connections. The US version exists: Emerald AI, with Nvidia behind it.
Europe's barely does, yet Ireland now requires new data centres to be dispatchable4 and the EU is targeting 230GW of flexibility by 20305. Regulation is doing the selling.
Founders to watch for: optimisation people leaving DeepMind, hyperscaler energy team leavers, engineers out of Octopus Kraken.
№3Watt ops for the mid market.
Hyperscalers solve power in house; Zendo is winning the colocation layer; below both sits a long tail of 1 to 20MW operators (enterprise server rooms, GPU clusters, edge sites) still run on spreadsheets and best guesses.
I advise businesses like these weekly: the willingness to pay is real and unglamorous, which is exactly why it's still open.
Founders to watch for: colo operations leads pairing with B2B software PMs.
04 · Reading Zendo backwards
The interesting thing about Zendo isn't that it raised. It's what was visible before it did.
Months ahead of the round, the inputs were already unusual and entirely public: an operator thirteen years deep at Octopus Energy going quiet, an enterprise product builder from Square joining him, a fresh Companies House filing, and a market whose constraint had just flipped from software to infrastructure. The £1.75m round, led by Fly Ventures, who wrote the first cheque into Wayve, with Octopus Ventures and Pact6, didn't create the signal; it confirmed a pattern that was sitting in plain sight.
Working that pattern forwards instead of backwards, across hundreds of companies a year, is exactly the muscle I'm applying to build. The next Zendo is incorporating in London this quarter.
05 · What I'd track from Monday
And what would change my mind: queues clearing faster than forecast, or hyperscalers pulling this software in house. Both are checkable every quarter.
06 · Why an internship, five years in
Building products taught me execution; working around investors taught me capital allocation; running my own consultancy is teaching me how businesses actually buy software.
The skill I'm missing isn't company building. It's pattern recognition, and it isn't something you learn alone. It comes from repetition and feedback: seeing hundreds of founders and markets before consensus exists, calibrated by people who've done it for years.
Pre-seed compresses that learning loop harder than any other seat, which is why I'm applying for this one. I'm not passing through; I'm repointing.