AI is the fastest-growing electricity load on the planet. The usual answer is to build more power, and plenty of capital is doing exactly that. We're part of it. Energy, grid, and fuels are a real part of what we fund. But, building is slow, expensive, and disruptive for local communities and ecologies.
Get our full visual explainer for Data Center Load Management.
The cheaper answer, and the cleaner one, is to use the power we already have, just more flexibly. We've spent the last few months mapping the software that lets you do that.
The scale is the first problem. One large data center pulls about 100 MW, the same as roughly 240,000 European homes, and the average new build size keeps creeping up, from 40 MW toward 60. Data centers already eat about 1.5% of the world's electricity. The IEA expects that to double by 2030, to roughly 945 terawatt-hours a year. That's more than Japan uses. And almost all of the growth is AI, whose slice alone is on track to more than triple.
The grid can't keep up. A new connection runs 4 to 7 years in the big markets, and the gas turbines you'd use to jump the queue are sold out into 2028. So the largest operators stopped waiting. They're building their own gas plants behind the meter, xAI's Colossus outside Memphis and Oracle's site in Abilene, Texas, both already running on it. Accenture reckons gas will cover almost 60% of US data center power by 2030. The fastest-growing load on the planet, wired to fossil fuel, right as emissions need to fall.
And you can't always build your way out anyway, because the people nearby increasingly say no. In the first three months of 2026, local opposition blocked or stalled 75 US data center projects worth around $130 billion, roughly the whole of last year packed into a single quarter (Data Center Watch). New York just passed the first statewide moratorium, a one-year freeze on permits for the big hyperscale sites, and it's on the governor's desk now for signature. Power, water, who foots the bill for grid upgrades: it's all turned into a local political fight. A project that can't get permitted is worth nothing, whatever the cheque size behind it.
"A project that can't get permitted is worth nothing, whatever the cheque behind it."
Europe hits the same wall from a different angle. Dublin has frozen new connections, Frankfurt and Amsterdam are jammed, and across the continent the real bottleneck is simply getting plugged in.
Now, the queue is about to get longer: the EU's Tech Sovereignty Package (June 2026) aims to triple domestic data center capacity - into a grid that's already constrained. In Europe, flexibility is the only way to get connected at all.
The headline numbers miss something, though. A grid is built for its busiest moment. Data centers don't really have one; they run significantly flatter than almost any other load class, all day, every day. If that load can flex even a little, the grid you already have holds far more compute than its nameplate suggests.
The size of that is surprising. Duke University's Nicholas Institute ran the numbers: shave just 0.5% off a data center's yearly load, and the existing US grid could absorb close to 98 GW of new data centers without building a single new power plant. In other words, a hundred power plants nobody has to build.
In Europe the payoff is a bit different. Flexibility defers new generation, sure, but the real prize is a connection date years sooner, which is the one thing money can't otherwise buy here. The EU's Energy Efficiency Directive already makes data centers report their energy use, and a flexible site is a far easier one to wave through. The AI build-out is a roughly $5.2 trillion bill globally by 2030 (McKinsey), and the power, grid and cooling slice of it runs somewhere between $0.9 and $2 trillion. Every megawatt you defer or avoid takes a bite out of both the bill and the emissions.
There are two ways to wring more AI compute out of a megawatt. You shift the work, either in time or to wherever the power is cheap and clean. Or you make the work itself need less power. We fund both.
If it were easy, it'd be done already. Data center load is intrinsically hard to flex. It runs flat and critical, at five-nines uptime, about as far from a load you can casually switch off as a grid ever sees.
The hyperscalers already cracked their own version of this. Google, AWS and Microsoft treat their global fleets as one big pool of compute, and they quietly shuffle deferrable jobs to wherever the power is cheapest and greenest that hour. Following the sun like this can cut a workload's carbon by roughly 15% (14.6–16.3% in the research), run through fleet managers like Google's Borg that treat every site as one pool. For them it's plumbing by now.
Everyone else is locked out, and not for reasons that fix themselves. GDPR for everyone, EU's Digital Operational Resilience Act (DORA) for the financial sector, and the rest mean you can move the compute, but the data it needs often can't legally leave the region. And colocations, which hold about 22% of the world's capacity, mostly can't even see which of their tenants' jobs are movable.
Shifting the data costs a fortune on top of that. Hyperscalers own the fiber and keep warm copies of data in every region, so when they shift a workload only the instruction moves. Colocations own none of that; they pay punishing egress and storage fees just to pre-position data, so in practice it stays put. The workloads themselves are stickier too: hyperscaler jobs are checkpointed and restartable in milliseconds, while a colocations tenant's legacy workload is "stateful", meaning they freeze the memory, ship it over the network, and a single dropped packet crashes it. Re-architecturing a complex legacy workload like this to move cleanly could take up to 7 years according to McKinsey. And the contracts are backwards: when a landlord saves money by shifting load, the tenant carries the migration risk and sees none of the upside. So nobody moves.
We see that gap as the whole opportunity. The hard part is real, the players who could build it in-house mostly haven't bothered, and the ones who actually need it (colocations, neoclouds, AI labs spread across sites) don't have it. That's where we want to be early.
We went through the European and global stack at Seed and Series A and sorted what we'd back into four buckets.
First, autonomous orchestration. Modern data centers are getting too dense and twitchy to run by hand, so we want control software that treats power, cooling and compute as one closed loop, and has actually proven it in a live building. Phaidra (Washington, USA) is an example of a later-stage solution with AI agents running all three; they're at Series B now. Lucend (Amsterdam, Netherlands) and Etalytics (Hesse, Germany) are earlier, coming at it from prescriptive control and energy management.
Second, grid-interactive data centers. The software opens up two-way power flow so a building can behave like a virtual power plant, throttling or rescheduling jobs when the grid asks and getting paid for it. The data center turns into a grid asset that happens to do compute on the side. Emerald AI (Washington DC, United States) and soma.energy (Vancouver, Canada) are out front; Loadwise (Berlin, Germany) is building the grid-signal layer they all lean on.
Third, workload mobility, which is really the colocations and multi-site problem. We want a confidential-computing layer that lets a tenant mark a job as movable without showing anyone what it is, plus schedulers that send work to whichever site has cheap, green power right now. Electricity Maps (Copenhagen, Denmark) already provides the carbon and price signal those schedulers would run on, and Pado (California, United States) is working the orchestration side. The confidential-computing piece, the bit that would finally pull colocations in, doesn't exist yet.
Fourth, power-aware infrastructure. Least glamorous, but probably the most clear-cut. It's the efficiency layer between the chip and the building that nobody really owns: tuning how hard the GPUs draw, scheduling around power, clawing back capacity that's sitting stranded. A rack that's already plugged in can give up another 10 to 20% with the right software, which buys the same headroom as shifting load and needs no new permits. Neuralwatt (Washington, United States) and Hammerhead AI (California, United States) work the GPU side; Corintis (Lausanne, Swizterland) cools the chip directly with microfluidics, and Nexalus chases the same gains at the rack.
One of these four is wide open. Colocations flexibility is just unsolved. Operators can't see their tenants' workloads, and even where they can, the contracts won't let them act on it. Our hunch is that the answer is confidential computing plus a genuinely new kind of tenant agreement that allows for, and we haven't found anyone building that yet. If that's you, we'd like the first call.
"We haven't found anyone building it. If that's you, we'd like the first call."
Here's the watchlist as it stands, Seed to Series A, Q2 2026:
- Autonomous orchestration: Lucend, Etalytics, Calcore, BayCompute*, Fluix AI*, Phaidra* (the vertical leader, now at Series B)
- Grid-interactive: Loadwise, soma.energy*, Emerald AI*
- Workload mobility: Electricity Maps, Pado*
- Power-aware infrastructure: Neuralwatt*, Haloid, Hammerhead AI*, Niv-AI*, Nexalus, Corintis
( headquartered outside Europe.)*
Who are we missing? Reach out and let us know.
Sources
Key facts and figures, in order of appearance. Render as inline links or an endnote list in the CMS.
Data centers ~1.5% of global electricity, doubling to ~945 TWh by 2030 (more than Japan uses today); AI demand tripling — IEA, Energy and AI (2025). https://www.iea.org/reports/energy-and-aiOne large data center ≈ 100 MW ≈ 80,000 homes; average new build 40 → 60 MW — dgtlinfra, Data Center Power (2025). https://dgtlinfra.com/data-center-power/Gas to supply ~60% of US data center power by 2030 — Accenture / Soben, Data Centre Trends 2026. https://www.accenture.com/content/dam/accenture/final/accenture-com/document-4/Data-Centre-Trends-2026-Soben-Part-of-Accenture.pdfUS local opposition blocked or delayed 75 data center projects worth ~$130B in Q1 2026 — Data Center Watch (10a Labs), 2026 — https://www.datacenterwatch.org/report ; coverage: NBC NewsNew York's first statewide data center moratorium (S.10462), awaiting the governor — Inside Climate News ; Greenberg Traurig (June 2026)Existing US grid could host ~100 GW of new data centers if load curtails ~0.5% of the year — Nicholas Institute, Duke University, Rethinking Load Growth (Feb 2025). https://climate.duke.edu/annual-report/items/rethinking-load-growth/AI data center capex ~$5.2T by 2030; power/grid/cooling ("energizers") slice ~$0.9–2T — McKinsey, The cost of compute: a $7 trillion race to scale data centers (2025). https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centersColocation ~22% of global data center capacity — Synergy Research Group, 2026. https://www.srgresearch.com/articles/the-worlds-total-data-center-capacity-is-shifting-rapidly-to-hyperscale-operatorsInference overtaking training as the bulk of AI compute — Deloitte, 2026 TMT Predictions. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.htmlEU Energy Efficiency Directive: data center energy reporting — European Commission. https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficiency-targets-directive-and-rules/energy-efficiency-directive_enCarbon-aware "follow the sun" workload shifting cuts emissions (double digits) — representative: A Guide to Reducing Carbon Emissions through Data Center Geographical Load Shifting, arXiv:2105.09120.McKinsey Digital / QuantumBlack, "AI for IT modernization: Faster, cheaper, better", December 2024.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/ai-for-it-modernization-faster-cheaper-and-better
