Everything-to-Grid: How AI Data Centers, EVs, and Batteries Are Becoming the Grid

Nepal exports its surplus hydropower to India across a single interconnection — a fact that's easy to take for granted here, but it's a reminder that a grid is only as intelligent as its ability to move power in the direction the moment actually demands. The rich world is now rediscovering, at continental scale and with considerably more hardware, something a small hydro-nation has always understood by necessity: the future grid has to be bidirectional, because the alternative is building enough spare capacity to never need to be.

The data-centre demand curve that changes the math

Start with the side of this story that's driving urgency: AI infrastructure's own appetite for power. The International Energy Agency's April 2025 "Energy and AI" report is the primary source worth anchoring to here, and its numbers are specific. Data centres consumed roughly 415 terawatt-hours in 2024 — about 1.5% of global electricity demand — and the IEA projects that climbing to just under 945 TWh by 2030, nearly tripling in six years. The regional concentration is stark too: as of 2024, the United States accounts for about 45% of global data-centre electricity demand, China about 25%, and Europe about 15%. In the US specifically, the IEA expects data-centre demand to grow roughly 130% by 2030.

That's the demand side. What makes 2026 different from a straightforward "the grid needs more power plants" story is that both the demand and the supply are increasingly the same kind of asset: distributed, software-addressable, and capable of shifting in real time. This piece is deliberately narrower than my earlier look at the coal-to-renewables crossover reshaping global electricity generation, which covers where the power comes from; this one is about how millions of small nodes — EVs, batteries, and the data centers I toured earlier this year — are becoming part of how it moves.

The virtual power plant, scaled to a nation

The Department of Energy has been tracking this shift formally since its original "Pathways to Commercial Liftoff: Virtual Power Plants" report in September 2023, updated in 2025. A virtual power plant, in DOE's framing, aggregates distributed energy resources — rooftop solar, home batteries, EV chargers, smart thermostats — into a single dispatchable resource a utility can call on the way it calls on a conventional power plant. The department's target is concrete: 80-160 gigawatts of VPP capacity by 2030, representing roughly 10-20% of expected peak demand, worth an estimated $10 billion a year in avoided grid infrastructure costs. Current North American VPP capacity sits at roughly 33 gigawatts as of the 2025 update — meaning the DOE's own target requires roughly doubling to quintupling that number in five years, against a backdrop where US peak demand itself is expected to climb from about 800 gigawatts in 2024 toward 900 gigawatts by 2030.

The department's projections for what feeds that growth are specific enough to repeat directly: 20-90 gigawatts of new EV-charging-related demand response capacity and 300-540 gigawatt-hours of EV battery storage capacity added between 2025 and 2030. Those are wide ranges, which is itself worth noting — DOE is explicit that the pace depends heavily on EV adoption rates and how many of those vehicles end up enrolled in bidirectional or managed-charging programs, not a settled trajectory.

When a data center becomes a grid asset

The most concrete illustration of AI infrastructure and grid infrastructure becoming the same conversation is the Microsoft-Constellation deal announced in September 2024: a 20-year power purchase agreement to restart Constellation's Crane Clean Energy Center — the undamaged reactor at Three Mile Island — at a cost of roughly $1.6 billion, delivering 835 megawatts, with a restart now targeted for 2027. It's a hyperscaler treating a specific power plant's capacity as strategic infrastructure worth a two-decade contractual commitment, which is a meaningfully different posture than simply buying power off the grid at spot rates.

A smaller-scale but structurally similar shift is happening at the meter level. Utilidata partnered with NVIDIA to build "Karman," a grid-intelligence module built on NVIDIA's Jetson Orin Nano chip, designed to give grid edge devices real-time AI processing capability; Aclara became the first manufacturer to embed it, announced March 12, 2024. Utilidata raised $60 million in a round that included NVIDIA's participation. The pattern across both examples — one at gigawatt scale, one at the level of a single grid sensor — is the same: AI compute is showing up as infrastructure inside the grid itself, not just as a customer sitting on top of it.

What I'm not pinning down yet

Vehicle-to-grid technology is real and shipping — Ford's F-150 Lightning supports vehicle-to-home power, Nissan's Leaf has offered V2G capability, GM's Ultium platform and Hyundai's Ioniq 5 support various forms of bidirectional power flow, and Tesla has discussed bidirectional charging plans of its own. I'm deliberately not attaching specific kilowatt ratings or model-year availability claims to any of them here, because those specs shift with trim levels and firmware updates faster than I could independently re-verify for this piece, and a wrong number on a well-known consumer product is a worse look than no number at all. The same discipline applies to the standards underpinning V2G — ISO 15118 for vehicle-to-grid communication, IEEE 1547 for distributed resource interconnection, OpenADR and OCPP for demand response and charging-station communication all exist and matter, but I'm describing their roles generically rather than asserting version-specific capabilities I haven't checked line by line. And specific megawatt figures attributed to Tesla's California virtual power plant program, Google's 24/7 carbon-free-energy initiative, and various Lawrence Berkeley National Lab aggregate V2G studies all appear in circulation with numbers that don't consistently match across sources — so they're left as "these programs exist and are meaningful" rather than pinned to a figure I can't stand behind.

The view from Kathmandu

The everything-to-grid shift reads, from here, less like a technology story and more like an accounting one: for decades, the grid treated demand as something to forecast and supply as something to build ahead of it. What DOE's VPP target and the IEA's data-centre numbers describe together is a grid where a meaningful share of both demand and supply are the same software-addressable asset, adjustable in real time rather than fixed months in advance. Nepal's single hydro interconnection to India is a crude version of exactly that idea — moving power toward wherever it's needed most, at the moment it's needed. The rich world is now trying to build a version of that instinct with millions of small nodes instead of one river, and the DOE's own numbers say it's only about a fifth to a third of the way there.

Sources

• energy.gov — "Pathways to Commercial Liftoff: Virtual Power Plants," original Sept 2023 report and 2025 update (80-160GW target by 2030, ~$10B/yr savings, current ~33GW capacity)
• Utility Dive — coverage of DOE VPP Liftoff update, including DER addition ranges (2025-2030)
• iea.org — "Energy and AI," April 10, 2025 (data-centre electricity demand: 415 TWh 2024 to ~945 TWh 2030; regional shares; US +130% by 2030)
• renewableenergyworld.com — DOE-sourced coverage of US peak demand trajectory (~800GW 2024 to ~900GW 2030) and current VPP scale
• CNBC; NPR; Data Center Dynamics — Microsoft-Constellation Three Mile Island (Crane Clean Energy Center) PPA, Sept 2024 ($1.6B, 835MW, 20-year term, 2027 restart target)
• utilidata.com; globenewswire.com — Utilidata + NVIDIA "Karman" module, Aclara integration (Mar 12, 2024), $60M raise

Written by Abhishek Kushwaha, founder and writer at Global Tech Search, based in Kathmandu, Nepal.