Journey The statement Sim time 00:00 Load Night · idle Freq 50.00 Hz Vault SOC 72 % ▮ Investor mode Next →
The firming layer · behind the meter

On-site power, made GPU-grade.

Turbines can reach an AI campus in months. What they can't do on their own is follow GPU load swings. Battery Vault builds the layer that closes that gap: a two-tier battery bank and PlantOS, our plant controller, designed into power islands our partners deliver.

GPU-grade, defined The envelope the reference design is sized against: bus frequency within ±0.5 Hz of nominal and rate of change of frequency below 0.25 Hz/s, held through the load steps a training cluster produces. The delivered envelope is set per site against the interconnect and the load.

See the island → Talk to us
Single-line diagram · Island-01 Island online
Supervisory control Generation on-site · mobile The Vault two-tier buffer · banks & releases AI data centre GPU load Generation supplies · steady GPUs demand · spiky Status ▮ Island online Freq 50.00 Hz Vault SOC 72 % Grid wait 0 months
Generation runs steady. GPUs spike. The Vault banks the difference. That difference is the layer we build.
The constraint

AI compute is ready years before the grid is.

Our customers have land, capital and GPUs on order. What they don't have is a grid connection before the 2030s.

~20%
Of planned data-centre projects are at risk of delay from grid constraints.
Source · IEA, Key Questions on Energy and AI, April 2026
485 → 950 TWh
Data-centre electricity consumption, 2025 to 2030. About 3% of global demand, with AI-focused consumption tripling over the same period.
Source · IEA, Key Questions on Energy and AI, April 2026
Grid interconnectionDepartures
Request filed 2025 —
Grid connection EST. 2031+ DELAYED
Island-01 MONTHS AVAILABLE
Illustrative · typical EU high-voltage interconnection timeline

Power is the binding constraint on AI build-out.

The solution

A behind-the-meter power island, delivered.

Partners deliver the megawatts. Our layer makes them GPU-grade: a two-tier battery bank and one supervisory controller, designed in from day one.

01 · Generation — partner-delivered

The megawatts

Turbine capacity from established energy partners, deployable in months rather than on a transmission timeline. We don't build generation — our layer sits in front of it.

02 · The Vault — ours

Two-tier battery

Energy is banked and released as GPU load swings, absorbing the transients that generation can't follow on its own. The load sees steady power.

03 · Control — ours

One plant controller

PlantOS, our supervisory software, runs generation and battery as one plant. Dispatch, buffering and islanding are handled in one place instead of being split across vendors.

Matthias Breidenbach Founder & CEO · Munich
“Another BESS startup?”

No. We don't trade storage in energy markets and we don't sell power. We supply the battery architecture and the control software that let a power island hold GPU-grade power quality.

What we are

We don't sell electrons. We make them firm.

Battery Vault is neither a battery manufacturer nor a utility. We design the firming layer, the vault and PlantOS, and deliver it inside power-island projects led by established energy partners.

Island-01 · Site view Drag to orbit · click our assets
Site view · static

The single-line diagram above carries the same island: partner generation, the vault on the bus, the AI load.

✕ Battery manufacturer ✕ Power utility · electron sales ✓ The firming layer · vault + controller
Size the buffer tier

What the machines cannot follow, the buffer has to.

A turbine can only take a load step so fast before frequency leaves the corridor. Everything past that point is the buffer tier's job. This sizes it.

100 MW 20–200 MW
17.5% of load 5–30% of load
Grid frequency
Assumptions, fixed
  • Aggregate inertia constant H4.0 s — a property of the machines the operator brings.
  • Frequency corridor±0.5 Hz of nominal.
  • Governor delay2 s.
  • Container rating2.5 MW per container.
  • BaseCampus load treated as MVA base, pf ≈ 1. An assumption, not a definition.
Derivation
  1. RoCoFmax = corridor / governor delay0.25 Hz/s
  2. ΔPpermissible = 2 · H · S · RoCoFmax / fnominal4.0 MW
  3. Buffer required = swing − ΔPpermissible13.5 MW
  4. Containers = ceil(buffer / 2.5 MW)6
Permissible machine step 4.0 MW What the generation can absorb inside the corridor on its own.
Buffer required 13.5 MW The step the machines cannot take. This is the fast tier.
Containers 6 + 1 6 for the swing, plus 1 for N-1 of the storage layer.
€

First-pass sizing basis. Not a quote and not a design.

Why now

AI power is going behind the meter, at gigawatt scale.

Grid connections take years AI operators don't have. So generation is moving on-site: turbines, engines, microgrids. A new asset class is being built at speed.

The operators with capital are no longer waiting for the utility. They're signing for on-site power today — and every one of those projects needs a firming layer to make it GPU-grade. That layer is what we sell.

~90 GW
Behind-the-meter generation announced across US data-centre projects. About 2 GW of it is operating.
Source · Cleanview, Behind-the-Meter Data Centers, 2026, a research firm's satellite-based analysis. RBC Capital Markets puts announced behind-the-meter gas capacity near 101 GW on DOE-compiled data.
Where we win

The trade-offs, side by side.

On-site megawatts arrive in months. Whether they're GPU-grade is decided by the firming layer.

Grid-forming BESS
Tesla · Fluence · Sungrow

Wins: cost per MWh, and installed references we cannot match. Thousands of megawatt-hours in the field.

Stops at: energy-market duty and per-project controls. Continuous sub-second swing duty against a turbine is a different service.

Supercapacitor E-STATCOMs
Siemens Energy SVC PLUS FS · Hitachi Energy SVC Light Enhanced · GE Vernova FACTSFLEX GFMe

Wins: response speed, and the validation that comes with three grid majors building for AI load. The problem is real at the highest level in the industry.

Stops at: seconds of energy. Every one of them.

Synchronous condensers
Rotating machines on the bus

Wins: real inertia, physically, with no control loop to argue about.

Stops at: no dispatchable energy, and two to three times the iron for the same job.

Bundled plant delivery
An operator selling generation, storage and control as one wrap

Wins: one contract, one throat to choke. It is live in Europe: Pure Data Centres Group and AVK launched a 110 MW data-centre microgrid in Dublin on 11 March 2026, three energy centres of up to 30 MW on Wärtsilä engines with a 20 MW battery, specified project by project.

Stops at: the plant it sells. That model moves a plant. We sell the layer into any fleet, including fleets a plant vendor will never own.

Everyone is fast, nobody brings the energy.

Why we can deliver Our layer ships inside projects delivered by established energy partners running field-proven turbine plant. Partner detail available under NDA.

Talk to us

Building an AI power island? Let's talk about the firming layer.▍

If you're building or powering an AI campus, or investing in the people who are, we'd like to hear from you.

For operators & power providers

Specify the firming layer

You're building or powering an AI campus. Tell us the load. We'll size the vault and the control layer.

Start the conversation →
For investors

Pre-seed, open now

We're raising to deliver the firming layer into the first islands. Open the vault for the headline terms.

Enter investor mode →
▮ Investor modeExit
The ask
€2M pre-seed
Open now. It funds the controller through the bench and into a plant, and reaches a funded seed.
Use of funds
Against dated milestones
Controller-in-the-loop bench, purchase order November 2026 · turbine governor and AVR model with the OEM packager from September 2026 · the non-stiff grid fixture as a parallel regression lane · the first controls hire.
Where we stand
PlantOS v0.2
Hardened and validated in software-in-the-loop. 9 scenarios, 26 assertions, each with a sensitivity proof. 325 tests green on main. Eight validated plant scenarios under fixed KPI gates, each with a counterfactual. A 1,188-point sweep.

The apparatus behind those numbers, the result it produced and the three boundaries on that result are public. Read the evidence.

Request the data room → Partner identities, the cap table and terms are in the data room, not on this page.

Firming layer · PlantOS v0.2, validated in software-in-the-loop — Munich · 2026