Power smoothing for AI racks

AI swings. The grid shouldn't.

G12 sits inside the rack and turns the spiky power draw of AI training into a steady load the grid can approve.

Talk to engineering Join the pilot
Power smoothing illustration A spiky orange line, representing GPU power draw, enters a box labelled G12 and leaves as a smooth blue line, representing what the grid sees. G12
The problem

Big AI loads swing faster than the grid can follow.

  1. GPUs move in lockstep

    Thousands of GPUs compute together, then pause together to swap data, all at the same instant.

  2. Megawatts swing in milliseconds

    A whole site's power draw jumps up and down by megawatts, many times a second.

  3. The grid pushes back

    Utilities see oscillations that stress turbines and equipment, so they slow down or limit new connections.

New AI sites wait for grid connection, and some are refused.

The solution

A shock absorber in every AI rack.

What the GPUs draw

Sharp, synchronized swings

G12

  • Power chip + firmware that react in the rack
  • Energy-storage module in the power shelf

What the grid sees

A steady, predictable load

G12 charges in the dips and discharges at the peaks.

What customers get

Steady power, without slowing the GPUs.

Faster grid approval

A smooth load profile is easier for utilities to study and approve.

No wasted energy

No dummy workloads burning power to fill the gaps. Energy is stored, then reused.

Lower peak power bills

Shaving peaks at the rack cuts demand charges on every bill.

Drops into the rack

Built for the power shelf makers already ship. No new building or site batteries.

Why now

Racks are getting bigger. The grid is the bottleneck.

600 kW

Where AI racks are heading, so each rack swings harder than ever.

Grid

Grid connection, not chips or buildings, is now the slowest step in bringing AI capacity online.

Up to 30%

Lower peak demand from in-rack storage on NVIDIA GB300, proof that the approach works.

How we work with makers

We're your supplier, not your rival.

G12 is a chip and a reference module that power-shelf and rack makers design into their own products. Their customers specify it. Makers sell it.

How design-in works
G12Power chip, firmware, reference module
Power-shelf and rack makersDesign G12 into their shelves
Design-in
Hyperscalers, AI labs, neocloudsAsk for G12 in the racks they buy
Spec-in
Comparison

The only option that checks every box.

How G12 compares with other ways to smooth AI power
Requirement Site batteries Software power burn GPU-vendor storage G12
Smooths in the rack — ✓ ✓ ✓
No wasted energy ✓ — ✓ ✓
Works with any GPU ✓ Partly — ✓
A chip any maker can design in — — — ✓
Roadmap

From validation to first silicon in a year.

  1. Validate

    Model real GPU power traces and confirm the design with makers.

  2. Prototype

    Working module and evaluation board.

  3. Live-rack pilot

    G12 smoothing a real AI rack at a pilot site.

  4. First silicon

    The G12 power chip, ready for design-in.

Team

Built by power engineers.

Gervin Kahunde

Founder & CEO

Emiru Egwell

Power IC lead

Jimmy Mangali

Control firmware

Mohan Raj Gopal

Data-center power

Meet the team
Pilot program

Waiting on grid approval? Join the pilot.

We're lining up sites for our Q2 2027 live-rack pilot. Tell us about yours and we'll be in touch about fit and timing.

Let's get your AI capacity on the grid.

Makers: design G12 into your next power shelf. Developers: join the pilot.

Talk to engineering Join the pilot