Nvidia stakes new startup that flips script on data center power
Add Axios as your preferred source to
see more of our stories on Google.

Illustration: Brendan Lynch/Axios
AI giant Nvidia and boldface names in tech and finance are backing a new startup that aims to transform data centers into flexible grid assets instead of liabilities.
Why it matters: Emerald AI is emerging from stealth today as the AI surge risks straining grids, and hyperscalers are limited by power availability.
Driving the news: Radical Ventures led the $24.5 million seed round with participation from Nvidia, AMPLO and others.
- Individual backers include ex-climate envoy John Kerry; Kleiner Perkins chair John Doerr; Google chief scientist Jeff Dean, and pioneering AI scientist Fei-Fei Li, to name a few.
- The founder and CEO is Varun Sivaram, a physicist whose prior stops include Ørsted and serving as a senior Kerry aide.
State of play: The software shifts AI computational loads to match regional grids' needs.
- That eases stress when demand is peaking, and lessens pressure for new generation projects and grid infrastructure.
Catch up quick: A recent field test in Phoenix with Oracle, Nvidia, the Electric Power Research Institute and the regional utility Salt River Project put the software through its paces.
- It showed that Emerald AI can reduce AI workload's power consumption by 25% over three hours during a "grid stress event," while ensuring acceptable performance, Emerald AI and Nvidia said.
The big picture: "Imagine a future in which AI data centers become an important solution, if not the most important silver bullet solution to better utilize our existing electricity system, which actually makes rates go down, not up," Sivaram said in an interview.
- Hyperscalers can avoid waits for interconnection that can stretch five to 10 years, because the software makes new power availability less of a constraint, he said.
How it works: Integrating Emerald AI's software with Nvidia chips and data center controls enables instant changes in AI workloads at and between facilities.
- That can mean redirecting queries away from data centers where local power use is spiking, or briefly slowing workloads, such as the training of large academic models.
The intrigue: There's growing interest in data centers' flexibility to lower power use for limited stretches.
- Duke University's Tyler Norris, lead author of a buzzy recent paper, is among Emerald AI's advisers.
- How much of this "curtailment-enabled headroom" is available and tapped could affect how much new gas-fired power is built.
Sivaram is seeking a "paradigm shift" that lets data centers become "the ultimate virtual power plant."
- The tech can also help integrate more renewables onto grids, execs say.
- Variable renewables are "easier to add to a grid if that grid has lots of shock absorbers that can shift with changes in power supply," Ayse Coskun, Emerald AI's chief scientist, said in a statement.
- "Data centers can become some of those shock absorbers."
What's next: A larger test in Phoenix, followed by more demos elsewhere in the country over the next six months.
- Emerald AI is eyeing commercial deployment in early 2026, Sivaram said.
