CEOInterviews.AI
Start App
Varun Sivaram
Varun Sivaram
Founder and Chief Executive Officer, Emerald AI
Watch

AI Factories: How Emerald AI Stabilizes the Grid

📅 May 14, 2026 SAIL Media and Makora 10 MIN 372 VIEWS 18 SEGMENTS · 2 SPEAKERS
Varun Sivaram discusses the critical power constraints facing the AI industry, the shift toward "power-flexible" data centers, and how Emerald AI is enabling massive compute clusters to operate in harmony with the electric grid. Varun explains how their "Conductor" platform helps "AI Factories" dynamically adjust their power consumption to prevent grid overloads—potentially doubling the capacity for AI infrastructure in the US without waiting years for expensive grid upgrades. Varun Sivaram is the CEO and Founder of Emerald AI, a startup dedicated to making large-scale AI infrastructure a fle...

What Varun Sivaram said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Varun Sivaram, CEO of Emerald AI, explained that his company makes AI factories flexible power users, enabling them to adjust power consumption on demand to stabilize the grid. He noted that over the next 3 years, America has demand to build 50 gigawatts of AI factories but can only plug in 25, while China will have 400 gigawatts of spare power capacity by 2030. Sivaram said the grid is utilized at roughly 50%, with transmission lines at 30%, and that AI factories can reduce power by 25% for 3 hours during peak loads. He cited demonstrations with Oracle, Nebius, and Nvidia, including temporal flexibility in London and spatial flexibility moving inference queries from Virginia to Chicago. Sivaram announced a partnership with Siemens and a 100 MW commercial deployment with Nvidia in Manassas, Virginia, integrating DSX Flex. He argued this technology makes AI factories attractive to utilities and communities.

Key takeaways

  1. Emerald AI enables AI factories to reduce power by 25% for up to 3 hours during grid peak events, fitting on today's grid without massive buildout.
  2. America can fit 100 gigawatts of AI factories on today's grid with this technology, versus 50 gigawatts demand and 25 gigawatts plug-in capacity over 3 years.
  3. China will have 400 gigawatts of spare power capacity by 2030, so it is limited by chips while America is limited by power.
  4. Emerald AI announced a 100 MW commercial-scale deployment with Nvidia in Manassas, Virginia, integrating DSX Flex, the world's first power-flexible AI factory.
  5. Sivaram said the grid is utilized at roughly 50%, with transmission lines at 30%, and that AI factories can be flexible 99% of the year.

Numbers and commitments

FigureWhat it refers toTypeAt
50 gigawatts Demand to build AI factories in America over next 3 years metric 1:50
25 gigawatts AI factory capacity that can be plugged in over next 3 years metric 1:50
400 gigawatts China's spare power capacity by 2030 metric 1:50
50% Grid utilization rate metric 2:56
30% Transmission line utilization rate metric 2:56
25% Power reduction during peak load events metric 2:56
3 hours Duration of power reduction during peak events timeline 2:56
100 gigawatts AI factory capacity that can fit on today's grid with this technology metric 4:13
100 MW Commercial-scale deployment with Nvidia in Manassas, Virginia commitment 7:51
75 MW Reduced power level for the 100 MW AI factory when asked metric 7:51

Chapters

  1. 0:00Emerald AI's mission and value proposition
  2. 1:50Grid constraints and the power deficit
  3. 2:56How flexibility works on the grid
  4. 5:12Demonstrations with Oracle, Nebius, and Nvidia
  5. 5:47Temporal and spatial flexibility examples
  6. 7:05On-site energy resources and Siemens partnership
  7. 7:512026 roadmap and Nvidia DSX Flex integration
  8. 9:22Team and autonomous AI agents

Questions asked in this interview

5
  1. 0:09Could you introduce yourself, like your name and where you work and the best place to find you online?
  2. 0:29Like what does Emerald AI do?
  3. 1:15Um but is it that like when the grid there's more demand, you you want to like decrease the amount of power, the amount of GPUs that are running so that like you're not stressing the grid when it's at higher demands?
  4. 4:24So, I'm like curious about like are you specializing towards certain types of data centers that are more likely to be kind of more agnostic about what time the load runs?
  5. 7:44What do you think kind of your 2026 roadmap, um what what's on top for you all?
Interviewer 0:09 ↗
Well, welcome to this researcher chat. I hope you're doing well today. Could you introduce yourself, like your name and where you work and the best place to find you online?
Varun Sivaram 0:18 ↗
Sure. Yeah. I'm Varun Sivaram and the CEO and founder of Emerald AI. Find us online at emerald.ai. But excited to join you.
Interviewer 0:29 ↗
Yeah. Could you start just like the pitch? Like what does Emerald AI do?
Varun Sivaram 0:34 ↗
Emerald AI makes AI factories flexible power users. The reason this is so important is we kill three birds with one stone. We enable more AI factories to connect with larger and faster power connections. We keep costs low for energy bills for communities locally. And we keep the electric grid reliable. We do all this by enabling the AI factory to change its power consumption on demand depending on what the power grid needs because historically AI factories or data centers have never responded to power grid needs, but that's possible now and Emerald AI is powering that technology.
Interviewer 1:15 ↗
Okay, so let me make sure that I understand. So you have some like data center that's running a bunch of GPUs that and I think AI factory what you mean by that is something that is just really good at producing tokens of AI models. And like traditionally you're just like running max utilization all the time. Um but is it that like when the grid there's more demand, you you want to like decrease the amount of power, the amount of GPUs that are running so that like you're not stressing the grid when it's at higher demands?
Varun Sivaram 1:50 ↗
Sort of. Let's back up a second. At a higher level, what's the problem we're trying to solve? Historically, if you wanted a data center, you would plug it into the grid, and then you would never think again about the grid. You would just run your CPUs and maybe GPUs whenever you wanted to, right? It's kind of like when you walk into a room, you turn on the light whenever you want. You don't think about the grid. Nowadays, energy has become the limiting constraint on data centers, and I say AI factories because as you said, the principal source of load growth is the token factories, right? These GPUs, specialized accelerator data centers. We call them AI factories. Actually, I don't. Jensen Huang calls them AI factories. And so now we have to care about the grid. Well, over the next 3 years in America, there's demand to build 50 gigawatts. That's a unit of energy. 50 gigawatts worth of AI factories, and we can only plug in 25 over the next 3 years, right? So there's a clear deficit. And the reason that's important is China doesn't have this deficit. China by 2030 will have 400 gigawatts of spare power capacity, so they'll be able to build plenty of AI factories. They're limited by chips, we're limited by power. Emerald helps to circumvent this in the following way.
The grid is kind of like a superhighway that's empty almost all the time, right? The grid is utilized at roughly 50%. In fact, the transmission lines are utilized at 30%. Most of the time they're carrying way less power than they otherwise could, and very rarely do they hit their peak load. But they're designed to meet that peak load, and so if an AI factory rocks up and they say, 'Hey, I'd like to pull 100 megawatts from the grid.' They say, 'Well, you might pull it at the absolute worst moment of the year. Hot summer day in Phoenix, all the air conditioners are running, and you ask for 100 megawatts, right?' And so today, AI factories may not be fully utilized. They may go up, they may go down. The problem is that they're unpredictable. With Emerald AI, we make them controllable and flexible in the moments that matter most for the grid. And so 99% of the year, the AI factory can use as much power as it wants. But 1% of the year, when you hit that peak load on the Phoenix grid, or as we showed in London, when it's half time of a soccer game and all of the tea kettles go on and spikes on the grid, that AI factory is able to be well-behaved and reduce its power temporarily. Let's say 3 hours, it'll reduce its power by 25%. That behavior means that that AI factory can fit on today's grid without triggering a massive and expensive buildout.
Yeah. Today in America, we can fit 100 gigawatts worth of AI factories on today's grid. I told you we need to build 50 gigawatts in the next 3 years. We can fit 100 today if we have this technology.

6 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Sivaram, V. (2026, May 14). AI Factories: How Emerald AI Stabilizes the Grid [Interview transcript]. SAIL Media and Makora. CEOInterviews.AI. https://ceointerviews.ai/interview/1435977/

MLA

Varun Sivaram. "AI Factories: How Emerald AI Stabilizes the Grid." SAIL Media and Makora, 14 May. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/1435977/.

BibTeX 
@misc{sivaram2026_1435977,
  author       = {Varun Sivaram},
  title        = {AI Factories: How Emerald AI Stabilizes the Grid},
  howpublished = {Interview transcript, SAIL Media and Makora. CEOInterviews.AI},
  year         = {2026},
  month        = {may},
  url          = {https://ceointerviews.ai/interview/1435977/},
  note         = {Speaker-attributed transcript with timestamps}
}