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

AI’s next bottleneck is energy. Varun Sivaram has a plan

📅 Sep 22, 2026 Axios Live and Axios 14 MIN 40 SEGMENTS · 2 SPEAKERS
Axios’ Amy Harder sits down with Emerald AI co-founder and CEO Varun Sivaram to discuss one of the biggest constraints facing the AI boom: energy. Sivaram argues that demand for AI compute is accelerating while the U.S. power grid struggles to connect new data centers fast enough. He explains why flexible AI data centers could reduce power use when the grid is under stress, helping utilities add more computing capacity without waiting years for major infrastructure upgrades. The conversation covers growing community opposition to data centers, Emerald AI’s new AI Energy Management Alliance,...

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, said the AI boom is not a bubble and that energy remains the bottleneck, with the US likely connecting only 30 of the 70 gigawatts of AI data center capacity planned by 2028. He highlighted a new political backlash against data centers and introduced the AI Energy Management Alliance, co-founded with Google and Nvidia, bringing together 20 companies including Anthropic, AES, Constellation, NRG, and National Grid to promote flexible data centers that can reduce power during grid stress. Sivaram proposed a 'grand bargain' where regulators offer faster grid access in exchange for data centers protecting rates and reliability. He declined to call Amazon's behind-the-meter natural gas plant a mistake, praising the company and inviting it to join the alliance. He said Emerald, now a unicorn, can unlock 100 gigawatts of existing grid capacity, worth $4 billion, and that a 100-MW power-flexible AI data center with Nvidia will open in Q4 2026. He dismissed acquisition talk, saying he is in it for the long run, and separated AI safety concerns from community opposition to data centers.

Key takeaways

  1. US will connect only 30 of 70 gigawatts of AI data center capacity planned by 2028, making energy the bottleneck.
  2. Emerald AI's alliance with Google and Nvidia aims to make data centers flexible, enabling 100 gigawatts of existing grid capacity to be used.
  3. A 100-MW power-flexible AI data center with Nvidia will open in Q4 2026, proving data centers can be good grid citizens.
  4. He invited Amazon to join the AI Energy Management Alliance, praising its leadership in sustainability.

Numbers and commitments

FigureWhat it refers toTypeAt
70 gigawatts AI data center capacity US wants to connect by 2028 metric 0:36
30 gigawatts AI data center capacity US will likely connect by 2028 metric 0:36
20 companies Companies in the AI Energy Management Alliance metric 2:09
100 gigawatts Available power that can fit data centers on existing grid if flexible metric 6:34
$10 million Profit per megawatt of grid capacity enabled by Emerald metric 6:34
12 Fortune 500 global investors in Emerald AI metric 6:34
$4 billion Value Emerald seeks to unlock metric 6:34
100-MW Power-flexible AI data center with Nvidia metric 8:10
Q4 2026 Ribbon cutting for the 100-MW data center timeline 9:25

Chapters

  1. 0:00AI boom and energy bottleneck
  2. 2:09AI Energy Management Alliance launch
  3. 4:08Amazon's behind-the-meter strategy
  4. 6:17Emerald AI's valuation and value
  5. 8:01Competition and secret sauce
  6. 9:14Data center delay and proof point
  7. 11:22Independence and long-term vision
  8. 11:41Pacing AI and community concerns
  9. 14:02Rapid-fire preferences

Questions asked in this interview

12
  1. 0:06So, Varun, what are the three biggest things that have changed since we talked in Davos in the AI boom in general?
  2. 1:41According to you, and how will it actually solve the problem of community opposition?
  3. 3:53Do you think Amazon's making a mistake?
  4. 5:29So, is Amazon making a mistake?
  5. 6:17What's the assumption embedded in that valuation that worries you the most?
  6. 8:01So, what's stopping Google or Microsoft or Amazon from doing what you're doing, using AI, and then they won't need you anymore?
  7. 9:14What's up with the delay, and when is the ribbon cutting?
  8. 10:23What's your secret sauce that you can't do with AI?
  9. 11:41Do you ever ask whether faster is always the right objective?
  10. 14:02Would you rather live without power for a year or without AI forever?
  11. 14:15Would you rather always know when someone is lying or when AI is wrong?
  12. 14:23Would you rather be a CEO 100 years ago or 100 years from now?
Amy 0:04 ↗
Great to see you, Varun.
Varun Sivaram 0:05 ↗
Great to see you, Amy.
Amy 0:06 ↗
I think the last time we chatted, it was January. It was a little chillier. We were across the pond in Davos. I was wearing shoes that were a little bit more covering of my ankles. And we were in Davos. And it seems like that was yesterday, but so much has happened since then. And what I like to do for audiences is try to connect the dots between how much change is happening throughout the year. So, Varun, what are the three biggest things that have changed since we talked in Davos in the AI boom in general?
Varun Sivaram 0:36 ↗
Look, it was such a pleasure to be at Axios House in Davos. And some things have stayed the same or intensified, and some trends are new. You know, first, you asked me in Davos, 'Hey, is this AI thing a bubble?' And I said at the time, 'I really don't think it's a bubble. When you have intelligence that is as or better than humans at virtually every task we can conceive, there will not be enough demand for that.' I think that trend has intensified. There's demand for compute. It's not a bubble. Second, the trend that's intensified is energy is the bottleneck. United States wants to connect 70 gigawatts of AI data centers between now and 2028, will only succeed in connecting 30, let's say. It's still the bottleneck for American competitiveness, American innovation. The third thing is has something really changed? If I had to pick one thing, it's the political backlash against data centers, AI data centers. Over the last week, there's been a safety phenomenon. And so, that's the topic that I think really has motivated us to double down on why AI data centers need to be good citizens, good citizens to power grids, good citizens to local communities.
Amy 1:41 ↗
So, you talk about wanting to be a good citizen, a good citizen, and you launched a new coalition, which we talked about earlier with Josh Parker from Nvidia. But, it's pretty wonky. I think it's called the AI Energy Management Alliance. There's a lot of alliances in Washington, D.C. People are laughing because it seems to be one every day. Why is this one different? According to you, and how will it actually solve the problem of community opposition?
Varun Sivaram 2:09 ↗
Well, I'm really proud of this new alliance. Thank you by the way for breaking the story. This alliance brings together Emerald AI, my company, which founded along with Google and Nvidia. And we brought 20 companies, Anthropic on the AI side. On the energy side, AES, Constellation, NRG, National Grid, energy giants. We brought the whole AI and energy ecosystem together for one very specific thing. Power flexible AI data centers. Sounds wonky, right? Super wonky. Making a data center that can control its power or reduce it during the moment when the grid is most stressed out. That moment, the hot summer day in Phoenix, Arizona, or the moment at halftime of a soccer game in England when everyone turns on their tea kettle in the kitchen. Those moments, if you can reduce your power at the data center, mean you can fit far more data centers onto the existing grid. What's really important is if you do that and you're a good citizen, you end up creating these societal benefits. You keep rates affordable for local citizens and local communities. You keep grids reliable so that they don't crash when lightning hits a substation. If the data center is a good citizen, it can control its power, as we've proven many times in six different demonstrations with Nvidia. If you can do this, you actually can have everything at the same time. So, the alliance is supporting a grand bargain. The grand bargain is regulators, governors, elected officials in America and elsewhere, they can offer faster and larger power access to the grid for data centers because if they're flexible, you don't need to build out so much new infrastructure. And in return, data centers that are good citizens are going to help to protect customer rates and keep grids reliable. That's the grand bargain. Massive societal impact for one very narrow wonky topic, flexible AI data centers.
Amy 3:53 ↗
But embedded in that is this idea that people will connect to the grid. We just heard from Kara Hurst from Amazon. They're building this massive natural gas plant along with some solar. That's actually behind the meter. It's not connected to the grid. Do you think Amazon's making a mistake?
Varun Sivaram 4:08 ↗
Look, there are two potential futures ahead of us. I want to be clear about how stark the choices, Amy. One potential future is where AI data centers, the AI industry says, 'My goodness, it's so hard to connect to the grid. They make me wait more than 10 years, and I have to understand all these crazy regulations. I'm just going to go off the grid.' Intuitively, that sounds like a good idea, right? Or I might even go into space with Elon Musk. I don't laugh. This is serious. It's a serious technology. It's going to do well. That's bad for everybody, in my opinion, right? It's bad, frankly, for the AI companies because it's expensive to run your own private utility. It's more expensive tokens of artificial intelligence. It's bad for the public power sector. You suddenly raise the prices of all the inputs, the turbines, and the switchgear that the public power system needs. And most importantly, you've taken away the biggest driver of American economic growth from the public power system, from the utilities, and from our local communities. We should want these data centers on our power systems, but they should be good grid citizens. Citizens that are able to respond to when the system is stressed out, able to help fund the improvements. Local communities need these data centers if they're also going to fund the upgrades that are going to bring on hospitals and advanced manufacturing plants, and neighborhood growth. That integrated future, that second future, that's the one we're pushing towards because the first one doesn't help anybody at all.
Amy 5:29 ↗
So, is Amazon making a mistake?
Varun Sivaram 5:32 ↗
Look, I'm a huge huge fan of Amazon. Amazon's been a long-time leader in the AI and in the data center space, and I think all of us would benefit from learning how they run data centers so well. Going forward, I'm really proud of this alliance, Nvidia, Google, Emerald, Anthropic, for what we're doing on making data centers good grid citizens for the 21st century.
Amy 5:53 ↗
Have you approached Amazon about joining that?
Varun Sivaram 5:55 ↗
I'm a huge fan. You know, Kara Hurst, who you just had, Kara was actually one of the founding members of the First Movers Coalition, which I spearheaded along with Secretary John Kerry and the World Economic Forum. I think that Kara, through many of her sustainability initiatives, has really been a leader. Huge fan of Kara.
Amy 6:12 ↗
So, it sounds like stay tuned if Amazon is going to join.
Varun Sivaram 6:14 ↗
I'd love for Amazon to join us. It would be fantastic.
Amy 6:17 ↗
That's an open invitation. I'll go ahead and ask her after this. So, I want to hear a little bit more about Emerald AI, your company. You've gone from a startup to a billion-dollar unicorn in barely a year. What's the assumption embedded in that valuation that worries you the most?
Varun Sivaram 6:34 ↗
Well, look, today, Amy, is a super special day for us. Two years ago to the day was when I went to investment committee for our founding investor, Radical Ventures. I think of my co-founder, Rob Teed, who's somewhere here in this audience. And that was the day that they took a real bet on Emerald. And over the last two years, as you mentioned, we just became a unicorn. I'm really excited. And you asked, 'Why do we deserve a valuation like that?' And I don't like to talk that much about why Emerald is so valuable. What I want to talk about is what Emerald unlocks trillions of dollars of value in the economy. Because there's 100 gigawatts of available power that we can fit data centers on our existing grid if they're just a little bit flexible. That's a massive number, right? That's 100 nuclear power plants. That's more data centers than we're trying to connect in the next three years. All of that you can connect overnight. Not make the data centers wait 10 years while they're good citizens. Emerald enables the data centers to achieve that revenue. It's $10 million of profit for every megawatt that we get on the grid. We enable the utilities to keep paying customers. Don't get disintermediated. The utilities actually get great customers. You know, National Grid, who's one of our investors, we now have 12 Fortune 500 global investors in addition to Nvidia. National Grid says, 'You're going to take data centers and convert them from this entity that I'm kind of scared to connect to an entity that I'd love to connect. I'm going to compete with other utilities to connect them because they're such good citizens.' That's the value that we're seeking to unlock, $4 billion.
Amy 8:01 ↗
So, what's stopping Google or Microsoft or Amazon from doing what you're doing, using AI, and then they won't need you anymore?
Varun Sivaram 8:10 ↗
You know, first of all, Google has been a pioneer in this field. We're really big fans of what Google's done. And we've just tried to put together the world's best team on AI-power flexibility. The Emerald Conductor platform, it's our software. We use AI agents in these data centers. It's almost like a recursive meta AI for AI. These AIs are helping to make sure that the data center can reduce its power when it's really hot in Phoenix, let's say, as we published in the journal Nature Energy, while we protect the performance of the mission-critical AI workloads. That's really hard for a human to do, but our AI agents are actually able to do this. So, we are just pushing forward the scientific frontier on what's possible, connecting the energy system and the data centers. We hope that we'll get to work with all of these different partners. Nvidia has been an absolutely wonderful one. Later this year, you broke the story, Nvidia and Emerald are launching a 100-MW AI data center, the world's first truly power-flexible AI factory. It's going to be a community-friendly, good grid citizen. I hope all of you will join us at the ribbon cutting.
Amy 9:14 ↗
So, when I first reported that, it was going to be the first half of 2026. It's now September 21st. What's up with the delay, and when is the ribbon cutting?
Varun Sivaram 9:25 ↗
Well, when you first wrote about it, the picture you published in the scoop that you broke was a picture of dirt, right? There was nothing there. And ever since, you watch this building take shape with massive power infrastructure and GPUs, Vera Rubins that are being wheeled in, it is a marvelous, it's an amazing exquisite machine. So, yes. It is coming online second half of this year. Q4 is going to be the ribbon cutting. We're very very excited. But, what I want to emphasize here is this is coming online in 2026. It's a real proof point. And if you're a regulator, if you're a utility, and you're wondering, 'Hey, can these AI data centers really be good grid citizens?' I know they say they can. This whole alliance just got created with three of the 10 world's 10 biggest companies. But, will they really show up when I'm in a pinch and it's a hot summer day? The answer is yes, and you're going to see the real proof. This 100 megawatt facility is going to respond to whatever the grid needs it to do perfectly.

20 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, September 22). AI’s next bottleneck is energy. Varun Sivaram has a plan [Interview transcript]. Axios Live and Axios. CEOInterviews.AI. https://ceointerviews.ai/interview/2916252/

MLA

Varun Sivaram. "AI’s next bottleneck is energy. Varun Sivaram has a plan." Axios Live and Axios, 22 Sep. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/2916252/.

BibTeX
@misc{sivaram2026_2916252,
  author       = {Varun Sivaram},
  title        = {AI’s next bottleneck is energy. Varun Sivaram has a plan},
  howpublished = {Interview transcript, Axios Live and Axios. CEOInterviews.AI},
  year         = {2026},
  month        = {sep},
  url          = {https://ceointerviews.ai/interview/2916252/},
  note         = {Speaker-attributed transcript with timestamps}
}