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

Newsom Signs Order Requiring AI Labs Develop ‘Kill Switch’ | Balance of Power

📅 Sep 18, 2026 Bloomberg Podcasts 40 MIN 63 SEGMENTS · 9 SPEAKERS
Emerald AI CEO Varun Sivaram. Stonecourt Capital Partner Rick Davis and Harvard Kennedy School Ash Center Visiting ...

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 how his company transforms AI data centers into flexible grid allies, reducing power during stress to avoid costly infrastructure and keep rates low. He cited partnerships with Nvidia and Google, including a 100-megawatt power-flexible data center in Manassas, Virginia, coming online later this year, and six demonstrations in London, Phoenix, and Virginia. Sivaram announced the AI Energy Management Alliance (AMA) with Google, Nvidia, and 20 partners like Anthropic and AES, aiming for a 'grand bargain' with regulators for faster grid connections. He noted the US can only connect 30 gigawatts of the 70 gigawatts needed by 2028, while China will have 400 gigawatts of spare capacity by 2030. He emphasized autonomous flexibility, as demonstrated in London during halftime tea kettle spikes, and dismissed concerns about AI security, focusing on energy as the defining issue.

Key takeaways

  1. Emerald AI's flexible data centers can use 100 gigawatts of otherwise unusable grid power, helping keep rates low and grids reliable.
  2. The US can only connect 30 gigawatts of the 70 gigawatts of AI data centers needed by 2028, while China will have 400 gigawatts of spare capacity by 2030.
  3. Emerald AI, with Nvidia and Digital Realty, is building a 100-megawatt power-flexible AI data center in Manassas, Virginia, coming online later this year.
  4. Emerald AI launched the AI Energy Management Alliance with Google, Nvidia, and 20 partners, including Anthropic and AES, to push for faster grid connections.
  5. Emerald AI and Nvidia demonstrated autonomous power reduction in London during halftime tea kettle spikes, protecting AI computations and grid stability.

Numbers and commitments

FigureWhat it refers toTypeAt
100 gigawatts Power that flexible data centers can use on today's grid metric 8:13
70 gigawatts AI data centers the US wants to connect by 2028 guidance 11:55
30 gigawatts AI data centers the US can actually plug in by 2028 guidance 11:55
400 gigawatts China's spare capacity for AI by 2030 metric 11:55
100 megawatts Size of the new power-flexible data center in Manassas, Virginia metric 9:21
20 Launch partners in the AI Energy Management Alliance metric 10:24
12 Fortune 500 companies that are investors in Emerald AI metric 9:21

Chapters

  1. 0:00Flexible data centers as grid allies
  2. 8:13How flexibility works in practice
  3. 9:21Retrofitting existing data centers
  4. 10:24Regulatory engagement and the grand bargain
  5. 11:55AI security and the energy race with China
  6. 13:44Autonomous flexibility in action

Questions asked in this interview

6
  1. 8:03Well, what gives though if you have to have flexibility?
  2. 9:11Do you have to retrofit all of them to be part of this new model or is this simply about the next generation?
  3. 10:09I just wonder what your conversations on the government side?
  4. 11:29How do we control this technology?
  5. 12:22But do you worry about AI security interfering with the work that you're trying to do on the energy side?
  6. 13:25Like, is the technology that is actually being powered by these data centers also factoring into the way in which they are flexible?
Narrator 0:01 ↗
Bloomberg Audio Studios, podcasts, radio, news. You're listening to the Bloomberg Balance of Power podcast. Catch us live weekdays at noon and 5:00 p.m. Eastern on Apple CarPlay and Android Auto with the Bloomberg Business App. Listen on demand wherever you get your podcasts or watch us live on YouTube.
Joe 0:25 ↗
We look to close out a wild week that has been in the markets. Of course, it was just this week that we got the Fed decision, which already feels like a lifetime ago. It was also just this week that we came into the week with conversations around potentially the pacing of artificial intelligence. The heads of some of the biggest companies in the industry all having agreed last weekend, of course, that that may be necessary. President Trump fervently disagreeing with that over the course of the week, especially in light of the Anthropic former employee's public resignation talking about how it could lead to the extinction of humanity. President Trump and many of those close to him have cast this notion as being a hoax, a conspiracy, a Chinese plot. Just listen to what President Trump has said about this throughout the course of the last week and change.
Donald Trump 1:12 ↗
Do you have any concerns about AI leading to human extinction? No, I don't have any. I have concerns that if we don't win AI, we're going to be put in a very bad position. We're leading China in AI. We can put guardrails. We can do this and that. But I think you have a lot of negative forces that are bringing it up that shouldn't be bringing it up.
Joe 1:33 ↗
Well, they're going to bring it up at least the issue of AI and security next week. As we mentioned, when President Trump meets with President Xi, the summit here in Washington, whether this will resolve in some framework or AI security agreement has to be yet to be seen and it's unclear exactly what the president's goals are. But it will be a who's who of AI executives as well, which is on brand for this White House. Sam Altman will be there, Jensen Huang, Cristiano Amon from Qualcomm, and it's being reported that Jamie Dimon will be there, and I'm sure that's going to be a much longer list. And Tyler Kendall will be on hand to cover it all on the heels of the United Nations General Assembly where the president will be as well at the start of next week. Tyler is our Washington correspondent and joins us now live from the White House. What are the president's goals when it comes to President Xi? Tyler, do we have a sense of what this agenda will look like?
Tyler Kendall 2:30 ↗
Well, Joe, at this point, we've heard from a few analysts that if the US and China are able to extend the trade truce, keep the calm, that in and of itself would be an accomplishment of a goal. As we learn more, including reporting from the Financial Times, that the US may only be looking to extend that truce by about 6 months amid what are heightened concerns with Beijing's loosening or lack of loosening of export curbs related to critical and rare earth minerals. If you went through it though, there are a few points of contention that will probably come up, including front and center when it comes to artificial intelligence because this White House's main argument against an industry-wide slowdown has to do with the competition with China. Just last week, we heard from three national security agencies alleged that Chinese AI firms are essentially stealing American IP through so-called distillation attacks. But if we put the fuller picture together, there's so many different things that this White House has to weigh when approaching this meeting with Chinese President Xi Jinping, including when it comes to those export curbs, not wanting to escalate on one front and have China retaliate on another. It is a balancing act next week. But the fact that Xi Jinping is expected to be here does play to the administration's broader goal of keeping relations with China, at least right now, on balance.
Kaylee 3:49 ↗
Well, and as you talk about an extension for some period of time of their trade truce, Tyler, Bloomberg is reporting that we won't see any tariff action until after this meeting is concluded.
Tyler Kendall 4:00 ↗
So, at this point, it's unclear exactly the reason behind this potential postponement of higher what was expected to be higher tariffs, but it could be perhaps a leverage tool for the White House as we go into the talks next week. A few points of context here. One, these were planned tariffs over allegations of Chinese overcapacity as the country ran a $1.2 trillion trade surplus next year. The tariff rate was supposed to come in at 7.5%. I bring that up because it would have brought the average effective US tariff rate against Chinese goods under the second Trump administration to 20%. Which of course is a relatively large number, but it actually hits that cap that has been previously established with China under the current trade truce. So, it actually wasn't really considered to be that major of an escalation. That is raising questions about exactly why these tariffs were postponed. Will there be a different rate depending on how the conversations go? Kaylee and Joe, we may get more details on Monday. The US Treasury Secretary is set to meet with his Chinese counterpart over the weekend.
Joe 5:04 ↗
Tyler, thank you so much. Live from the White House, Bloomberg's Tyler Kendall with much more to discuss ahead of next week. It'll be a very busy one around here with UNGA at the start of the week and of course the Xi Summit at the end of the week here on Balance of Power. We'll have it all covered. Tyler is on her way up to Turtle Bay shortly after the close of business today. As a matter of fact, Kaylee mentioned one of the biggest and most important stories of the week when you talk about an unproductive Congress. Kaylee, it was this Congress that passed the data center bill. They actually got it done and it was an overwhelming vote as it turned out in an effort to control the impact on people's utility bills. But if you remember, this thing doesn't really take effect for 2 years and it's not going to do much between now and November.
Kaylee 5:48 ↗
Yeah, it may be more signal than anything trying to show voters that lawmakers are serious about making sure that there aren't detrimental effects on communities due to data centers. While the House passed this with a massive bipartisan margin, if and when it passes the Senate, on what timeline is another question, and again the timeline for it to actually have an impact is an open one too. But we are starting to hear consistently from lawmakers of both parties that if you want to have data centers built, you have to do it in a way in which it's beneficial to communities. They bring their own water, they bring their own power. This seems to be what everyone is coalescing around.
Joe 6:23 ↗
That's true, which was not the way it went when data centers started popping up out of the earth. You're from Virginia. I live in Virginia, the data center capital of the world. We've learned a lot over the past many years on this, but this has become a huge issue. And there's a slightly more clever solution that didn't come from Washington that we want to talk to you about and it involves a company called Emerald AI and two that you may have heard of, Nvidia and Google. I'm glad to say that we're joined right now by the CEO and founder of Emerald AI, Varun Sivaram, who's with us right now in our studio in Washington. Varun, thank you for joining us here on Bloomberg TV and radio.
Varun Sivaram 6:59 ↗
Joe, Kaylee, thanks for having me.
Joe 7:00 ↗
So, we've illustrated the problem. I want to hear how you've managed to solve it because it's a more nuanced approach than what we've heard from lawmakers and regulators here in Washington.
Varun Sivaram 7:11 ↗
Exactly. Emerald AI, which we founded under two years ago, aims to transform AI data centers into flexible allies of the power grid. And the reason this works is if an AI data center can reduce its power consumption during those moments when the grid is really stressed out, a hot summer day in Phoenix, Arizona, let's say, then that data center can actually fit onto the existing power grid without having to build out expensive infrastructure that can have the effect of raising rates. In fact, flexible AI data centers, all else equal, can help to keep rates low and affordable for local communities. And they can help to keep grids reliable all while helping America gain a technological edge in the race for AI innovation as we get more data centers connected on the grid. So, it's a win-win-win thanks to this technology of flexibility when you need it.
Kaylee 8:03 ↗
Well, what gives though if you have to have flexibility? If there are moments at which the grid is strained and they have to bring it down, I would imagine that that means that you aren't necessarily powering all the compute you would be otherwise.
Varun Sivaram 8:13 ↗
There are many ways to get flexibility. I'll give you some examples. So, our company, Emerald AI, has worked with giants like Nvidia to enable a data center to slow down some of the computations that it's running. You might have a fine-tuning job that you can slow down or pause, or you can take a job and move it at the speed of light from one location to another. So, Kaylee, you're talking to ChatGPT and suddenly you're getting your answer not from Virginia, but from Chicago, as we've demonstrated at two different Oracle data centers on NVIDIA chips. In addition, you could run a battery that happens to be located on site. There are many ways to achieve this flexibility, but the goal is there's actually a lot of power on today's power grid that we just can't utilize because today data centers require the use of power potentially 24/7. If they're able to be flexible and what we call good citizens on the power grid, well, they can use 100 gigawatts that we otherwise can't use on the power grid today. And that helps with the affordability concerns and with the speed to power.
Kaylee 9:11 ↗
This sounds really smart, Varun, but what about all of the existing data centers? Do you have to retrofit all of them to be part of this new model or is this simply about the next generation?
Varun Sivaram 9:21 ↗
Well, so the way that Emerald AI works is we work both with existing data centers and with new data centers. We are, you know, Nvidia is a big partner and investor in Emerald as are 12 other Fortune 500 global companies. And together with them, we have models where for an existing data center, our software, the Emerald Conductor platform can help an existing data center change how it consumes energy. And we're building a new data center with NVIDIA and Digital Realty, the world's first from the ground up power flexible AI data center, AI factory. It's 100 megawatts in Manassas, Virginia. And it comes online later this year. So new data centers can get faster connections to the grid. Think of a fast lane. And existing data centers can get even more power much more quickly.
Joe 10:03 ↗
It's interesting, Kaylee, because we talk so much about taking data centers off the grid, was behind the meter, now we need them on the grid.
Kaylee 10:09 ↗
Well, and I've actually spoke to the chair of FERC about this and the ways in which the grid needs to adapt to all of this. Are you in conversations with regulators? I just wonder what your conversations on the government side?
Varun Sivaram 10:24 ↗
Yeah, absolutely. We think that there is a grand bargain to be had where if data centers can work with regulators in the energy industry and say, look, you can actually have it all. You can have more data centers, more innovation, you can have lower rates for communities and you can have more reliable grids. There's a grand bargain to be had. Let's connect these flexible data centers faster. That's why this week we're so proud to have launched the AI Energy Management Alliance. Emerald, Google, and Nvidia are the founding members of the AMA, this alliance, and we're joined by a cohort of 20 launch partners: Anthropic, leading AI lab, and Analog Devices, leading semiconductor firm, and leading energy companies AES, NRG, Constellation, RWE, National Grid. And together we're presenting a unified front. You mentioned FERC, Kaylee, we're talking to the FERC commissioners, we're talking to state regulators, we're talking to the administration to say, here's this grand bargain, let's make sure that AI data centers can be good citizens, good citizens to grids, good citizens to communities. And in return, they should get faster and larger connections to the power grid.
Joe 11:29 ↗
What do you make of some of the AI stories that we woke up to today? I know this isn't data center related, but you've been in this industry since it really started to kick up. I met you at the Nvidia conference here about a year ago in Washington DC. We weren't talking about stories like these about Anthropic for instance, self-learning. Claude is doing 26% of its own research and development. Claude is also hacking ChatGPT. How do we control this technology?
Varun Sivaram 11:55 ↗
Well, Joe, as you know, the reason I got into this business to begin with is because I think energy has now become the defining issue for the AI revolution. If you look at China, for example, China has plenty of energy. By 2030, they're going to have 400 gigawatts of spare capacity for AI. The United States, however, does not. We want to connect 70 gigawatts of AI data centers over the next 2 years between now and 2028. We can only really plug in 30.
Joe 12:22 ↗
But do you worry about AI security interfering with the work that you're trying to do on the energy side? People are changing their opinions as we're having this conversation.
Varun Sivaram 12:29 ↗
Well, look, it is absolutely critical that if the United States is going to compete, we're going to need data centers here in this country. And those data centers, we want to make sure through this new alliance and through Emerald AI's work, we want to make sure that those data centers are good citizens. The communities that host them deserve as much. If you're a community and you've got a data center coming to town, you want to be sure, hey, that data center is going to bend over backwards so that I don't have blackouts. I don't have reliability issues and my bills can stay affordable because if not, we start to face trade-offs. Data centers could come to town and you might say, well, do I have the space on the grid to also bring in advanced manufacturing or hospitals or neighborhood energy growth? Now, with flexible AI data centers, you can have it all. And that's now not just Emerald AI saying this. It's now 20 leading companies bringing the same unified message of this grand bargain to regulators across the country.
Kaylee 13:25 ↗
Can I just ask, as one of the stories Joe's referring to is that AI is improving upon itself without necessarily human direction. How much of this flexibility would be driven by automated decision-making? Slow down here, send it there. Like, is the technology that is actually being powered by these data centers also factoring into the way in which they are flexible?
Varun Sivaram 13:44 ↗
It's such a great question. How do we make these new devices, these large data centers, how do we make them flexible? I first want to say this is not science fiction. This is actually proven. You know, Google, one of the founding members of this new alliance, has done a gigawatt of demand response for its data centers. Emerald and Nvidia have now done six demonstrations around the world in London, in Phoenix, in Virginia. And we now have, as I mentioned, this 100 megawatt facility coming online that can flex. I'll give you an example.
Kaylee 14:10 ↗
Are humans directing the flexion?
Varun Sivaram 14:12 ↗
Great question. I'll give you an example. In London, it turns out at halftime of the soccer game, everybody watches in Britain these soccer games, football they call it, and everyone turns on their tea kettles at halftime, right? And so there's a spike of a gigawatt of power onto the grid at that moment. Now, in that moment, or you know, a lightning strike hits the grid and takes out a substation or a line. In that moment, no human can act quickly enough in order to modulate the power in that data center. And so what we demonstrated, Emerald, with Nvidia and the local utility National Grid, we demonstrated that data center was able to autonomously reduce its power consumption, but do it in a very special way. Do it in a way where the computations, the AI computations, would continue to run, the mission critical computations continue to run at the right performance level. And so we protected the AI, we protected the grid, rates stay low and communities have resilient electricity.
Joe 15:02 ↗
We ran out of time to ask you when you're running for office or when you're going public. Let's do that next time.
Varun Sivaram 15:07 ↗
Joe, thank you for having me.

31 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 18). Newsom Signs Order Requiring AI Labs Develop ‘Kill Switch’ | Balance of Power [Interview transcript]. Bloomberg Podcasts. CEOInterviews.AI. https://ceointerviews.ai/interview/2044060/

MLA

Varun Sivaram. "Newsom Signs Order Requiring AI Labs Develop ‘Kill Switch’ | Balance of Power." Bloomberg Podcasts, 18 Sep. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/2044060/.

BibTeX
@misc{sivaram2026_2044060,
  author       = {Varun Sivaram},
  title        = {Newsom Signs Order Requiring AI Labs Develop ‘Kill Switch’ | Balance of Power},
  howpublished = {Interview transcript, Bloomberg Podcasts. CEOInterviews.AI},
  year         = {2026},
  month        = {sep},
  url          = {https://ceointerviews.ai/interview/2044060/},
  note         = {Speaker-attributed transcript with timestamps}
}