1. Podcasts
  2. This Week in AI
  3. Data Centers Aren't Killing the Grid. They Could SAVE It.
This Week in AI

8 Oct 2026 · 45m · 19 chapters

Episode artwork

Data Centers Aren't Killing the Grid. They Could SAVE It.

In short

Emerald AI argues that data centers aren’t “killing” the power grid; instead, they can help stabilize it by acting as flexible, “good grid citizen” assets that respond to peak-demand signals.

Key claims

(1) The main blocker for new AI data centers is energy availability, not land or construction labor. (2) The grid lacks standardized electronic signaling to data centers; Emerald developed a standard for grid-to-data-center communication. (3) Flexibility—not just efficiency—is required to fit more data centers without building “the ninth lane” of new transmission/substations. (4) Data centers should use intelligent workload orchestration (throttling, rescheduling, moving inference) plus batteries/generation as needed.

Notable examples

Phoenix peak-demand throttling; London halftime “tea kettle” spike; Virginia cold snap moving load to Chicago; Portland heat dome; California “duck curve”; a planned 100MW flexible AI data center in Virginia with NVIDIA and Digital Realty.

Guests

Varun Sivaram, founder/CEO of Emerald AI (ex–Fortune 500 energy producer strategy/tech leader). Also discussed: Vibha Viswaathan, co-founder/CEO of Pascal AI (builds context-aware AI agents for institutional finance; product in production with 12+ customers; ~$500M AUM among customers).

Written by AI. May contain mistakes.

Chapters

0:45

The Evolution of Data Centers

Exploring the changing perception of data centers and their energy consumption.

“But key issue is, hey, is this going to affect my energy bill?”
1:28

Emerald AI and Energy Efficiency

Discussion on how Emerald AI's platform helps data centers manage energy usage.

“Or with the Emerald Conductor, because that's your product.”
3:09

Flexibility and the Grid

Understanding the importance of flexibility in data centers for grid reliability.

“And that means large swaths of batteries.”
5:59

Dynamic Response and Standards

Exploring the need for standards in grid communication and workload management.

“because suddenly the grid says, oh my God, I can fit you.”
11:14

The Grand Bargain for Data Centers

Advocating for policies that encourage flexible data centers to enhance grid reliability.

“There's a Duke University research paper.”
14:00

The Role of Data Centers as Good Grid Citizens

Learn how data centers can positively impact local grids and communities.

“but you also would love for the communities to welcome them.”
17:43

America's Energy Challenges and AI Data Centers

Understand the challenges America faces in expanding its energy capacity for AI.

“And just so you know, China is going to have 400 gigawatts of spare capacity at peak in the year 2030 for AI, whereas the United States is already at a deficit unless we start to.”
23:14

Transformative Potential of AI Data Centers

Discover how AI data centers can transform energy utilization and efficiency.

“What percentage do you think of efficiency is there to wring out of the existing grid with software, with solutions like yours?”
26:18

Community Concerns and Political Responses

Explore community concerns about data centers and the political landscape surrounding them.

“And we're really proud to work with NVIDIA and others on them.”
28:03

Data Center Sentiments and Community Concerns

Exploring the apprehensions communities have regarding AI data centers and the response from the industry.

“Are they the most anti now, or is it New York?”
Show all 19 chapters
30:00

Introduction to Pascal AI

Introducing Vibha Viswaathan, co-founder and CEO of Pascal AI, and the focus of the discussion.

“Next up on the program, we've got a really interesting AI startup that's designed to help hedge fund managers and asset managers.”
30:13

Building AI Agents for Institutional Finance

Discussing the development of AI agents specifically designed for hedge funds and asset managers.

“He is the co-founder and CEO of Pascal AI.”
31:11

Contextual Understanding in AI Solutions

The importance of context in building AI solutions for finance and how Pascal AI addresses it.

“Because hedge fund managers have been using machine learning and AI for decades.”
32:45

Real-Time Decision Support for Investors

How Pascal AI provides real-time insights and decision support for portfolio managers.

“So we take decision traces of various names in your portfolio.”
34:14

Differentiating Pascal AI from Generic AI Tools

Explaining how Pascal AI differs from tools like ChatGPT by focusing on institutional knowledge and context.

“So maybe you have 20 different sources, which is part of your institutional memory.”
35:38

Temporal Awareness and Financial Insights

Understanding how Pascal AI uses temporal awareness to enhance decision-making for hedge funds.

“So what's the question being asked here?”
37:39

Data Sources and Proprietary Information

Discussing the various data sources Pascal AI utilizes, including proprietary data from hedge funds.

“and different ways of working in the institutional finance world is what is your time horizon for investing?”
40:40

Customer Adoption and Competitive Edge

Insights into why hedge funds are choosing Pascal AI over building their own solutions.

“So that really informs the agent on what the internal context and thinking is and basis that we're able to pull out signals from the license feeds and the world, you know, the internet.”
42:01

The Future of AI and Hedge Fund Managers

Explore how AI is transforming the roles of hedge fund managers and their work processes.

“I think that was the way people phrased it pre-LLMs.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On This Week in AI, what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00All right, everybody, welcome back to This Week in AI. With us today, Varun Sivaram, Sivaram, got it, is the founder and CEO of Emerald AI. They're building software to help data centers cut their power. Varun, welcome to the program. Jason, thanks for having me. Yeah, so data centers are hated. Suddenly, they went from locations begging and giving tax breaks for people to build a data center. They were being courted, and then something changed. Maybe there were too many data centers, maybe a little bit of noise pollution on the margins, maybe a little bit of gas pollution on the margins. And of course, key issue, aesthetics and water on the margins.

0:45But key issue is, hey, is this going to affect my energy bill? So here we are. You have a company that's helping people lower their energy usage, which I think would address some of those issues. But the big blocker anyway for data centers is energy consumption. It's easy to find land. It's not easy to find construction workers, but you can get that done. But the energy seems to be the high order bit. So tell me a little bit about what you're building at Emerald Conductor. Thanks to our friends at PayPal, the exclusive sponsor of This Week in AI. Pay zero processing fees on your first 100 ,000 ineligible PayPal payment volume.

1:25Find out more at paypal.launch.co. Absolutely, Jason. Or with the Emerald Conductor, because that's your product. That's right. Emerald AI is the company. We make the Emerald Conductor software platform. And I want to be clear up front, Jason, that the goal is not for AI to use less energy. Actually, AI is going to use a whole lot more energy. The goal is to enable us to unlock the power grid for AI to use a whole lot more energy, while at the same time keeping local communities' bills affordable and the electric grid reliable. And there's only one way that you can achieve all three of these things at the same time, more AI data centers on the grid, but that act as good citizens, friendly to local communities.

2:07And that's for those data centers to be flexible. So of course, efficiency is important. The NVIDIA Verirubin generation today produces a million times more tokens per watt of input energy than the Kepler generation in 2012. But flexibility is a different concept altogether. Flexibility means that every so often, rarely, when the grid is really stressed out, when it's a hot summer day in Phoenix, Arizona, and everyone's air conditioning is turned on and the grid is facing its peak demand, during those moments, if the data center can reduce its power consumption because it's flexible and controllable using our software, then that data center can actually fit on the existing grid that we already have.

2:50and that data center's revenues that it brings into the power system help to go to pay for the electric grid we already have and the upgrades that it will need to serve the local community. So again, the only way you get more data centers, affordable bills and reliable grids is for data centers to be flexible. And that's the technology that we do to make data centers grid assets and not liabilities. And that means large swaths of batteries. And so is what I'm seeing here in this looped video on your website, Emerald AI, I, are those giant batteries I'm looking at right here, these big black tanks that you have?

3:25To be clear, it's not just about batteries. You're looking at GPUs. But in addition to, of course, batteries are one way to do this, right? So you can have a data center. And during that hot summer day in Phoenix, the data center has got to reduce its power consumption. And one way you could do that is by having massive banks of batteries, right? But another way you can do it is you can slow down some of the processing that's happening on site during that rare moment. or you can even, you can pause jobs, you can slow down GPUs, or you can move those jobs from that location, say Virginia or Phoenix to a different location, say Chicago at the speed of light.

4:00And if you orchestrate the compute intelligently, we, we at Emerald believe that that is one of the key ways that data centers can be flexible grid assets. And in addition, you can have batteries, for example, that can supplement that flexibility from the computations. So you're more in the software business than the battery business here. You're looking at it saying, hey, the grid is telling us, the grid intelligently tells these data centers, hey, we're at peak consumption. There's some sort of input. Is there like a standardization for how that occurs and how it gets communicated to people using electricity?

4:38Yeah. Believe it or not, there isn't, Jason. So if I walk back. No standard. The grid is over a century old. And the grid also has 3 ,000 different utilities across America. And so it's very balkanized. In some parts of the grid, in fact, many parts of the grid, so many that I'm embarrassed, the way you do this is you pick up the phone. So somebody will pick up the phone and call you and say, hey, we'd love for you to reduce your power consumption. What? That's insane. Now, there's no reason that it has to be that way. And so we'll walk through what Emerald's done. But in the field, we have demonstrated through our software and signaling platform, we've actually developed a standard for the grid to electronically communicate with data centers.

5:19We showcase this, for example, in London, where the national grid utility sent signals via our portal and the data center using our software received those signals. And the Emerald Conductor made sure that the data center was reducing its power during peak demand. Now, in London, this was not on a hot summer day. This happened to be at halftime of a soccer game, where in London, at halftime, all the British people go and turn on their tea kettles. And at that moment, it spikes a gigawatt of load on the grid, and the data center we proved could reduce its power consumption during football halftime.

5:53So wherever you are, there'll be some idiosyncratic reason that the grid needs you to power down a little bit. And that ability to dynamically respond is how we make the data center the grid's best friend. because suddenly the grid says, oh my God, I can fit you. And I don't have to over-provision my grid for that crazy worst case peak. Now, this would require some intelligence and on the application layer. So if you're going to drop down, you know, the ability of those GPUs in real time, well, that could have a profound impact on, I don't know if it was Cursor whose servers were on there. Now every developer in the world, their response time goes down 50%.

6:36Well, that's not acceptable either, but going down 10 % might be unnoticeable. So is that what's happening here? It says to the people who are using it, hey, on an intelligent level, if this was chat and GPT or Gemini, obviously you probably don't have Google as a customer, but I don't know that you don't. they might not have a problem with Gemini being a second slower, or ChatGPT might not be so offended if images, let's say, or videos came in five seconds less. Is that what's going to happen here on the back end? Yeah, Jason, you said this exactly right. Intelligence is very important. If you didn't do this with intelligence, you might, for example, unplug some servers, or you might blindly throttle the entire data center, and you would have these impacts on particular workloads that were just unacceptable for the customers, as you mentioned, right?

7:27And so you need the intelligence. We think about Emerald AI as an AI for AI data centers. You're intelligently throttling workloads that can tolerate some of that throttling. Maybe, you know, this fine tuning run for this LLM can be throttled or rescheduled. Maybe this batch inference process has some temporal flexibility. And maybe even there's a serving inference query. You mentioned Jason, you mentioned, hey, a chat GPT query. We've showcased with two different Oracle data centers that we can move inference queries from one location to another, again, at the speed of light. That's how fast the query moves in such a way that the ultimate latency that the end customer faces is completely acceptable, right?

8:07The time to first token basically didn't change. The query response time basically didn't change. And so the intelligence is the emerald conductor saying, what are all of the different workload types that I'm confronted with? I've received a grid signal. And what's the best way to preserve the customer's quality of service while still precisely meeting what the grid needs me to achieve? That level of intelligence is the win-win opportunity, right? It keeps these AI workloads are, you know, America's now most valuable product will be tokens of artificial intelligence. So you don't want to compromise on the - In some cases, in some cases, it's goofy nonsense that has absolutely no imperative to be done immediately.

8:49So is there a standard on the other side? We talked about the standard for the grid letting the data center know, hey, could use this amount of alleviation or just needs alleviation, but maybe it could be even more intelligent how much it needs. But then on the other side, is there a standard for workflows to say, I'm optimizing here for price and time at this ratio, or this is a consumer free workload. This is a consumer$20 a month account. This is a consumer$200 a month account. And then this is a corporate customer with an SLA. Those are like five different levels of usage. And is there a standard there for somehow to have the jobs know one, two, three, four, five in terms of SLA?

9:40Yeah, the short answer is partially. But Jason, your point's very valid. We're trying to reinvent two different industries practices at the same time. So we'll start with the electric power industry. There, today, the product is the utility gives you 24-7 always on power, and you will never be throttled. That product's worked reasonably well, right? So you at your house, when you turn on your light switch, you know you're allowed to turn that on. Now, data centers would take the deal. If you could get connected 10 years ahead of schedule or get an additional 100 megawatts overnight, they take the deal that every so often on that hot summer day, you would have to be throttled.

10:15But that's a new class of service that the electric grid has to offer you. And we, Emerald, are working with some of our partners, NVIDIA, Google, and others, to advocate for policies that create this new class of service. What is that class of service called? Contingent-based energy? What is it called? Yeah, it differs around the country. I'll just share. We just launched the AI Energy Management Alliance with NVIDIA and Google and Emerald to advocate for exactly this. You can call it interruptible service, curtailable service, non-firm service. We're developing the jargon right now. But that's one side, right?

10:48Interruptible is the best. Interruptible service. So we'll take the energy, but under the condition it can be interruptible. And then the frequency and duration of the interruption, I guess, is the detail there. We're interruptible for 100 hours for X gigawatts. How do you quantify acceptable, interruptible. Great question, Jason. And then I'm going to come to the other side, the AI side, and talk about your question about is there a standard there for different classes service? But just on your question here, how do you quantify this? There's a Duke University research paper. There's also a McKinsey paper.

11:20What they've quantified is there's about 100 gigawatts of available power on our electricity grid today with reasonably minimal curtailment, 2 % or less of the time. So there's close to 9 ,000 hours in the year and just a couple hundred hours of that time, if you are able to reduce your power consumption by a small amount, it could be 25%, let's say, for a couple hours at a time during that hot summer afternoon, that's what unlocks all this capacity. Think of the grid, Jason, as a highway, right? The highway faces peak rush hour two times a month. And the rest of the time, there's plenty of power plant capacity, plenty of power line capacity.

11:56So these lanes are empty. And so when the next data center comes, you can either make it reduce its power during those rare hours, or you can build a ninth lane of the highway. The ninth lane of the highway is the thing, the new power lines, the new substations. That's what takes 10 years and makes a data center wait in line. And that's what raises everybody else's bills because we all pay for that ninth lane of the highway. And in fact, there were efforts during rush hour in major cities to convince and give discounts and tax breaks to major employers to have people start at 6 or 7 or 8 a.m. as opposed to 8, 9 or 10 a.m.

12:31or start at 10 a.m. or 11 a.m. and finish at 7 p.m. and just try to load balance it. So it's not an uncommon concept here in terms of load balancing. And I would think if you're building a data center and somebody came to you with an interruptible offer to start now, and it was for of the 900 hours a year, let's say 5 % of the time, 500 hours or so, you could say, well, wait a second, even if those 500 hours occurred, I don't know, 10 hours in a row, which probably isn't the case, but if they even did, you could build enough battery power or redundant power, gasoline, kerosene, whatever the generators are.

13:14You could have multiple different types of generators, I suppose. You could actually build for that inevitability, yeah? Kind of. We continue to believe that a combination of approaches is probably optimal. That's why this AI Energy Management Alliance that we just launched advocates for technology neutral approaches, whether it's computational flexibility, a bunch of batteries, right, other generators, or you put them all together. Maybe you have a small battery, plus you're also flexing your GPUs, which might be the most economical way to prepare for all of these potential events over those couple hundred hours in the year.

13:46The goal is, yes, we can figure out the way in which we will achieve that throttling every so often. But there's a grand bargain to be had. If you're a regulator, if you're the Trump administration, if you're a state official, a governor, you want more data centers in your territory for the economic development, the innovation competitiveness. but you also would love for the communities to welcome them. And for the communities to welcome them, these data centers have to be what we call good grid citizens or good citizens to communities. So that's why we're championing this grand bargain. Allow these flexible, friendly data centers to jump the line, to get a fast pass onto the grid and get faster and larger connections.

14:23And in return, they will act in this way, as you just laid out, they'll act in this way for whatever your local idiosyncratic thing is. We've now done six demos around the world, right, at real data centers. In Phoenix, it was a hot summer day, right? In Virginia, it was a cold snap in the middle of the winter where we had to throttle a load and move it to Chicago and other data center. In Portland, Oregon, there was a heat dome. In London, it was the tea kettle thing, right? In California, it's something called the duck curve where so often the solar shuts off, every day it shuts off. And there's a spike in electricity that needs to be met.

14:57And we showcase of the day. Everybody comes home. I mean, in California, the duck, and we have it here in Texas as well. And they'll pull up the duck electricity curve here in a second because there's a million drawings of it online. People come home from work. They do a load of laundry. They open their refrigerator and they plug in their electric vehicle in the same two hours. And then they turn up, they turn the AC back on. Now, this sounds crazy. Like, why would they do that. But in the middle of the day, you're not home. So most people have, if they're not home, their AC is turned off or they put it at some reasonable number.

15:33The duck is the morning. People wake up, do a load of laundry, and people come home. 6 p.m., look at that spike. And that informs the duck. So if you know you have a duck issue in California and Texas, but you have a different formation in Virginia, the same company could build five data centers and just load balance between them. And they just have to make the commitment to the grid that, hey, yeah, if you throttle us, we're going to accept the throttle and not sue you or be otherwise upset. Yeah? Exactly, Jason. So there are, of course, legal complexities here. But yes, that's exactly right.

16:13There is this bargain. So the utility, the regulator can offer this new deal of what we'll call non-firm service. Not very often, by the way, very rarely will you actually have to be interrupted. And then on the data center side or the AI industry, there's a new class of service, a new kind of SLA or service level agreement, where the compute customer says, you know, 95 % or more of the time, I'm going to get firm power to those GPUs. But every so often, I might need to be gracefully throttled. And so long as I label what it is that's mission critical and that doesn't get throttled, I can be sure that my workloads, either they'll be running at full speed here or somewhere else, as you mentioned, load balanced.

16:53I can be sure that they'll be running to my satisfaction. So that's the two differences. The electric power industry has to offer a new product. And the data center, AI industry, the cloud industry has to offer a slightly new SLA type. And you put those two together and suddenly you pack 100 gigawatts on the grid. And just to let you know magnitudes here, between now and 2028, America wants to plug in 70 gigawatts worth of AI data centers. We will successfully only plug in like 30. Put that in context that people can understand. They hear 30, they hear 70 gigawatts. What does that mean? Just if you were to come up with an analogy, enough power to small town, yada, yada.

17:31Yeah, some analogies, you know, that's a gigawatts about the same as a big nuclear power plant. So there's 100 nuclear power plants worth of data centers that I could put on the grid right now if they were flexible. And just so you know, China is going to have 400 gigawatts of spare capacity at peak in the year 2030 for AI, whereas the United States is already at a deficit unless we start to. And that's through their massive investment in nuclear and solar and batteries. Yeah, those are the big three or am I incorrect? And wind and coal, China has a massive, China last year brought online 500 gigawatts, again, 500 nuclear power plants worth of power compared to America's 50, right?

18:14So we will not win the build our way out of the AI race challenge in the near term, right? In the near term, we're not going to beat them on building speed. And so this is an intelligent way to unlock the infrastructure we already have. And it has a nice knock-on effect, Jason. And as these friendly, good citizen data centers come to town, they win over communities. They also end up bringing more revenue into the community because you're not triggering that ninth lane expansion. And so that additional revenue you bring in actually helps with your, Jason, your bill going down, right? It can help to subsidize your own bill.

18:48So that's a nice consequence where that additional revenue can subsidize your bill. It can also subsidize the building of infrastructure to serve a factory or a hospital or a neighborhood service. I mean, there's plenty of money in the mix here. The question is, is there enough energy and is there like too many blockers? And, you know, people have the ability to go to their local town halls and stop things. People forget just how powerful in America local government is. If local governments want to stop something, they'll be able to do it. They have the balance of power here. The federal government is purposely throttled in what it can do, no pun intended here, because just how we built these United States of America, not just America.

19:36It's the United States. All these decisions were made locally, and so you have to figure out a way to compromise with them. Take us to the other side of how jobs are going to be eventually categorized. I'm shocked that there isn't a system like this, but I guess we haven't needed to have a system if the states were begging for the next data center. And there was plenty of energy near some of these, you know, hydro, nuclear, gas facilities. There was more than enough locations to solve these problems. So here we are. We're in a different era. Well, Jason, the AI boom has taken the energy world by surprise, right?

20:20The energy world is kind of sleepy. I've been in it for two decades. I used to be the chief strategy and technology officer of a big Fortune 500 energy producer. And for the last two decades, energy demand has been flat, right? So we haven't been building in forecasts for growth. And suddenly, here comes AI, here also come some other trends like electrification, and suddenly, we're growing year on year. And so it took it by surprise. And now, as you correctly mentioned, we have a choice to make. Data centers are saying, wow, there's a 10-year wait ahead of me. Do I really want to connect to this grid?

20:55These regulations are really complicated. Sometimes utilities, as I mentioned, have to pick up the phone. Do I want to connect to a power grid or can I find another way of doing this? Maybe I'll just go off grid. I'll build my own little private utility. You see these deals happening, you know, in the Mountain West, the American West, et cetera. I'll just go build my private utility. I'll build generators right around me. I'll even go to space. So, you know, there are large public companies now that are valued based on putting data centers in space because that completely bypasses. SpaceX. Yeah.

21:24Right? Yeah. I bypass all the regulatory hurdles and the painfulness of connecting to the grid and the local opposition. There's no local opposition in space, right? And it's so intuitive, right? Especially as communities say, hey, don't raise my bill. An intuitive answer is, well, just go off the grid and you're not going to bug the community anymore. And I really don't like that future, Jason. And I'll tell you why. In that, first of all, I believe space-based AI is a wonderful idea and it should happen. Yeah. That's three to five years out at scale. And it should also happen alongside getting every last megawatt we can squeezed out of our existing grid.

22:03Yeah, because that's the right. And we're right now we're on path to not do that. We're on path to not squeeze out the grid and skip that step and go straight to space. And so the reason I don't like that future is because America's economic output is increasingly going to move to AI. Whether or not, Jason, you think that some of these goofy videos are valuable economic output or not, America's GDP will increasingly move toward AI. That's the revenue source that can help to support our public power system, the one that you and I depend on every day. So there are two futures. Future number one is the one where AI decouples from the grid and there's an archipelago of islands.

22:38And each of these islands doesn't support the public power grid anymore. Worse, they actually raise the price of the power grid because all the switchgear that they buy ends up raising the price of that same component for the public power grid. The better future is the one where we're very integrated. Data centers can connect to the grid. They can keep rates affordable. They actually pay for the grid expansion we need for all the other stuff we want, like car factories and TSMC advanced chip factories and neighborhood growth. And the AI data centers have the cheapest cost of token production because the American power grid is the best way to get reliable power.

23:14What percentage do you think of efficiency is there to wring out of the existing grid with software, with solutions like yours? How much extra can we get out of the grid ballpark? 10 %? 5 %? Yeah, great, great question. So today, the American grid is utilized at half or less, 50 % or less. So on average, you think about that highway, think about that highway being 50 % or less covered by cars over the course of a year. And I mentioned also that there's 100 gigawatts, which is more data centers than we'll need for, you know, through the end of the decade, that we can unlock today by driving that utilization up from 50 to 51 to 52 to 53 to 55 to 60%.

23:51So we can absolutely wring out a whole lot more utilization out of our grid. 10 % sounds like an easy target. Yeah. Look, I wouldn't call it easy. I certainly wouldn't call it easy, but it is absolutely achievable, Jason. A possible target. Okay, great. So easy might be two or three percentage points. 10 % might be a good stretch goal for you. Fantastic stretch goal. And I think the transformative flexibility of data centers arises because they're different from any load we've ever seen in history, right? They're big, they're controllable electronically. Like think of trying to spin down a car factory.

24:28You might damage equipment. But with AI data centers, I can throttle AI workloads from just using software. And they can move workloads at the speed of light from one part of the country to another. I can't charge my EV in Indiana one moment and then in California the next moment, right? So this is a transformative new load to help the grid work better instead of drain it. And we've seen small examples of this before. You can set your Tesla to charge at 1 a.m. when it's cheap. So there's a financial. So we didn't even get into today. You know, there might also be the opportunity if you can correctly say, hey, this job is not time dependent.

25:08I don't care what time of day it occurs. run it when the 50 % of the grid isn't in use, or when stuff's cheap, run it overnight when everybody's asleep and tucked in their beds. And while the cars are charged, whatever window that is. And then, yeah, when people are running their air conditioners and people are charging their EVs, we can run it slower, right? You could actually make those kind of trade-offs intelligently in the near future. Listen, Varun, go ahead. Yeah. To your point, Jason, these are small examples so far, right? Later this year, here's one example, a hundred megawatt AI data center that Emerald, NVIDIA, and Digital Realty are launching in Virginia.

25:50One single data center, that's going to be a massive load that is flexible. It'll be the world's first truly power flexible AI factory. And it's the reference design for the next thousand flexible AI factories. So whereas you're right, Jason, demand management has been a niche, relatively marginalized resource, Nest thermostats, for example. Now we're at the precipice of gigawatts, like a massive scale of flexible resources. It's going to transform our grid and it couldn't have happened without AI data centers. And we're really proud to work with NVIDIA and others on them. Amazing job. Where can people find more?

26:22I know you guys just raised a bucket load of cash and you're doing great. I'm sure you're hiring. So tell us where they can learn more and where the career page is in all of this. I think you're - No, we're grateful. You can find Emerald at emeraldai.co. We just raised a Series A. We have 12 Fortune 500 global investors, including NVIDIA. You just went to our careers page. Thanks a lot. And in addition, we just launched the AI Energy Management Alliance with Google and NVIDIA and Anthropic. So that's the organization advancing the grand bargain. And we're hiring for engineering roles, product roles, and several others.

27:01So we'd just be delighted to work with any of the folks listening to this podcast. And Jason, thanks for the shout out. Of course. I mean, trust me, I know how hard it is to hire. And hey, listen, even if you want to be corporate counsel or do policy work, that's a big part of this right now is we've somehow got to figure out how to get alignment with all these communities and be good actors. And I think a lot of this anti-data center stuff has more to do with maybe the fear-mongering around AI safety and job loss, which are completely manageable issues in the world and have very little to do with data centers, which, yeah, if done correctly, are an amazing asset for communities.

27:48And if one community doesn't want it, man, that's the other thing about 50 amazing states. You can pick which states. And you have Pennsylvania, I guess, fighting this now? Or what's the game on the field with Pennsylvania and Texas? We hear about those two. Are they the most anti now, or is it New York? And what's the sentiment, and what's your take on it as we wrap here? Is this performative by the politicians because they know it gets them votes, and it's a popular way to get people to vote for you? Or do they actually have valid concerns for their communities? Jason, I think it's the absolute wrong response for any of us in the AI industry to look at the apprehensions that folks face about AI data centers and dismiss them out of hand, say they're performative, say they're playing to the politicians, et cetera.

28:33I think communities are apprehensive, they're right to have anxieties, and it's our job to prove that AI data centers can be good grid citizens. You asked about some states. The question yesterday was, or last year was, which states might pass a moratorium on data centers? The question today is, which states won't pass a moratorium on data centers? Answer it, which ones won't? Which ones are taking a really thoughtful approach to this and appreciate the business? Is Texas still one of them? Well, again, it's red and blue states that are considering pauses. Even Texas now has a pause on its what's called batch zero process to connect data centers to the grid.

29:08Governor Abbott has a pause on it, likely through the midterm season. So again, it is incumbent on us to prove that AI data centers can be good for communities. and I have spent a lot of time, especially with this new alliance launch with Google and NVIDIA to show a different side of data centers. Data centers bending over backwards, being flexible just to keep your bill affordable and so that you'll welcome them. If we can prove this, then I want to move to a regime where states compete with each other, where governors say, I'd love to have these friendly data centers in my state. It used to be.

29:40Like it was five years ago and pretty straightforward. Don't screw up the grid and there's so many ways to do that. And yeah, maybe build a couple of parks and stadiums and maybe lower people's electrical bills. Super easy to do. Varun, thank you for your efforts here and continued success with EmeraldAI.co. EmeraldAI.co. We'll see you soon. All right, everybody. Next up on the program, we've got a really interesting AI startup that's designed to help hedge fund managers and asset managers. Today on the program, Vibha Viswaathan. He is the co-founder and CEO of Pascal AI. I think that's one of the first programming languages I learned, Vassal.

30:22Tell us a little bit about what you're building, who it's for, and maybe the stage of the company, where you're at with the company. Products launched, products in development, yeah. So we are building AI agents for institutional finance. So that includes hedge funds and asset managers primarily. We are early stage. So we have launched the product. We've raised a first round of institutional finance. It's in the hands of now about over a dozen customers across the world, many of them in full production. So if you look at the accumulated AUM of all of our customers, it's about half a billion dollars right now.

31:00So very large enterprises, lots of hedge funds and asset managers who are using the product on a daily basis. Okay. And what's their goal with the product? What problem are you solving for them? Because hedge fund managers have been using machine learning and AI for decades. They are the people who have been most attracted to this. People like Jane Street, I understand, are putting out$20 billion worth of data centers, in fact. So these are the smartest people in the world with computer scientists on board, with AI PhDs already working for them. So what product are you building for them that maybe they didn't already have in their companies?

31:44Yeah. So a lot of the work that these managers have done has been pre-LLM. So a lot of the development is on the models and neural networks and so on that existed before generative AI came about. So what we do is actually build out the context graph for these enterprises. So think of these portfolio managers, analysts, and so on, who have taken a lot of different decisions over the years. They've got that stored in different parts of their enterprise data stack. We've stitched all of those together. So if you wanted to know, look, how did I think about this particular name in 1999? right and how does that compare to how i think about the stock today a lot of that requires data plumbing and contextual understanding across a siloed set of data and that's the problem that we're solving okay vibhob show us uh what you've built here at pascal what we do is like i said we build ai agents for institutional finance so this is context aware infrastructure uh behind these workflows.

32:52So we take decision traces of various names in your portfolio. The system is able to parse your IC memos, various different reports that are written. You have a living thesis. So if you own, for instance, 15 or 20 different stocks in your sleeve of the fund that you're managing. It looks through every single investment thesis and vectors that really matter to you. And as new information flows in, it's able to sift through volumes of information contextual to your thesis and tell you if it's positively impacting your company, negatively impacting your company, is it neutral? So you have a view every single minute of what's really happening to your portfolio.

33:42And this is not noise. This is actual signals that you're capturing from across various different sources. The same agent lives across different work surfaces, including Excel. So if you had an Excel model and you needed the latest information to be updated, it'll update it. It'll create your valuation charts for you. You can create various different skills on top of it. So this is almost like an always-on digital teammate for you. So as and when new information comes in across your work surfaces, it's able to real-time make changes, right? The most important part of this that our customers really appreciate is the fact that once you have sort of walked the model through what your thinking is and how you really look at excels and stocks and, you know, whatever work surface you're on, it starts to remember and capture the institutional memory, right?

34:41So maybe you have 20 different sources, which is part of your institutional memory. So all of those sources are combined together. And this can be in any format, right? PDF, excels, it really doesn't matter. So you could take all of that in, you could ask it a question the way you'd ask, you know, maybe your analyst or somebody on the cell site or anybody or an expert. So it literally is able to sift through your thesis, your information, get you an answer. Now, all the answers that we have are source cited. So you can take any section of your answer, literally look at the ground truth and trace back to where the information comes from.

35:23So you've got accuracy, you've got completeness of the answer. So what is the question being asked here? And how is this different than, say, me just using ChatGPT, Claude, et cetera, because I do this kind of stuff myself looking at companies. So that's always the question investors have for founders. So what's the question being asked here? How is this different than me asking the latest Claude and Frontier model? And how do you compete against those? Yeah, I'm glad you asked that. It's a question that our customers ask, everybody asks us. So the question here is, can you summarize my thesis on Honeywell compared with management guidance and the company's performance since then, suggests recalibrations to my thesis based on the management's latest forward guidance and highlight major changes to my thesis based on Honeywell's latest development.

36:12The important thing to note here is you're not uploading any document. There is no additional information given to the system barring the question that you're asking. So if you were to ask somebody at your firm the same question, the individual who is looking at this would have to know where the thesis is. Make sure the thesis is the latest thesis on Honeywell, has to compare it with management's guidance. Since the thesis was written and the thesis might have been written last week, it may have been written two weeks ago. You may have eight different versions of the thesis. And you need to really look at the guidance and company's performance since that particular time.

36:52So if you've got something called a temporal awareness, meaning we understand timelines really, really well. And then we also need to look at, you know, what's the forward looking guidance. So not just you have the performance, but you have to look at quarterly performance since the thesis is written across perhaps your the company filings, but also on what the brokers have said, what you've seen on X, what you might have seen on podcasts, you know, various different sources as a hedge fund that you might track. So we stitch all of that together and give you information that meaningfully confirms or disconfirms your thesis, right?

37:32It's not like we're giving a data dump, but it really understands your thesis and basis that gives you an answer. Now, to your question on how is this different from what the frontier models do, so we use a lot of the frontier models and it's super, super exciting. The easiest way that I explain this is if you have a super intelligent human being, let's say Albert Einstein, and you ask Albert Einstein to work at Meta or you ask Albert Einstein to work at Amazon or ask Albert Einstein to work at Alphabet, Einstein will have to work differently at each of those places because they have different ways of working.

38:13and different ways of working in the institutional finance world is what is your time horizon for investing? What data sources are you looking at? Who do you talk to? How do you invest? How do you trade in and out of stocks? So there are funds that do quarterly and their big edge is to try and figure out what the next quarter's earnings are. And then there are asset managers that want to invest for the long term for maybe the next five years, right? So you've got this super intelligent agent that Claude and GPT and Grok and the open source folks have, but you need to really fine tune that with your institutional context and memory, and that's where we come in.

38:54So our technology is, we take all of the data sources you have, we extract out the key entities, build relationships between these entities, and make sure that when you ask a question, We give you a complete and accurate answer. Got it. And this is based upon the data from which sources, because a lot of that data is proprietary. And so you have the ability to look into their Bloomberg tournament or watch their desktop. How are you getting information? Because if people are making trades based on podcasts, that's a little concerning for hedge funds, or maybe not. I don't know. Maybe there's interesting things being talked about on some podcasts.

39:33But I don't understand how this is much different in terms of the data sources. So maybe that's where the difference really comes in is, hey, there's some proprietary data or the hedge fund has its own data set that's helping it make answers that that data is not visible to Claude or to OpenAI's models. Exactly. Exactly. You hit the nail on the head. So we classify the data sources into three parts. So there is the proprietary enterprise data that the hedge funds have that they will not share with anybody else. Then there are these licensed market feeds. So this could be like, you know, all the licensed feeds that one might consume.

40:12This could be a CapIQ feed. It could be a fact set feed. It might be a Bloomberg and so on. So you pay dollars to get this data. And then there's everything in the, you know, wild web, right? So this could be, you know, all the different things that are being spoken about on X, on podcasts, et cetera. So what we have access to that the frontier models don't is everything that's proprietary. So we have a top 10 hedge fund in the world that shared every single part of their internal thinking with our agents. So this would be their IC notes, it would be their portfolio data, it would be all of the management meetings that they've had, it could be all the internal morning meeting discussions that they've had, and so on and so forth.

40:56So that really informs the agent on what the internal context and thinking is and basis that we're able to pull out signals from the license feeds and the world, you know, the internet. And you have customers using this. Who are some of the customers and why are they using this as opposed to building it? That's always, you know, in today's era, it's very easy to build software. These big hedge funds have software developers. So tell me who's using it. Why aren't they building it themselves and trying to get a proprietary edge is always what we try to determine buy or build, right? That's a classic decision.

41:35So who's the customers of this? You have specific names who are using it? Yeah, these are secretive hedge funds, so I'm not allowed to give the names, but they're like one of our marquee customers, like a$40 billion Tiger Cup. So they've got, if they want it, they've got the best of data sources. They've got a massive budget. Now, the question even for them is, is a build or buy? I think that was the way people phrased it pre-LLMs. I think post LLMs, what we are seeing is it's build and buy, right? So what they're doing is they're using Pascal for the graph. So we get all the institutional context in, and then we build a certain set of agents and they build other sets of agents.

42:21So we've actually worked very closely with our internal tech team, where we give them the entire infrastructure plumbing and their internal tech team can build specific applications on top, or they could use our application library to build or use some of the agents that we already have pre-built. When are humans going to be antiquated here? Like, it seems to me like there's a lot of people who have these jobs who are, you know, basically, if the software is doing what you're saying, it feels like they're going to be disrupted because they're getting information late. People are making trades much earlier.

42:59And this information is being commodified super quick. So what's the future of these hedge fund managers and like their futures and how are they going to work? Because this is, you know, bolting some AI onto their, you know, existing systems to try to help them, you know, shore up their thesis. I get it. Makes sense. But what are their jobs going to look like in three or four years? Are they going to even have jobs or is it going to be better to just have an AI tell you these are the best stocks to buy? you need to get out of this stock, you need to be in this stock? I'll tell you what we've seen practically.

43:35What we see with everything in Pascal is everything that's digital, right? So you've got stuff that is digitally captured and AI with the exponential, the way it is progressing. A lot of that is probably being already captured by the AI models. But what we're seeing practically is these folks are traveling a lot more, they're doing more channel checks, they're out in the field and in the ground gathering information, right? So they're getting a lot more of the primary work done. And that's where the alpha is shifted into. So we had a customer come and tell us that their travel bills has gone up to X, right?

44:13So that sort of is indicative - So they can get their paperwork done quicker, which means they get more time to go out and be an investigative journalist, investigative reporter to try to get an edge. Makes total sense. Vibhov, thanks so much for coming on the program. Continued Success, where can people find out more about the company? They can log on to www.pascalealabs.com. We've got a bunch of information there. And then they can schedule a demo, or they could just write directly to me, vibav at pascalalabs.com. Continued Success, you've got a very deep-pocketed customer base, and you're obviously helping them solve some real problems here.

44:53So congratulations. It's a pretty good start. and you're in a very interesting space that's ripe for disruption, which means opportunity. So keep that peripheral vision open and we'll see you next time. Bye-bye.

From the publisher

Thanks to our partner, @Paypal! Pay zero processing fees on your first $100K in eligible PayPal payment volume. Learn more at https://paypal.launch.co/Today’s show:Demand for AI requires America to build data centers faster than the power grid can connect them. Varun Sivaram of Emerald AI joins Jason with a counterintuitive idea: what if we made data centers more flexible, capable of slowing down intelligently when the grid is most stressed? (That only happens a few hundred hours each year.)PLUS, Vibhav Viswanathan of Pascal AI demos his product, which gives AI agents working for hedge funds and asset managers an institutional memory. Why can't Claude or ChatGPT do the same job? And what happens to analysts when bots handle the back office admin?Guests:Varun Sivaram on X: https://x.com/vsiv Emerald AI: https://www.emeraldai.co/Vibhav Viswanathan on LinkedIn: https://www.linkedin.com/in/vibhav-viswanathan-13550015/ Pascal AI: https://www.pascalailabs.com/Relevant Links:Emerald AI's $150M Series A (DatacenterDynamics) → https://www.datacenterdynamics.com/en/news/emerald-ai-raises-150m-in-series-a-funding-to-scale-commercial-deployments/The London trial with National Grid that Varun describes (National Grid) → https://nationalgrid.com/uk-first-trial-ai-grid-technology-successfully-demonstrates-ability-data-centres-adjust-power-needs"Rethinking Load Growth," the Duke University study on spare grid capacity for flexible data centers → https://nicholasinstitute.duke.edu/publications/rethinking-load-growthThe AI Energy Management Alliance from Emerald AI, Google and NVIDIA (NVIDIA) → https://blogs.nvidia.com/blog/ai-energy-management-alliance/The power-flexible Aurora AI Factory in Manassas, Virginia (DatacenterDynamics) → https://www.datacenterdynamics.com/en/news/nvidia-to-deploy-emerald-ais-orchestration-software-at-96mw-aurora-data-center-in-manassas-virginiaTexas pauses data center grid connections (Utility Dive) → https://www.utilitydive.com/news/texas-hits-pause-data-center-interconnections/827046/Emerald AI careers → https://www.emeraldai.co/careersPascal AI's $3.1M seed round (Unite AI) → https://www.unite.ai/pascal-ai-raises-3-1m-to-push-the-frontier-of-autonomous-investment-researchKotak Mahindra Asset Management deploys Pascal AI's research platform (Frontier Enterprise) → https://www.frontier-enterprise.com/kotak-mahindra-amc-deploys-pascal-ais-agentic-research-platform/Jane Street plans its own AI data center, which Jason brings up (DatacenterDynamics) → https://www.datacenterdynamics.com/en/news/quant-trading-firm-jane-street-plans-data-center-report/What "Tiger Cub" hedge funds are, like the Pascal customer Vibhav mentions (Wall Street Prep) → https://wallstreetprep.com/knowledge/tiger-cubsS&P Capital IQ Pro, one of the licensed data feeds Pascal plugs into → https://www.spglobal.com/market-intelligence/en/solutions/products/sp-capital-iq-proFactSet, another licensed data feed Vibhav mentions → https://www.factset.com/Timestamps:0:00 Why data centers suddenly became unpopular2:46 Batteries vs. throttling GPUs vs. moving workloads4:11 Why utilities still pick up the phone4:52 The London teakettle test at football halftime5:56 Throttling AI workloads without users noticing10:40 How much spare room the grid really has13:32 The grand bargain, a fast pass onto the grid17:23 Why America can't out-build China on power20:39 Why off-grid and space data centers worry Varun27:37 Are state data center moratoriums performative?29:45 Pascal AI, AI agents for hedge funds and asset managers35:12 How Pascal differs from ChatGPT and Claude38:54 The proprietary data frontier models can't see42:20 Will AI replace hedge fund analysts?Subscribe to our newsletters on Substack:TWiST: https://twistartups.substack.com/TWiAI: https://www.thisweekinai.ai/TWiVC: https://thisweekinvc.substack.com/Check out all our partner offers: https://partners.launch.co/

More from This Week in AI

All episodes