Vikhyat Chaudhry, CTO, COO & Co-Founder of Buzz Solutions, leads the company’s technology development and operations as it applies artificial intelligence, machine vision, and predictive analytics to the inspection and maintenance of critical energy infrastructure. He co-founded Buzz Solutions in 2017 after developing experience across data science, machine learning, embedded systems, computer vision, and autonomous technologies. Before focusing full-time on the company, Chaudhry worked as a data scientist at Cisco, developed machine learning and sensor-based systems at Altitude Co., and conducted Stanford University research involving autonomous drones, wind-velocity sensing, and computer vision for wind farms. His earlier research at India’s Defence Research and Development Organisation explored landslide forecasting using simulations and remote-sensing software, providing him with a technical background spanning infrastructure risk, geospatial analysis, and applied AI. Buzz Solutions currently identifies him as its co-founder, CTO, and COO.
Buzz Solutions develops visual intelligence software designed to help energy companies inspect, maintain, and secure transmission, distribution, substation, and generation infrastructure. Its PowerAI platform analyzes imagery collected through drones, helicopters, fixed-wing aircraft, and ground inspections, converting visual data into diagnostics that help utilities identify defects, assess asset condition, prioritize repairs, and integrate findings into existing geographic information system, asset-management, and work-order workflows. The company positions its technology as a way to detect infrastructure risks earlier, reduce the manual work required to analyze inspection imagery, and lower the likelihood of equipment failures contributing to outages, shutdowns, or wildfires. Buzz Solutions also offers tools for substation monitoring and security, reflecting its broader goal of helping utilities build safer, more resilient, and more data-driven grid operations.
When we interviewed you in 2024, our conversation focused on using AI to move utilities from manual inspections toward predictive maintenance. What has changed most over the past two years, both across the energy sector and in your own thinking, that led you to write Powering Intelligence and frame electricity as a potential constraint on the future of AI?
When we last spoke, the conversation was mostly about the grid’s back office, moving utilities off manual inspection and toward predictive maintenance. That’s still the work. But something shifted underneath all of us in the two years since, and it shifted faster than almost anyone in my world expected. Back then, AI was a demand story we talked about in the abstract. Now it shows up as large load growth numbers for utilities. I started hearing the same thing from people I’ve known for years in this industry: interconnection requests for loads that would have been unthinkable a few years ago, single facilities asking for the kind of power a small city draws and asking for it on a timeline that has nothing to do with how the grid actually gets built. That was the moment it clicked for me. We spend all our energy debating how capable these models will get, and almost none on the boring physical question of whether we can feed them. I wrote Powering Intelligence because I kept looking for the book that connected those two worlds and it didn’t exist. Everyone was writing about the AI models. Almost no one was writing about the electrons needed for them. And I was sitting in a strange spot, close enough to the AI side to understand the ambition, and close enough to the grid to know exactly how long a new transmission line and a power generation plant development really takes. The gap between those two clocks is the whole story, and I wanted to put it in one place.
You argue that the AI industry scales in months while the power grid changes over decades. Where is the most immediate bottleneck today: electricity generation, transmission capacity, substations, transformers, interconnection queues, permitting, or the availability of skilled workers?
I think there are multiple bottlenecks here, and the answer is that they’re all tangled together. But if you make me point at the tightest knot right now, it’s relatively less about generating the electrons and more about delivering them: transmission and the interconnection process, with transformers and other electrical equipment as the physical pieces that ties up everything. Here’s the thing people often underestimate. You can announce a power plant and a data center in a press release. You cannot press-release a large power transformer into existence, some of those have lead times measured in years, and a lot of them aren’t even built domestically. So even when the will and the capital are there, you hit a supply-chain wall on components. Then, layered on top, is the interconnection queue, which in a lot of regions is a multi-year line just to get studied. And underneath all of it is the workforce problem for the power industry. The people who actually know how to build and maintain this equipment are retiring faster than we’re replacing them.
So, it isn’t one bottleneck. It’s that the slowest-moving physical pieces, the transformers, the transmission corridors, the trained lineworkers, are exactly the ones you can’t accelerate with software or money alone. That’s what makes this hard.
AI data centers can introduce enormous, highly concentrated loads into areas whose grids were not designed for that level of demand. How are these facilities different from the large industrial loads utilities have managed historically, and what new risks do they create for reliability and capacity planning?
Utilities have handled big loads for a century. A steel factory, an aluminum smelter, automobile manufacturing plants, these are enormous loads, and grid planners know how to think about them. So, the instinct is to treat a data center like just another big customer. I think that instinct is wrong, and the differences are what creates the new risk.
Three things stand out to me. First is the concentration and speed. A traditional heavy-industrial load usually grew alongside the grid in that region over decades. These facilities want to land huge, all at once, in places whose local grid was never planned around them. Second is the load profile. A lot of these are not smooth loads, they can ramp hard and fast in ways that stress equipment differently than a factory running a steady process. Third, and this is the one that worries planners most, is the speculative pile-up, which I know is really your next question.
The risk it creates is that utilities are being asked to plan around demand that behaves differently than anything in their historical models, and to commit capital and infrastructure on the AI industry’s timeline rather than the grid’s. When your planning assumptions are built on a hundred years of one kind of load and the new load breaks those assumptions, reliability planning gets genuinely harder. That’s not a reason to say no. It’s a reason to plan differently.
Utilities are receiving ambitious power requests from data-center developers, but not every proposed facility will necessarily be completed. How can utilities distinguish committed demand from speculative requests and avoid building costly infrastructure for projects that arrive late or never materialize?
This is one of the most important operational questions in the whole space right now, and I’m glad you asked it directly, because it doesn’t get enough attention next to the flashier topics.
A developer can submit interconnection requests for the same project in multiple territories at once, shopping for whoever can deliver power fastest. Some of those projects are real and funded. Some are options, placeholders, and a way to hold a spot in line. Some of those are being called “Phantom loads” which might never come through. If a utility treats every request as real and starts building generation and transmission for all of it, they risk stranding enormous cost, infrastructure built for a data center that shows up years late or never shows up at all. And guess who ends up paying for stranded assets, the existing ratepayers. On the other hand, if the utilities build less, then they risk not being able to fulfill the demands of the incoming load which can lead to reliability issues on the grid.
The way through is to make speculation expensive and commitment cheap. That means real financial skin in the game before a utility commits capital, meaningful deposits, phased commitments tied to construction milestones, contract terms that put the risk on the party making the request rather than on the ratepayer base. Some jurisdictions are starting to move this way. The principle I keep coming back to is simple: the entity asking for the power should carry the risk of the power not being needed. Right now, too often, that risk quietly lands on everyone else.
Who should ultimately pay for the generation, transmission, and substation upgrades required by AI data centers? What rate structures or contractual safeguards could prevent households and existing businesses from subsidizing infrastructure built primarily for hyperscale technology companies?
This is the question that’s going to define the politics of the next decade of the grid, and I don’t think it has a purely technical answer. I think it’s about fairness. My honest view is that the cost should sit, as much as possible, with the load that causes it. If a hyperscaler’s facility requires a new substation and a transmission upgrade, the default assumption should be that the facility pays for what it triggers, not the retiree down the road whose bill quietly climbs to cover it. That sounds obvious, but the way rate structures have historically worked, big new infrastructure often gets socialized across the whole customer base, which made sense when the new load was a factory that employed the town. It makes a lot less sense when the beneficiary is one of the most valuable companies on earth.
The tools do exist. Large-load tariffs designed specifically for this class of customer. Minimum-take contracts so a data center can’t get infrastructure built and then walk. Long-term commitments that match the life of the assets being built. What matters is that regulators and utilities actually use them, and use them before the build, not after. Because once the concrete is poured and the cost is in the rate base, it’s very hard to claw back. I write about this in the book because I think it’s one of the places where getting the incentives slightly wrong could quietly erode public trust in both the utilities and the AI industry.
Could AI data centers become flexible grid assets rather than inflexible sources of demand through workload shifting, demand response, batteries, thermal storage, or on-site generation? Which AI workloads can realistically be moved across different hours or locations without compromising performance?
Yes, and I think this is one of the more hopeful ideas in this topic, so I want to be careful not to oversell it, because the flexibility is real but it’s not unlimited. Start with the distinction that matters: not all AI compute is the same. There’s inference, which is the compute after these large models are trained (the live models), and you asking a model a question and needing an answer now. That’s latency-sensitive and hard to move around. And then there’s training, the massive, batch-style workloads that build the models. Training is far more flexible. It doesn’t necessarily care whether it runs at 2pm or 2am, or in this state or that one, as long as it gets done. That flexibility is the opening. If a big chunk of a data center’s load is the kind that can shift in time, then it can lean into the hours when the grid has surplus clean power and back off when the system is strained.
Now layer on batteries, thermal storage, and on-site generation, and a facility that looks like a rigid block of demand can start to behave like a resource that actually helps balance the grid, soaking up excess solar in the middle of the day, riding through peaks on stored energy. The honest caveat is that this requires the operator to actually want to cooperate, and it requires rate structures that reward them for it. The technology is the easy part but the incentive design is the hard part. But a world where data centers are flexible partners to the grid instead of adversaries of it is genuinely possible, and it’s a lot better than the alternative.
What combination of energy sources can realistically meet AI’s electricity requirements over the next decade? How do you see renewables and storage, natural gas, conventional nuclear power, small modular reactors, and geothermal energy fitting into that mix?
I’ll give you the answer people don’t love, which is that it’s all of it, and anyone selling you a single-source silver bullet is selling something. Renewables and storage are the fastest thing we can actually deploy; solar plus batteries can be built on timelines that look almost sane compared to everything else, and that speed matters enormously when demand is rising this fast. So that’s the near-term workhorse. In fact, a lot of Buzz’s utility customers (Dominion, Ameren, AEP, NYPA, Southern and others) are continuously increasing their solar footprint in order to meet the demands of the data center loads. But renewables are variable, and these loads run around the clock, which is why storage is doing so much of the heavy lifting in that sentence, and why firm power still matters. Natural gas is going to be part of the bridge whether people like it or not, simply because it’s dispatchable and it’s here. Nuclear is fascinating right now, we’re watching plants that were headed for retirement get a second life specifically because a data center will sign a long-term contract for their output, which almost nobody predicted. Small modular reactors (SMRs) and geothermal are the genuinely exciting longer-plays, but I try to be disciplined about timelines, they’re a meaningful part of the story by the back half of the decade and beyond, not the thing that saves you in 2027.
So, the realistic mix is a layered one: renewables and storage carrying the growth, gas bridging the firmness gap, existing nuclear extended and newly valued, and the next-generation sources coming online as they mature. The mistake is ideological purity in either direction as the grid doesn’t care about our preferences. It cares about kilowatt-hours delivered reliably and on time.
One of the central ideas in your book is that the same AI industry placing pressure on the grid could also provide the tools needed to modernize it. Where can AI create the greatest near-term gains: load forecasting, predictive maintenance, transmission optimization, vegetation management, outage restoration, digital twins, or autonomous grid controls?
This is the heart of the book, and honestly the part I find most hopeful, the same industry straining the grid is handing us the best tools we’ve ever had to fix it. If I’m ranking by where the payoff is real today rather than in a slide deck, I’d put it here.
Predictive maintenance, continuous monitoring and inspection first, because that’s where Buzz has spent years, and I’ve watched it work. Using computer vision and visual AI to find the defective components before they fail, the cracked insulator, the overheated transformer, the rusted hooks, that’s not speculative, that’s deployed and saving real money and preventing real outages right now. Right alongside it I’d put load forecasting and demand response, because the entire data center problem we’ve been discussing gets more manageable if you can actually predict demand with more precision instead of planning blind. Then transmission optimization, squeezing more capacity out of the wires we already have, which matters enormously given how slow it is to build new ones. Vegetation management and outage restoration are close behind as both are areas where AI takes something slow and manual and makes it fast and efficient but at the same time providing enough insights and situational awareness to provide resiliency to the grid in order to prevent incidents like wildfires, storm damages and power outages.
Digital twins and fully autonomous grid controls are the further horizon, they’re coming, and they’re powerful, but they carry the trust and safety questions that I suspect is your next question. So near-term, my honest ranking leads with the unglamorous stuff: see the grid better, predict its behavior better, and get more out of what’s already in the ground by also making the grid system more aware. That’s where AI pays for itself first.
Our previous interview discussed Buzz Solutions’ reliance on real utility data and human-in-the-loop review. As power systems become more autonomous, which decisions could safely be delegated to AI, and where must human judgment remain mandatory? How should utilities prepare for opaque model decisions, cyberattacks, or failures that could propagate across critical infrastructure?
When we last talked, I described our own approach as human-in-the-loop, subject-matter experts giving feedback to the AI models on anything below a confidence threshold or bad predictions, and I want to start there because my view on this hasn’t softened, it’s hardened.
AI is extraordinary at the tasks of perception and recommendations, look at all this data, find the anomaly, flag the risk, rank what needs attention. Delegating that is not only safe, but also better than what humans can do alone at that scale. Where I get much more cautious is consequential action on critical infrastructure and mission critical projects, the decisions where being wrong takes down power for a hospital or cascades across a region. Those need a human (subject expert) accountable, not because the model is necessarily worse in the average case, but because critical infrastructure is exactly the place where the rare catastrophic case is the whole ballgame.
So, in my opinion, the preparation is threefold. First, keep humans meaningfully in the loop on consequential decisions, not as a rubber stamp, but with real authority to override. Second, insist on explainability for anything touching critical operations, if you can’t understand why the model did what it did, you can’t be accountable for it, and “the AI decided” is not an acceptable answer when the lights go out. Third, and this is the one I think the industry is dangerously behind on, treat these autonomous systems as an expanded attack surface from day one. The more we let software run the grid, the more the grid inherits software vulnerabilities. A cyberattack on an autonomous control system isn’t a data breach, it’s a physical event. We have to design for that reality before we become full autonomous, not after. The positive news is that utilities are now taking data security, infosec and AI governance very seriously and have also started setting up AI and data governance councils within their organizations. We, at Buzz, have published cybersecurity and AI governance playbooks for the industry to learn more about the successful deployments of AI systems securely and at scale.
Looking five years ahead, what will separate the countries and regions capable of supporting continued AI expansion from those constrained by electricity availability? What would a successful strategy for aligning AI development with grid modernization look like by 2031?
I think five years from now we’ll look back and realize that the constraint on AI was never the chips or the talent. Those move around the globe easily. The thing that couldn’t move, the thing that actually decided who won, was power and infrastructure. Electricity becomes the strategic resource of the AI era the way oil was for the last century, and I don’t think that’s over-exaggeration; I think it’s just becoming visible.
So, the regions that pull ahead will be the ones that treated grid capacity as a competitive asset and started building and reforming early, the places that streamlined interconnection, that modernized permitting without gutting the safeguards that matter, that invested in transmission before the demand fully arrived rather than scrambling after. The ones that fall behind won’t lack ambition or capital. They’ll have the data center announcements and the investment dollars, and they simply won’t be able to deliver the electrons, and the projects will quietly go somewhere that can.
A successful strategy by 2031 looks like the two clocks (referenced in the book) I keep coming back to finally being set to work together instead of against each other. It means aligning the speed of AI ambition with the reality of energy infrastructure timelines, using AI itself to get more out of the grid we have, while we build the grid we need, getting the cost allocation right so that the public sentiment around power affordability doesn’t become extremely negative, and taking the workforce problem seriously enough to actually train the next generation of people who build and run this system. None of that is a technology problem as we have the technology. It’s a coordination and a will problem. In my opinion, the good news is that those are the problems we know how to solve; it would just require the technology industry, the energy industry and the policy makers and regulators to sit down together and have the relevant conversation.
Thank you for the great interview, readers who wish to learn more should visit Buzz Solutions.
