Andrii Garanin, Chief Energy and Infrastructure Officer at Silicon Foundation – Interview Series

Andrii Garanin is Chief Energy and Infrastructure Officer at Silicon Foundation, where he develops AI data center and battery storage projects across major U.S. power markets, including PJM and ERCOT. With more than 20 years of experience across the energy value chain, including power generation, grid integration, and storage, he is one of the few infrastructure leaders bridging the worlds of power generation and AI compute. He has developed and operated energy assets spanning renewables, gas, and battery storage and has held leadership roles at international energy and technology companies, including Scatec Solar and Elementum Energy.

Silicon Foundation is an infrastructure-focused organization building at the intersection of energy, AI computing, and capital markets. The foundation’s core thesis is that future AI infrastructure will rely on smaller, distributed data centers directly connected to energy assets rather than massive centralized facilities. To support this vision, Silicon Foundation develops and operates GPU data centers for AI workloads, invests in large-scale battery energy storage systems (BESS) to improve grid reliability, and is creating RAW Protocol, a market infrastructure layer designed to connect physical energy and compute assets with institutional capital markets. The organization is currently developing a large-scale compute and energy campus in West Virginia that combines grid access, energy storage, cooling infrastructure, and AI compute capacity as part of its broader goal of building an energy-abundant future.

You have spent more than two decades building energy infrastructure across renewables, methane-to-energy, utility-scale solar, battery storage, and now AI data centers. How has that journey shaped your view that power availability is becoming just as important to AI development as GPUs and foundation models?

I spent most of my career on the power generation side, but it was working on the consumption side that changed how I perceive the whole system. You quickly realize that matching demand to supply is not a secondary concern. It is the concern. The real constraint isn’t generation in the abstract; it’s the grid itself, which has finite capacity to move and balance power. To accommodate the rise of renewables and the congestion it causes, you must treat consumption as controllable and flexible, not fixed.

AI sits right in the middle of that. It is the most consequential technology shift of this century, and it is already a major and fast-growing source of electricity demand. The IEA projects that electricity use by data centers will roughly double by 2030, with AI-optimized facilities more than quadrupling; in the US, data centers alone are expected to account for close to half of all electricity demand growth this decade.

Power availability is now a first-order input to AI development, on the same level as compute. The work ahead is finding intelligent ways to make that power available, and that’s a power-systems problem as much as a technology one.

At Silicon Foundation, you are developing AI data center and battery storage projects across major U.S. power markets such as PJM and ERCOT. What are the biggest misconceptions technology leaders have about what it actually takes to secure power for AI infrastructure at scale?

A few stand out. The first is the belief that you can drop a gigawatt-scale data center almost anywhere, as long as the site looks promising on paper. In reality, you have to account not just for the infrastructure that exists today but also for what’s coming: the other loads and generation that will connect around you and reshape the local grid.

The numbers make this concrete: ERCOT’s large-load interconnection queue jumped to over 200 GW in 2025, up from around 63 GW at the end of 2024, with the large majority of it data centers. You are not sitting in a static system; you are sitting in a queue that is being completely reshaped.

The second misconception is treating it as a real estate play. It isn’t. A serious developer needs genuine expertise across the entire value chain, power generation, transmission and grid integration, and compute management, not just one slice of it. The third is having too narrow a field of view. You can’t just look at regionally available power. You have to track national policy, work with the communities you’re entering, and partner closely with the local grid operator, because they are the ones who ultimately decide whether and how you connect.

Many investors still evaluate AI companies primarily through software, chips, and model performance. Why do you believe energy infrastructure may become one of the most important bottlenecks and opportunities of the AI economy?

Once consumption reaches this scale, you have to look beyond chips and software to the infrastructure that makes compute possible, and that upstream system is far more complex than many investors assume. The compute layer cannot be evaluated in isolation from what’s happening on the power supply side; they’re the same system.

Grid access has become the binding constraint on data center construction, ahead of land or permitting, and connection timelines now stretch for years. That’s the bottleneck, and it’s also exactly where the opportunity is for anyone who can actually solve it.

There’s a second factor people underestimate: the load profile of an AI data center is not flat. Training and inference workloads can swing power draw sharply, sometimes within seconds. That intermittency has serious implications for how a facility operates and, more fundamentally, for how it has to be designed and integrated in the first place.

A site that can’t manage those swings or turn them into a feature the grid can use is going to struggle. The investors who understand that the energy layer is where value and risk concentrate will see the opportunities the chip-and-model lens misses.

We often hear about data center operators seeking grid connections, but you have argued that grid access alone is no longer enough. What additional capabilities will successful AI infrastructure operators need over the next decade?

As I said, a grid connection on its own does not guarantee a successful project, because the grid is never static. New generation and new load are connecting all the time, and both change what your connection is actually worth. The operator and developer have to understand those implications: whether the local grid can absorb rising demand and what limitations the operator may impose down the line once they recognize they can’t keep the system stable.

These constraints are real and increasingly common. PJM, which serves 67 million people, recently came up short of its reliability target in a capacity auction for the first time, with prices hitting the regulatory cap.

Renewables are the cautionary example here. Many projects were suspended or curtailed after they had already been built. Regulators may manage those events at the system level, but the financial burden usually falls on asset operators, directly weakening the investment case. Worse, it can reduce investor appetite for the next wave of projects.

So the capability that matters over the next decade isn’t just getting connected; it’s the ability to anticipate and absorb that volatility through flexibility, storage, and the right contractual and operational design so your project stays bankable even as the grid around it changes.

Battery Energy Storage Systems (BESS) are increasingly being discussed alongside AI infrastructure. How do you see battery storage evolving from a backup solution into a strategic asset for AI data centers?

I see this as one of the major shifts of the past decade, and it comes down to two factors. The first is cost. Battery costs have fallen dramatically, by roughly 90%, while installed utility-scale costs have been declining at around 20% per year. That trend is still continuing, and it is changing what is economically possible.

The second is the defining technical feature of BESS: it can react almost instantly to changing conditions on the grid. That speed matters simultaneously for generation, for AI data center consumption, and for overall grid stability. It’s the one asset that sits across all three. BESS has now proven it can stack multiple revenue streams: energy arbitrage, frequency regulation, capacity, and ancillary services, while also providing the cushion the grid needs during stress. That combination is what turns it from a backup box into a strategic asset.

The real challenge from here is distribution. A battery only delivers its full value when it sits where the grid actually needs it — at the constrained nodes, next to the load and the generation it’s balancing. Getting these assets sited and operated as a coordinated, distributed resource, rather than as isolated installations, is the work that unlocks the strategic value for AI infrastructure.

During your time at Hiveon, you worked extensively on demand response, frequency regulation, and energy optimization programs. What lessons from the Bitcoin mining sector can AI data center operators apply today?

I see the crypto industry as the bridge to flexible consumption, not because it invented demand response, but because it brought a genuinely new kind of load that forced grid operators to build better rules around it. Interruptible-load programs for heavy industry have existed for decades; that part isn’t new. What mining added was a load that is large, concentrated, and almost perfectly discretionary: you can power down with no cost beyond the revenue you forgo, no equipment damage, and nothing to restart.

That’s a rare profile, and it’s why ERCOT created a voluntary curtailment program aimed specifically at large flexible customers like mining facilities and why these loads can register as controllable load resources eligible for demand response and ancillary services.

But I want to be precise about what’s actually transferable, because the speed question gets oversimplified. Having worked on ASIC fleets myself, I can tell you that shifting a large mining load in seconds is genuinely hard. Where mining is exceptionally strong is fast economic curtailment and emergency load reduction, responding to price signals or grid stress over minutes, reliably and at scale.

Frequency regulation, the near-instant, automatic balancing the grid uses to hold its frequency steady, is a different and much harder problem because it requires continuous, proportional control across the whole fleet. It can be done: I was involved in integrating ASICs into frequency regulation in one Nordic market, where reaction times were under five seconds. But it grew more complex over time, and I have not seen miners deliver true frequency regulation in the US the way they’ve delivered demand response and curtailment.

So the real lesson for AI operators is not to “copy the curtailment profile.” It is that the market mechanisms, regulatory frameworks, and operator relationships already exist, and crypto played a major role in proving that they can work at scale.

What is striking is how little of that groundwork traditional data center operators have done with their grid operators. AI also has an important asymmetry: model training is far less interruptible than mining, given checkpointing requirements and uptime commitments, even if inference workloads can be more flexible. The lesson to adopt is the engagement model and market participation, not the assumption that AI load can be curtailed like a mining rig.

AI data centers are often portrayed as massive consumers of electricity. Could they eventually become grid-supporting assets that help stabilize energy markets rather than simply add demand?

Absolutely. I am confident that data centers, as a category, significantly underutilize their ability to operate flexibly. These are programmable machines. By their nature, they can be scheduled and shaped to run in particular patterns. I do not want to overgeneralize, because not every workload is interruptible. But even for the portion that cannot be interrupted, there is infrastructure on site, including backup generation and, increasingly, batteries, that can support the grid or shed load when the grid needs it.

The underlying reality has changed. Consumption is no longer a flat line, and with renewables now making up a significant share of generation, supply is no longer flat either. There are peak periods and low-demand periods, and data centers with the right tools and grid incentives can help absorb that variability rather than simply adding stress to the system. ERCOT has already shown the precedent, with flexible large loads providing regulation and reserve services during periods of grid stress.

For most operators, the gap isn’t capability; it’s willingness. And the case for being flexible goes beyond the extra revenue. It’s also about long-term security of supply. If you flex, you keep the system stable and avoid running your own supply chain at the absolute limit, where a single failure upstream can take you down. Flexibility is both a revenue stream and an insurance policy. The operators who treat it that way will be the ones the grid wants to keep.

You have a track record across the full energy value chain, and have overseen projects financed by major international institutions. What separates an AI infrastructure project that is attractive to long-term infrastructure investors from one that is not?

The clearest distinction is maturity. The AI data center industry is still very young, and it has to go through a period of establishment in which revenue models, operational risk frameworks, the skillset base, and the supply chain all mature. Right now, many players are entering the market from very different backgrounds, but very few have full value chain expertise across power generation, grid integration, financing, construction, and compute management. That kind of end-to-end coordination is rare, and it is precisely what long-term infrastructure investors are looking for, because it is what makes cash flows more predictable.

What separates a fundable project from one that is not fundable is whether the entire chain has been thought through and coordinated under one coherent plan. It cannot be a strong site with a weak power strategy or a good power deal with no real understanding of how the compute load will actually behave. Infrastructure capital underwrites durability and predictability over decades. The projects that win are the ones where every link has been genuinely integrated, so the investor is not being asked to bet on five separate things coming together by luck.

There is growing discussion around on-site power generation, microgrids, natural gas, renewables, and hybrid energy systems. Which energy strategies do you believe are most likely to support the next wave of AI infrastructure growth?

“Bring your own generation” has become close to a precondition for new data center development, because grids are facing demand spikes they cannot safely meet on their own. We are already seeing this in the data. Developers in the US are increasingly moving ahead with on-site gas generation, and the IEA estimates that 15 to 27 GW of on-site gas capacity could be powering data centers by 2030.

But I would push back on answering this too generally, because the right strategy is genuinely region and site-specific. Location determines what is available, what can be permitted, and what the grid will allow. That said, when speed of deployment is the priority, and right now it almost always is, gas generation is clearly the most attractive option. The alternatives are either intermittent, such as wind and solar, or take years to bring online, such as nuclear, hydro, or coal.

In practice, the near-term wave will likely be led by gas, often paired with storage and renewables in a hybrid configuration. Firmer and cleaner sources can then phase in over a longer horizon as they become available. It is also worth being honest that on-site gas is not a free lunch. Reliable firm supply for a variable AI load can require significant overbuilding, and gas turbines themselves are in short supply. That is why the hybrid approach is what makes the model more durable.

Looking ahead five to ten years, what do you think the AI industry still underestimates about the relationship between compute, energy, grid flexibility, and infrastructure finance, and what should leaders be preparing for today?

It is hard to play oracle in an environment moving this fast, but a few things look clear to me.

First, all data centers will become grid interactive. This will not be a nice-to-have feature. It will become a condition of getting connected at all. We can already see grid operators moving in this direction, with PJM and ERCOT creating new categories of large loads that accept curtailment in exchange for faster access. Flexibility is becoming the price of admission.

Second, developers will increasingly chase stranded power and move toward vertical integration. That may mean acquiring generation and grid assets, or it may mean building that expertise in house. When the interconnection queue is the bottleneck, owning more of the value chain is how you gain more control over your own timeline.

Third, the negative public image around data center development will start to change, but only if developers and operators learn to integrate properly with the communities where they operate. That means sharing infrastructure costs fairly, being transparent about water and power use, and creating clear local benefits. A lot of that pressure will also come from the financial side. Infrastructure investors and lenders will demand stronger ESG performance and community alignment as a condition of capital, and that will force the industry to professionalize how it develops.

The projects that treat community integration and environmental performance as core to financing, not as public relations, are the ones that will get funded and built. My advice to leaders today is simple: build flexibility, value chain control, and genuine community alignment into your projects now. In five to ten years, all three will be entry requirements, not differentiators.

Thank you for the great interview, readers who wish to learn more should visit Silicon Foundation.