DCF Show Video: Powering AI When the Grid Can’t

This special edition of The Data Center Frontier Show Podcast is a recast and 2026 update of the DCF Trends Summit 2025 panel moderated by Fengrong Li (FTI Consulting): “From Grid to Onsite Powering: Optimizing Energy Behind the Meter for Data Centers.”

The premise is simple and increasingly unavoidable: utility interconnection timelines are colliding with AI’s buildout clock. With multi-year waits for firm capacity in key markets, the industry is treating “onsite” less like backup insurance and more like primary infrastructure—a way to buy schedule certainty when the grid can’t deliver it.

Across the discussion, the panelists converge on several themes:

• Behind-the-meter is still early—and messy. There’s no standard architecture yet, and every project becomes a bespoke integration problem.

• Modularity is the common denominator. Whether it’s distributed fuel cells, 75 MW nuclear blocks, or phased turbine plants, “build in blocks” matches how AI campuses actually scale.

• Load volatility is a design driver. AI’s transient swings force new thinking about buffering, smoothing, and “grid-like” behavior inside an islanded microgrid.

• Contracts are becoming infrastructure. Take-or-pay, deposits, and longer-term PPAs are emerging as the mechanisms that unlock equipment queues, fuel commitments, and financing.

• The endgame is unresolved. Some see eventual grid reintegration as inevitable; others argue many sites will never voluntarily return—because SLAs, control, and economics don’t allow it.

PODCAST TRANSCRIPT

Matt Vincent (Data Center Frontier):
Hello and welcome to a special edition of The DCF Show Podcast. I’m Matt Vincent, Editor in Chief with Data Center Frontier.

At the 2025 DCF Trends Summit last August, one of the defining conversations focused on a central AI-era question—how do you power data centers when the grid can’t keep pace? Since then, constraints and interconnection delays have only intensified, pushing onsite and behind-the-meter strategies to the forefront of infrastructure planning.

Given how urgent this shift has become, today we’re recasting and updating that power discussion for 2026—bringing back much of the original panel and adding fresh perspective. That summit conversation was led by Fengrong Li, Senior Managing Director at FTI Consulting, and we’re pleased to have her back moderating today. Fengrong, over to you.

Fengrong Li (FTI Consulting) (Moderator):
Thank you, Matthew—and welcome to our panelists. Today we’re focusing on one of the most urgent topics shaping digital infrastructure deployment: onsite power and the rise of co-located, integrated power and AI data center campuses.

These models are accelerating data center expansion beyond traditional hubs, changing redundancy strategies, and reshaping how data centers interact with the grid. Let’s start with quick introductions—your role and location. I’m in Washington, D.C., and happy to moderate today.

Yuval Bachar (ECL):
Hi, I’m Yuval Bachar, CEO of ECL. We’re headquartered in the San Francisco Bay Area.

Brian Gitt (Oklo):
I’m Brian Gitt, SVP and Head of Business Development at Oklo. We’re an advanced nuclear company—we design, build, and operate power plants to power AI factories and cloud data centers.

Marty Trivette (AlphaStruxure):
Hi everyone—I’m Marty Trivette, SVP of Energy Solutions at AlphaStruxure. We build onsite, behind-the-meter power plants using a full-lifecycle model: design, finance, build, own, operate, and maintain. We’re headquartered in Boston.

David Blank (Siemens Energy):
Thank you. I’m David Blank, Director, Distributed Sales North America at Siemens Energy, based in Orlando, Florida.

Fengrong Li:
Great—let’s jump in. Onsite generation is increasingly competitive, but it’s still fragmented—no standardized solutions. How would you categorize the current state of onsite power for large-scale AI data centers, and what trends are shaping adoption today?

David Blank (Siemens Energy):
From the OEM perspective, it’s still in its infancy. Behind-the-meter as a primary-power application has really accelerated in the last year or so, and customers are still figuring out what works. There isn’t a standard blueprint. What works for one project doesn’t necessarily work for the next.

And the driver is straightforward: everyone would prefer grid power. But in many cases, reliable access isn’t available for five, ten, even ten-plus years—especially at the scale AI is demanding. So behind-the-meter becomes the only way to hit timelines.

Brian Gitt (Oklo):
In some ways, this is back to the future. Historically, large industrial loads built their own plants because they needed to. Data centers have always had onsite generation, but it was mostly backup. Now it’s evolving toward primary power at a much larger scale.

The scale is the headline. We’re talking about multi-gigawatt campus master plans. The grid can’t meet that on these timeframes. But I also think “behind-the-meter forever” is hard politically and maybe hard operationally in some regions. Speed-to-power matters now—but over time, there will be pressure for these assets to integrate with the grid and contribute value.

Yuval Bachar (ECL):
I think many sites that go behind the meter will never go back. Utilities are saying five to seven years to catch up—and they’re planning for a future demand profile that’s a moving target.

There’s also a fundamental issue: data center operators have SLAs. If the grid can pull power back from a behind-the-meter campus, what happens to availability guarantees? Control matters. And once operators realize they can own the system, control their destiny, and potentially improve economics, the incentive to stay behind the meter is strong.

Marty Trivette (AlphaStruxure):
I agree—and I’d add financing and rate-base separation. One reason behind-the-meter is gaining momentum is that it cleanly separates costs. If the data center offtaker contracts for the power onsite, you avoid the perception that residential ratepayers are subsidizing upgrades.

Historically, some sites built generation and later mothballed it when utility service became cheaper. I think we’ll see a mix. But once you have an effective onsite operating model—and it’s reliable—there’s less reason to move back.

Fengrong Li:
Let’s extend that. Onsite generation changes everything downstream—redundancy strategy, fuel choices, water, permitting, backup power. How are you seeing onsite power reshape these secondary and tertiary design elements?

David Blank (Siemens Energy):
It’s a complete shift. You’re effectively building a utility-scale power plant. You need an EPC. You need long-lead equipment. You need permits. You need gas and water plans. And then you have to make the plant behave like what the data center expects—reliability levels, redundancy, and load dynamics.

AI load swings are real. When load moves quickly, you need a system that can handle that without risking stability. There isn’t a single standard approach yet. That’s why each project ends up being a custom integration exercise.

Yuval Bachar (ECL):
A lot of this depends on architecture. One approach is a centralized plant. Another is distributed generation. We build smaller blocks—each with localized generation close to the load. That reduces dependence on high-voltage infrastructure, long transmission runs, and some of the supply chain choke points like large transformers and switchgear.

It also opens the door to different sustainability outcomes. If you can eliminate local emissions and reduce water impact, that changes the conversation with communities—especially as inference sites move closer to cities where power is scarce and timelines are long.

Brian Gitt (Oklo):
Every project will be bespoke for a while. A lot of teams are optimizing for speed, not lowest cost. And when you’re recreating grid-like capabilities on a site, that adds cost and complexity. Over time, we’ll standardize more. But in the near term, it’s going to be an ecosystem of different configurations, phased builds, and hybrid solutions.

Marty Trivette (AlphaStruxure):
And constraints differ by location. Urban inference sites have emissions and noise constraints. Some projects lean toward fuel cells. Larger sites can support turbines. Others might be mixed DER. A common thread is modularity—repeatable building blocks, then adapt to the site’s permitting, fuel, and uptime profile.

Fengrong Li:
Yuval—ECL is known for a modular, distributed hydrogen-powered fuel cell approach. What’s the key advantage versus a centralized, larger plant?

Yuval Bachar (ECL):
Localized generation reduces the risk of concentrating all energy in one place, and it gives you flexibility. AI is changing fast—nine to twelve-month cycles. Power generation hardware doesn’t evolve on that cadence. With block-by-block buildouts, the blocks you deploy in 2027 don’t have to look like the blocks you deployed in 2025.

Hydrogen also enables a different environmental profile—zero local emissions, and the byproduct is water. That can change how communities respond. And because we generate at low voltage or DC close to the load, we reduce dependency on some of the high-voltage supply chain bottlenecks.

Fengrong Li:
What happens when you move closer to cities without hydrogen infrastructure?

Yuval Bachar (ECL):
For smaller sites, delivered liquid hydrogen can work—especially in markets where grid pricing is high. But scaling delivery has limits. That’s why we’re also focused on flexible architectures—mixing sources where needed, including hydrogen, natural gas with carbon capture, and grid inputs.

Urban environments often have constraints on natural gas pressure and flow too, so the future is going to be blended. We’ll use every energy source we can, because demand is simply too high.

Fengrong Li:
Brian—Oklo announced a significant deal with Meta in Ohio. What does that signal about hyperscaler procurement models—and replication?

Brian Gitt (Oklo):
It signals a move toward additionality and synchronized buildouts. In markets like PJM, the direction of travel is clear: large loads will need to enable corresponding generation—one-for-one in some form, whether onsite or offsite.

Our model is modular—75 megawatt plants scaled in blocks. That matches how data centers actually expand: phased buildings, phased loads. The advantage is the ability to coordinate power and compute growth in a way that supports time-to-power, while avoiding the mega-project complexity that has historically created schedule and cost risk in the power sector.

Fengrong Li:
David—turbines are in record demand. How do you see gas turbines serving as near-term backbone—and a bridge to lower-carbon firm power? And what can you say about lead times?

David Blank (Siemens Energy):
Gas turbines are proven, reliable, and power-dense—so they’re a major enabler for near- and mid-term growth. We also have pathways for alternative fuels, including hydrogen blends, depending on what customers can access.

Lead times have extended because demand has surged. Larger units are pushing into the early 2030s. Generally, the smaller the turbine, the faster the lead time.

Fengrong Li:
Marty—AlphaStruxure integrates multiple technologies. What are you deploying most often, and how do you handle load-following and reliability?

Marty Trivette (AlphaStruxure):
Today it’s mostly combinations of engines, fuel cells, and turbines—sometimes combined cycle—depending on size, location, and timeline.

Load-following is a real design driver. When you have sharp swings, you may need fast-responding systems—supercapacitors, flywheels, battery storage, and load banks—to shape the curve. Data hall tier matters too. Tier 3 and Tier 4 architectures can smooth load inherently via UPS systems. Tier 2 environments can be more challenging and may require more aggressive buffering.

We’re also seeing movement toward smoothing at the server level—filler workloads that reduce spikes. That helps compute economics and power stability. Ultimately, the requirement is simple: deliver reliability that meets or exceeds grid expectations.

Fengrong Li:
We’re close on time. Onsite power requires significant capital and integration complexity. How are contract structures—take-or-pay, duration, credit requirements—affecting the pace of development? How early are customers stepping in to provide certainty?

David Blank (Siemens Energy):
We’re typically not contracting directly with the end user, but we are very selective. We want to see clear offtake commitment before major equipment decisions move forward.

Brian Gitt (Oklo):
We use capacity reservation deposits or prepayments to lock in capacity and support long-lead commitments. That customer certainty helps unlock critical inputs and project sequencing.

Yuval Bachar (ECL):
Long-term fuel agreements matter. Growth can require shared investment in expanding supply. But as the cost of compute continues to rise, power infrastructure can become a smaller portion of overall project cost, and customers are willing to participate if it accelerates deployment and ensures availability.

Marty Trivette (AlphaStruxure):
Financing requires creditworthy offtake. The sooner the ultimate offtaker is committed and backstopping the project, the faster it can move—subject to equipment and fuel lead times. Credit profile affects pricing too.

Fengrong Li:
Thank you all—excellent discussion. Matthew, back to you.

Matt Vincent (Data Center Frontier):
Thanks, Fengrong—and thanks to all of you for joining us for a timely, insightful panel. We’ll look to revisit this topic down the road and keep our finger on the pulse of onsite and behind-the-meter energy for the data center industry in this AI era. Thanks to all.

At Data Center Frontier, we talk the industry talk and walk the industry walk. In that spirit, DCF Staff members may occasionally use AI tools to assist with content. Elements of this article were created with help from OpenAI's GPT5.

Keep pace with the fast-moving world of data centers and cloud computing by connecting with Data Center Frontier on LinkedIn, following us on X/Twitter and Facebook, as well as on BlueSky, and signing up for our weekly newsletters using the form below.