The Forum was moderated by John Geraghty, National Practice Leader of Energy & Infrastructure at Marshall & Stevens. Panelists included Bobby Majumder, Partner and Co-Chair of the Energy Industry Team at FBT Gibbons; Ken Malik, Head of Project Development at Grupo Cobra; Brent Nelson, Senior Managing Director of Markets and Strategy at Ascend Analytics; and Fahad Siddiqui, Director of Structured Finance at TotalEnergies.
“ERCOT is talking about 400 gigawatts of data center demand. That’s just not happening. Ten percent, maximum.”
Ascend Analytics’ Brent Nelson said it flatly, and the panel’s estimates were all well below the announced pipeline.
At the Forum, the data center power discussion produced some striking estimates, not about the potential scale of demand, but about how much of the announced pipeline will actually be built and what kind of power it actually needs. The discussion also highlighted infrastructure constraints that may still be underappreciated in parts of the market.
The announced pipeline is likely overstated
The most provocative exchange of the forum was the rapid-fire question on what percentage of the announced data center pipeline will actually come online.
Bobby Majumder estimated roughly 25% of announced projects ultimately get built, while Ken Malik and Fahad Siddiqui placed the number closer to 40–50%. Brent Nelson was more skeptical about ERCOT specifically, estimating roughly 10% there, while suggesting regulated utility territories outside ERCOT may ultimately support a higher percentage of announced demand, particularly where large hyperscaler-utility arrangements are already emerging.
“There’s not the capital, there’s not the labor, there’s not the chips,” Nelson said of ERCOT’s projections specifically.
For investors underwriting energy exposure to the AI build-out, the announced pipeline number is not the investment thesis. The deliverable percentage may be the more relevant question.
Training-model and inference-model data centers are not the same energy problem
Brent Nelson emphasized a distinction that became increasingly important throughout the discussion: training-model and inference-model data centers do not create the same energy demand profile. Training model data centers, which are building the AI models, need speed to power above everything else.
The competitive pressure is significant:
“If you fall behind on the exponential curve, you fall behind fast. The cost of losing the AI race for any given company is much higher than the cost of whatever they have to pay to get power.”
Inference model data centers, which serve queries at scale, are a different story. As AI matures and cost competition increases, inference models will become more willing to accept latency, more willing to shift compute geographically, and more cost-conscious.
“I think 10 years from now, we’re going to be very much more in that space,” Nelson said.
Infrastructure designed around today’s training-driven demand may not perfectly align with a more inference-heavy world by 2032.
Battery storage is becoming operationally central
Some large-scale data centers are already installing battery storage as operational load buffers because AI training creates “hundreds of megawatts to gigawatt-level variations in load” that are difficult for grids and on-site systems to absorb without some form of buffering. During the discussion, Brent Nelson described battery storage as increasingly connected to the operational realities of managing large-scale AI power demand.
That operational role also changes how storage is evaluated economically. As Nelson explained:
“If you already have storage as a load buffer, you might as well think about what else you can do with it.”
Ken Malik added operational context to the discussion, noting that many data center load profiles involve concentrated periods of very high demand, which increases the relevance of battery storage as part of the broader power strategy.
Transmission and water may be among the most underappreciated constraints in the data center power equation
Ken Malik named transmission scarcity as the most underhyped risk in the data center power equation, specifically the structural gap between where generation is available and where data centers are located or want to be.
Bobby Majumder flagged what he sees as an underappreciated variable in the data center buildout:
“Nobody is talking about cooling. Nobody is talking about water.”
He warned that large-scale data center development in water-constrained regions could face increasing scrutiny around cooling requirements and aquifer usage:
“The farmers are not going to be happy with you at all about you pumping down their aquifer for cooling.”
More broadly, the panel emphasized that the data center demand story is real, but that the timeline, geography, and infrastructure requirements behind that demand are likely more constrained and region-specific than headline pipeline figures suggest. Transmission access, interconnection, and existing power availability repeatedly emerged as key differentiators in what projects are most likely to move forward.
The full conversation on what it takes to develop and finance projects in 2026 is available on demand. Watch the recording or download the 2026 Energy Outlook Report for our complete analysis.
