AI’s next bottleneck isn’t computing. It’s infrastructure.
The data center industry may build more capacity in the next three years than it did in the previous 30 combined.1 That was the starting premise at a recent panel on AI infrastructure at the KeyBanc Capital Markets Technology Leadership Forum.
Moderated by Bill Benjamin, Managing Director, Mosaic, KeyBanc Capital Markets, the discussion featured Dean Nelson, CEO of Cato Digital and Founder and Chairman of Infrastructure Masons; Mike Nguyen, Founder and CEO of Inflect; Lei Pan, Chief Architect at Inflection AI; and Zachary Smith, Co-Founder and CEO of Datum. The panel explored how the rapid expansion of AI is no longer just a technology story. It is increasingly an infrastructure story.
Nelson set the stage with data from Mason’s 2026 State of the Digital Infrastructure Industry Report, a professional association that includes more than 6,000 professionals representing $1.5 billion of infrastructure across 130 countries. For three consecutive years, he noted, the group’s forecasts have proven too conservative. In 2024, the expectation was that the industry would double in five years and triple in 10. By 2025, that had accelerated to doubling in two years and tripling in five. This year, he said, the industry had tripled in three years, with 10 times growth now in view.
Even if half of that materializes, we’re going to build more capacity in the next three years than we have in the last 30.
Dean Nelson
Cato Digital / Infrastructure Masons
373 gigawatts — and what it will take to get there
Global capacity in the development pipeline has grown from 105 gigawatts (GW) to 373 GW, said Nelson, including 71 GW active, 22 GW under construction, and 280 GW in development. For its part, the U.S. accounts for 228 GW — or approximately 62 percent of global capacity.1 Nelson said Texas alone accounts for approximately 100 GW of the capacity tracked in the report, much of it associated with proposed development rather than operating facilities.
Building at that pace is not without its challenges. Nelson pointed to five active constraints:
- Power. Time to power, including the time required to secure generation and grid connections, is a critical bottleneck.
- Pushback. Community and regulatory opposition is rising. Data Center Watch identified at least $156 billion of infrastructure projects blocked or delayed in 2025.
- People. There may not be enough skilled construction laborers, including electricians, plumbers, and welders, to keep pace with demand.
- Policy. Permitting timelines and regulatory decisions can add years to a project or fast track it.
- Planet. Data centers are heavy consumers of energy and water, and that toll is coming under increasing scrutiny.
Power is the first question
For many developers, investors, and operators, power availability has become more of a pressing concern than sheer cost. Utilities are facing interconnection queues that cannot keep pace with demand. In Texas alone, Nelson cited more than 300 GW of requests, while acknowledging that some of that figure reflects multiple developers pursuing the same opportunities.
Nguyen, who works across data center, fiber, GPU as a Service, and large-scale infrastructure transactions, said the first diligence question is to ask point blank whether power is available, or not.
Do you have it? That’s question one.
Mike Nguyen
Inflect
The answer, he said, is rarely simple. Buyers need to know whether the power is grid-connected or behind the meter, whether long-lead components are secured, whether multiple connections exist, and whether other constraints such as water or network availability could still break the deal. Deals can fall apart for any one of those reasons, Nguyen said, with power closely followed by other infrastructure dependencies.
Nelson described a market in which data center operators are becoming active grid participants out of necessity. He pointed to Crusoe’s Abilene, Texas project as an example of speed at unprecedented scale: 200 MW delivered “from dirt to token” in 12 months,2 supported by upfront procurement, power access, and thousands of construction workers on site.
Inference changes the architecture
While training large foundation models still drives much of today’s power demand, the panel said inference — the process of running AI models in production — will reshape infrastructure requirements over time. Nguyen said the market is still catching up with training-driven demand, with GPU cloud capacity heavily committed for the next 12 to 18 months. But he pointed to what comes next: lower-latency deployments, metro markets, and the network required to connect data, storage, and compute.
You have to deliver the tokens to places that are not data centers at some point.
Mike Nguyen
Inflect
Pan added the practitioner’s view from an AI lab. Training and inference, he said, are fundamentally different workloads. Training is predictable and can be scheduled for throughput, while inference is marked by variable demand, latency sensitivity, and complex movement across compute, network, and memory. In agentic systems, where multiple models and reasoning loops may be involved, networking and memory become part of the compute itself.
Networking and memory are really a part of the compute. You can’t separate those things.
Lei Pan
Inflection AI
That shift has meaningful implications for infrastructure design. The future is likely to be more heterogeneous, with different workloads served by different combinations of hyperscalers, neoclouds, specialized chips, CPUs, GPUs, memory architectures, and network configurations. As Pan noted, the software stack will need to evolve alongside the physical buildout, especially as AI systems orchestrate across multiple providers and hardware environments.
That heterogeneity extends beyond hardware to geography — where computing happens, not just what kind.
Not every AI workload belongs in a gigawatt campus
The panel revisited a long-running infrastructure question: whether edge computing, which places processing closer to users and data, will materialize at scale. Smith argued that the answer is yes, but not necessarily by placing AI infrastructure on every phone pole or in every home. Instead, he pointed to metro-level intelligence, where computing power and data move closer to users, enterprises, and local networks.
Smith also tied the shift buildout to a broader societal shift. AI has made digital infrastructure more visible and more personal, which means communities, states, and countries are likely to demand more transparency, governance, and localization. In his view, technology is now geopolitical, and enterprises will face growing complexity around where data lives, who accesses it, how it is secured, and which jurisdictions shape its use.
Technology is a geopolitical thing now.
Zachary Smith
Datum
Community acceptance is becoming part of the build
Public acceptance, several panelists agreed, is no longer a secondary issue. Data centers, once largely invisible, are now physical symbols of AI’s impact on jobs, energy, water, and local communities. Smith said the industry needs to think less like a purely private buildout and more like a public-private partnership, with infrastructure companies becoming part of the communities where they operate.
Nguyen went further, arguing that neutrality itself isn’t enough. Communities do not simply want data centers to avoid harm. They increasingly expect them to contribute tangible local benefits, whether through investment in power systems, water infrastructure, jobs, training, or other community priorities.
Nelson said the industry has been “very, very quiet for very long,” even as many operators and hyperscalers have invested in community efforts like parks, water treatment, and local development. The challenge now, he said, is to make those efforts visible, credible, and coordinated — especially with paused or stopped projects already creating financial and strategic risk.
Efficiency helps, but may not reduce demand
The panel cautioned against assuming efficiency alone will solve the infrastructure challenge. Better software orchestration, hardware design, and workload coordination can reduce waste, said Pan, noting that some customer environments operate at less than 20 percent GPU utilization.
Better utilization could reduce unnecessary demand and improve economics across the stack, he added.
But Nelson warned that efficiency can cut the other way, invoking the Jevons paradox: as AI becomes cheaper, faster, and more useful, demand may expand rather than decline. Inference, he argued, will be more pervasive than many expect, and infrastructure will need to be built more broadly than just gigawatt-scale campuses.
Pan closed with a related point for enterprises and investors: the industry may be moving beyond simple metrics such as cost per token or number of GPUs. The more meaningful measure, he suggested, may become cost per useful result, or cost per unit of intelligence delivered by an AI system.
AI’s future all comes down to infrastructure
The opportunity remains enormous, but if this panel made one thing clear, it’s that the future of AI may be determined less by who builds the most advanced models, and more by who can successfully build and sustain the infrastructure that makes those models possible while managing the execution risks that come with them.
About the Technology Leadership Forum Conference
The 27th annual Technology Leadership Forum, held in Deer Valley, Utah, brought together nearly 1,000 technology leaders to network and share insights across a diverse audience of public and private companies, industry leaders, and institutional, venture capital, and private equity investors. The forum was packed with thought-provoking presentations and insights on emerging trends. If you would like more information about attending this conference, contact our Corporate Access team.
To learn more, visit our Technology Investment Banking and Institutional Equity pages.