"Not Your Grandfather's Data Center:" Reflections on Bechtel's Case for AI Infrastructure

On April 8, 2026, Deep Tech at Duke hosted an AI Infrastructure Working Group headlined by remarks from Provost Alec Gallimore with a keynote presentation by Catherine Hunt Ryan, President of Bechtel’s Manufacturing & Technology. Her presentation, titled “Deploying AI Infrastructure at Scale: A Perspective from Engineering & Construction,” offered a rare vantage point of those at the forefront of building necessary AI infrastructure. Through five critical challenges, Hunt Ryan made a compelling case that the physical and commercial realities of AI infrastructure are outpacing the frameworks the industry currently uses to deliver it and closing that gap requires a shift in how we think about AI development as a whole. 

Challenge 1: AI Infrastructure is a System, Not a Building

Hunt Ryan’s central provocation was deceptively simple: “This is not your grandfather’s data center.”¹ AI infrastructure is no longer just a data center; it is a mega-integrated system that encompasses power generation, water treatment, substations, backup power, carbon capture, and cooling infrastructure, all working in concert with the computing facility itself.² What began as a cloud data center connected to a substation is evolving into a campus rivaling the complexity of a major industrial facility, potentially complete with small modular reactors, combined cycle gas turbines, and wastewater treatment.³ 

For students at Duke who might be accustomed to thinking about AI in terms of software, algorithms, and occasionally chips, this framing is a useful corrective. The intelligence of AI systems runs on physical infrastructure, and the decisions made today about how to build infrastructure will shape the trajectory of the technology for decades. As Hunt Ryan explained, the decisions made early in design regarding cooling architecture, rack density, power compatibility, and electrical load assumptions propagate through a system whose site and shell may stand for thirty or forty years, even as the IT hardware inside it is replaced every four to six years.⁴ She framed this as a three-layer lifecycle problem with long-lived physical assets, medium-cadence facility systems forced to adapt faster by AI's density demands, and rapidly refreshing IT hardware. Allow misalignment within those layers, and the costs compound quickly.⁵ Hunt Ryan’s presentation served as a reminder that AI computation is a megaproject and should be viewed as an interconnected system—not just a building. 

Challenge 2: Capital Intensity is New to Software-Minded Market

For an industry built on software, where marginal costs approach zero and scale is achieved through replication rather than construction, the physical demands of AI infrastructure represent a novel challenge. Hunt Ryan offered figures that illustrate the scale of the shift. Specifically, a single gigawatt of AI compute requires roughly $15 billion in engineering, procurement, and construction costs, plus approximately $35 billion in chips.⁶ The same gigawatt of dedicated power generation adds more than $2 billion in additional engineering, procurement, and construction (EPC) costs.⁷

The workforce implications are equally striking. A one-gigawatt AI data center may require more than 5,000 craft professionals at peak construction and around 500 field professional staff.⁸ That figure is one that requires regional labor market planning, workforce development pipelines, and procurement strategies that must begin years before a shovel breaks ground. Rebecca Kujawa, CEO of Zerra Partners, later reinforced this point from an energy perspective, noting that states such as North Carolina are only beginning to grapple with what it means for the energy system when customers shift from being capital-light to capital-intensive at this scale.⁹ The permitting requirements alone for hyperscaler facilities present a degree of complexity that regulators and utilities are still working to understand.¹⁰ Hunt Ryan similarly observed that hyperscalers accustomed to scaling through software deployments now carry “more concrete on their balance sheets than a construction company.”¹¹ The companies sponsoring these systems are learning, sometimes uncomfortably, that physical assets require a different set of competencies than software products.

Challenge 3: Speed is the Strategic Variable—and the Hardest Problem

Hunt Ryan’s data on the financial stakes of construction timelines reframed how I think about the urgency surrounding AI infrastructure. A one-gigawatt AI data center generates – using a rather conservative estimate - an estimated $29 million in daily revenue once operational.¹² That is more than a semiconductor fab, an LNG train, a major airport, or a copper mine. She noted that a six-month shift in schedule, in either direction, can swing more than two billion dollars in cumulative revenue impact.

Yet, as Hunt Ryan made clear, the commercial frameworks and operating models governing how these facilities get built have not caught up to that reality. Contracts are still being negotiated, risk is being allocated, and projects are being staffed using frameworks designed for a slower-moving industrial world. The speed premium is real and quantifiable, but the institutional response to it is not. This gap between what the technology market demands and what the construction industry can deliver is not purely an engineering problem. It is a coordination problem, a commercial problem, and in some respects, an educational one—one that Hunt Ryan argued will not be solved by any one discipline acting alone.

Challenge 4: AI Infrastructure Has to Expect Change

As AI continues to push the demand for power and cooling design, physical infrastructure must be built not just to serve today’s AI workloads, but to accommodate the workloads of the next generation of hardware, which will be denser, hotter, and hungrier for power. Hunt Ryan cited Masayoshi Son of SoftBank, who observed that computing components will need to be refreshed roughly every five years, at an estimated cost of $300 million each time.¹³ At a multi-gigawatt campus scale, cumulative refresh spend can reach tens of billions per cycle.

Hunt Ryan’s framing here was particularly memorable. She stated that change is “not a monster under the bed, it’s the new alarm clock.” The ability to manage change by absorbing scope shifts, incorporating technology updates, and designing iterations without grinding a project to a halt is not a nice-to-have in this environment. She argued that this openness to adaptation is the primary source of competitive differentiation for engineering and construction firms. 

Challenge 5: Power is the Binding Constraint

Perhaps the most consequential reframe in Hunt Ryan's lecture was her treatment of power. For most of the data center industry's history, power has been a utility input where businesses select a site, connect to the grid, and pay the bill. Hunt Ryan stated that the model is ending with grid interconnection timelines that are now stretching into multiple years. Existing grid capacity is dwindling as AI-driven demand accelerates. 

AI workloads create sharp, non-linear demand patterns that stress traditional grid assumptions, and hyperscalers and sovereign AI customers expect near-continuous, five-nines availability that cannot tolerate those stresses. The result is a push toward behind-the-meter generation, adjacent power plants built specifically for the data center campus, and in some cases, islanded energy systems. Hunt Ryan’s conclusion was stated without hedging: for AI, power is no longer a utility input; it is a megaproject that must be phased, designed, and delivered in lockstep with the data center. Power and compute, at a gigawatt scale, are not two projects. They are one asset.

The Case for Multidisciplinary Thinking 

What runs beneath every argument Hunt Ryan made is a conviction that these problems cannot be solved within any single discipline. The people making investment decisions about AI infrastructure often have limited intuitions about physical systems, permitting timelines, labor markets, or the mechanics of power generation. The people building those systems often lack visibility into the commercial logic driving schedule pressure or the technology roadmap shaping future refresh cycles. Neither group has historically had strong structural reasons to sit in the same room. Hunt Ryan is arguing that they need to—urgently.

Her five considerations for adapting to AI infrastructure challenges are each arguments for breaking down those silos: integrated EPC delivery with a single data model, change management as a competitive differentiator, programmatic rather than bespoke approaches to delivery, unified power-and-compute planning, and her fifth and perhaps most important point of multidisciplinary agility.

That last consideration deserves particular attention. Hunt Ryan argued that engineering and construction optimization must be intimately linked with the commercial models of delivery, and that understanding the core infrastructure asset requires systems-based, multi-disciplinary education and experience. You cannot design a data center well without understanding power. You cannot negotiate a construction contract well without understanding technology refresh cycles. You cannot site a facility well without simultaneously solving for permitting, water access, and grid constraints. Ultimately, the workforce that will build and govern AI infrastructure does not yet fully exist, and creating it requires deliberate investment in interdisciplinary education and experience.

Why This Matters Beyond Engineering

Countries, companies, and institutions worldwide are confronting a version of the same question: How do we build the physical foundations of a technology-dependent future, quickly enough and wisely enough to matter? How do we coordinate across industries, timelines, and expertise domains that have never had much reason to overlap? And how do we train the next generation of leaders to think in systems rather than in silos?

Hunt Ryan’s lecture does not pretend to answer all these questions. But it names them with unusual clarity. The data centers built today will shape where AI capacity exists, who can access it, what it costs to run, and how resilient it is to disruption for decades to come. Getting those decisions right is not a purely technical challenge. It is organizational, commercial, regulatory, and educational all at once.

For those of us engaged in technology and policy, sessions like this are a reminder that the most consequential work in the AI era will live at intersections between nations and industries, between software and concrete, between the millisecond world of compute and the multi-decade world of infrastructure. Understanding those intersections, as Hunt Ryan makes clear, is no longer optional. It is the work. 

1. Catherine Hunt Ryan, “Deploying AI Infrastructure at Scale . . . A perspective from engineering & construction” (presentation, AI Infrastructure Working Session, Durham, NC, April 8, 2026). 
2. Ibid. 
3. Ibid. 
4. Ibid. 
5. Ibid. 
6. Ibid. 
7. Ibid. 
8. Ibid. 
9. Rebecca Kujawa, “The Infrastructure Imperative: AI’s Physical Infrastructure” (panel discussion, Duke University’s AI Infrastructure Working Group Public Panel, Durham, NC, April 8, 2026). 
10. Ibid. 
11. Ibid. 
12. Catherine Hunt Ryan, “Deploying AI Infrastructure at Scale . . . A perspective from engineering & construction” (presentation, AI Infrastructure Working Session, Durham, NC, April 8, 2026).
13. Ibid.