Unblocking AI Infrastructure Will Define AI’s Next Chapter More Than Models

This Industry Viewpoint was authored by Tom Traugott, SVP of Emerging Technologies at EdgeCore Digital Infrastructure

For most of the AI boom, the limiting question was who could secure the most advanced chips. In August of this year, NVIDIA signaled that the challenge has moved beyond silicon. It announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR intended to mobilize more than $500 billion of third-party capital for AI infrastructure over time. The goal is to finance not just accelerators, but the broader systems required to turn them into usable compute.

At almost the same moment, BloombergNEF highlighted the other side of the equation. The company indicates that powered U.S. AI data center capacity could fall behind available AI server chips as early as 2027, with a gap of roughly 42 GW by 2030. BNEF’s broader AI-chip scenario extends that mismatch to 63 GW by 2033. The market will adapt through faster construction, accelerated refresh cycles, overseas deployment, or fewer chip purchases. But the underlying message is clear: the industry may soon have more silicon than places to plug it in.

That is not a bearish signal for AI. It is a deployment signal. Capital and chip output can scale faster than interconnection studies, transmission upgrades, substations, equipment lead times, local approvals, water and cooling systems, fiber routes, and commissioning. AI’s next chapter will be defined by who can convert hardware orders and capital commitments into reliable megawatts.

The Bottleneck Has Moved Down the Stack

BNEF raised its forecast for installed U.S. data center capacity to 118 GW by 2030 and 194 GW by 2035, yet even that higher buildout falls short of what expected chip shipments could support. This distinction matters because the market often treats announced capacity as though it were operational. A land parcel is not capacity. An interconnection request is not capacity. A GPU reservation is not capacity. Capacity is a commissioned environment with power, cooling, fiber, security, operations, and a workload ready to run.

Execution risk is rising with scale. BNEF found that 52 of the 100 largest announced U.S. projects are being advanced by companies that have never built a data center, representing 87 GW. Experienced developers are also attempting to build campuses that are many times larger than anything they have previously delivered. The market will increasingly separate announced gigawatts from deliverable gigawatts.

Capital Is Necessary; Deliverability Is Scarce

NVIDIA’s financing initiative is important because it brings long-duration institutional capital into compute and infrastructure. That can broaden access to GPUs, reduce financing friction, and support customers that cannot fund multibillion-dollar campuses entirely from their own balance sheets. It also confirms that compute is being treated less like a short-lived IT purchase and more like productive infrastructure with usage-linked revenue.

But capital cannot eliminate the physical critical path. Utility studies, grid upgrades, transformers, switchgear, generators, chillers, liquid-cooling systems, and fiber construction still take time. The strongest development models will combine capital with phased delivery, credible customer commitments, early equipment procurement, utility and community partnerships, and multiple pathways to energization. BNEF estimates that the U.S. grid has never connected more than 10 GW of new data center demand in a single year and concludes that, absent faster grid connections, on-site gas generation would need to play a significant role in meeting its base-case forecast.

Agentic AI Changes the Shape of Demand

The rise of agentic AI compounds the challenge. Training remains compute-intensive, but agentic inference is persistent, interactive, and increasingly continuous. Agents can reason through multi-step tasks, call tools, communicate with other systems, and execute processes over time. That shifts infrastructure demand from occasional bursts toward always-on environments that must deliver consistent performance, low latency, network resilience, and around-the-clock availability.

For telecom and internet infrastructure providers, this puts connectivity back at the center of the AI stack. Large clusters require high-capacity campus fiber, diverse long-haul routes, robust east-west fabrics, and low-latency links among compute, storage, users, and other data centers. Power without network is stranded capacity just as surely as chips without power.

The Winning Infrastructure Model Is Flexible

Not every AI workload belongs in the same facility. Frontier training may favor very large, power-rich campuses, while inference may increasingly distribute closer to enterprises, population centers, and data sources. Infrastructure that can accommodate multiple accelerator generations, direct-to-chip liquid cooling, changing rack densities, and both training and inference workloads will retain more value over time. Geographic diversity, staged energization, and flexible load operation can also turn a single-site constraint into a portfolio optimization problem.

Deployment Will Define AI’s Next Era

AI’s first chapter was defined by model innovation and chip scarcity. The next will be defined by deployment discipline. The industry has the capital, hardware roadmaps, and demand signal to build at unprecedented scale – but not enough “ready” infrastructure. The winners will align silicon, power, cooling, connectivity, capital, and operations, converting them into dependable compute before the chips arrive.

About the author:

Tom is the SVP of Emerging Technologies at EdgeCore Digital Infrastructure where he spearheads the company’s understanding and adoption of emerging technologies with a current focus on the impacts and demands of AI on the data center ecosystem.

While interacting with a broad spectrum of technologists, industry thought leaders, and community stakeholders, Tom ensures that EdgeCore stays at the forefront of innovation, geographic expansion, and future trends in order to deliver on the company’s promise to provide safe, sustainable, and futureproof data center solutions to the world’s largest cloud and technology companies. Tom’s industry experience began in the post-dot com early 2000s period and stretches through the rise of enterprise wholesale colocation in the 2000s, through the rise of hyperscale cloud in the 2010s, and now generative AI driven world.

Prior to joining EdgeCore, Tom worked for Amazon Web Services, where he was accountable for strategy and execution for new and existing regions across EMEA, APAC, and the Americas, along with diligence and strategy for additional regions under evaluation. Prior to Amazon, Tom was co-practice leader of Cassidy Turley’s (now Cushman & Wakefield’s) Data Center Advisory Practice, focused on end-user representation and capital markets transactions. Tom previously worked at CoreSite Realty Corporation as Regional VP of Sales.

Tom holds a B.A. in Social Studies from Harvard College, with honors.

If you haven't already, please take our Reader Survey! Just 3 questions to help us better understand who is reading Telecom Ramblings so we can serve you better!