Building the Future of Al Infrastructure

Every AI application getting attention right now — the copilots, the agents, the generative tools — sits on top of a layer almost nobody outside the industry thinks about: the physical and computational infrastructure that makes any of it possible.

At a recent GCN founder spotlight, we sat down with a founder building in that infrastructure layer — compute orchestration, power procurement, and the unglamorous plumbing that determines whether an AI company can actually scale or whether it hits a wall the moment demand outpaces supply. The conversation was a useful reminder that the most durable opportunities in this cycle may not be the applications everyone can see, but the infrastructure underneath that almost nobody can.

The Problem Nobody Was Solving

The founder's origin story is a familiar one in infrastructure: they didn't set out to build a compute company, they ran into a wall while building something else. After spending over a year trying to reliably provision GPU capacity for a previous machine learning project, they became convinced the actual bottleneck in AI wasn't algorithms or talent — it was physical capacity.

"Everyone was competing for the same fixed pool of high-end chips, and the allocation process was opaque, slow, and relationship-driven rather than efficient," the founder explained. "We'd get quoted six-month lead times for compute that a hyperscaler's biggest customers could get in days. That gap is where the company came from."

The insight wasn't just that compute was scarce — plenty of people had noticed that — but that the scarcity was being managed inefficiently, and inefficiency in a market this large is where venture-scale companies get built.

Why Compute Demand Keeps Outrunning Supply

The founder walked through the structural dynamics driving the gap between AI compute demand and available supply, and why they believe it isn't a temporary, cyclical shortage that resolves itself in a year or two.

Model sizes and training runs have grown at a pace that chip manufacturing, data center construction, and power grid buildout simply cannot match on the same timeline. A new fabrication plant takes years to come online. A new data center campus takes years to permit, build, and energize. Meanwhile, model architectures and inference demand are compounding on a timeline measured in months. That mismatch, the founder argued, is structural, not cyclical — and it means the companies solving allocation, efficiency, and access will matter for a long time, not just through the current hype cycle.

Why Power Became the Real Constraint

Perhaps the most striking part of the conversation was how quickly it moved away from chips entirely and toward energy. "Everyone assumes the bottleneck is silicon. Increasingly, it's electricity," the founder said. "You can have all the GPUs in the world sitting in a warehouse, and it doesn't matter if you can't get reliable, sufficient power to the site to run them."

This has pushed infrastructure builders toward unconventional strategies: co-locating with existing power generation rather than waiting years for new grid capacity, striking direct agreements with utilities and independent power producers, and in some cases evaluating on-site generation entirely separate from the public grid. The founder was candid that this part of the business increasingly resembles energy infrastructure more than it resembles traditional software.

"We didn't expect to become experts in power purchase agreements when we started this company," they said. "But if you're serious about compute at scale, you can't outsource that problem to someone else and hope it works out. Access to power and reliable energy has become mission-critical to how we build the business, not a footnote."

Betting on a Distributed Future

Looking further out, the founder pushed back on the assumption that AI infrastructure will consolidate into a small number of massive, centralized hyperscale facilities. Their bet is on a more fragmented landscape, shaped by three forces pulling in different directions at once.

Data sovereignty requirements are pushing certain workloads toward regional or national infrastructure rather than a handful of global mega-campuses, particularly for regulated industries and government-adjacent work. Latency-sensitive applications — real-time inference embedded in physical products, industrial systems, and consumer devices — are pushing compute closer to the edge, physically nearer to where it's actually used. And highly specialized workloads in sectors like biotech, defense, and financial services increasingly need infrastructure purpose-built for their specific compliance and performance requirements, rather than generic capacity.

"The future of AI infrastructure isn't one shape," the founder said. "It's going to be distributed across edge deployments, sovereign regional capacity, and industry-specific builds, all at once. Betting the whole business on one of those categories is the riskier position, not the safer one."

Building a Team for a Business That Doesn't Fit One Mold

One of the more unexpected threads of the conversation was about hiring. Most early-stage AI companies are optimizing for machine learning talent — researchers, applied scientists, engineers who can fine-tune and deploy models. This founder's hiring priorities looked almost nothing like that.

"Our first ten hires included someone who spent fifteen years doing utility-scale power procurement, and someone who'd built data centers for a hyperscaler," they said. "We needed almost no traditional ML talent in the early days. What we needed were people who understood how to get large amounts of physical infrastructure built, permitted, and energized on realistic timelines, because that was the actual constraint on the business."

This created real recruiting challenges early on. Experienced infrastructure and energy operators weren't used to evaluating equity-heavy startup compensation, and many were skeptical that a young company could offer the stability their prior roles at established utilities or hyperscalers provided. The founder said winning this talent required leaning hard into mission and technical ambition — framing the work as building the physical foundation the entire AI industry depends on, rather than pitching it as a typical fast-growing software startup.

The blend has become a defining feature of the company's culture: software and infrastructure engineers working alongside former utility executives and site-selection specialists, two professional cultures that rarely overlap but that this category increasingly requires in the same room.

Lessons From the Hardest Year

Asked about the most difficult period since founding the company, the founder didn't point to fundraising or early product struggles — they pointed to a specific site development project that fell through after nearly a year of work, when a planned power agreement collapsed late in negotiations due to a change in the utility's regulatory environment.

"We had already committed significant capital and team time to that site. Losing it wasn't just a financial setback, it forced us to rethink our entire approach to how much we relied on any single power partner or single jurisdiction," they said. "It's actually part of why we now believe so strongly in a distributed approach rather than betting everything on one or two flagship locations. That failure directly shaped the strategy we described earlier."

It's a useful reminder that in infrastructure-heavy categories, the biggest risks often aren't competitive or technical — they're regulatory, logistical, and dependent on counterparties the company doesn't fully control. Founders evaluating this space should expect that some of their hardest problems will look nothing like a typical software startup's hardest problems.

What Investors in the Room Wanted to Know

The Q&A portion of the evening surfaced the questions investors are actively wrestling with when it comes to infrastructure plays in this category. Capital intensity came up immediately — infrastructure businesses require real balance-sheet capital, not just software-style venture rounds, and investors wanted to understand the founder's approach to blending equity, project finance, and strategic partnerships to fund physical buildout without over-diluting early.

Defensibility was the second major thread. In a category attracting enormous capital and competition, investors pressed on what actually protects the business once larger, better-capitalized players notice the opportunity — the founder pointed to long-term power agreements, specialized siting relationships, and deep technical expertise in an unglamorous operational layer as harder-to-replicate advantages than pure software moats.

Timing was the third. Several investors asked directly whether this was a multi-year window or a multi-decade one. The founder's answer: as long as model capability keeps improving and inference demand keeps growing — which shows no sign of slowing — the underlying infrastructure gap isn't closing anytime soon.

Advice for Founders Considering This Space

Asked what advice they'd give a founder eyeing an entry into AI infrastructure, the founder was blunt about the tradeoffs involved. This is not a category for founders looking for a fast, capital-light path to product-market fit. Every advantage in this business is earned slowly — through relationships with power providers, through site selection expertise built over years, through operational reliability that only shows up in a track record.

"If you're choosing this space because it's hot right now, that's the wrong reason," they said. "Choose it because you're genuinely willing to spend years becoming an expert in something unglamorous — permitting timelines, transmission capacity, industrial real estate. The founders who win here are the ones who find that stuff interesting, not the ones tolerating it on the way to the exciting part."

They also pushed back gently on founders who assume infrastructure necessarily means massive capital requirements from day one. Their own company started with a much narrower, capital-efficient wedge — reselling and optimizing access to existing capacity — before expanding into direct site development once they had revenue, credibility, and a track record to support larger capital commitments. "You don't have to start by building a data center from scratch. Find the smallest version of the real problem, prove you can solve it better than the status quo, and let that earn you the right to go bigger."

The Broader Signal for Founders and Investors

This conversation captured something GCN has seen repeatedly across recent events: some of the most durable opportunities in the current AI cycle sit below the application layer, in infrastructure that's harder to build, harder to explain in a two-minute pitch, and correspondingly less crowded with competition. For founders building in this space, the lesson is that the operational complexity — power contracts, siting, hardware logistics — that makes these businesses hard to start is precisely what makes them hard to copy once built. For investors, it's a reminder that the flashiest application-layer pitch isn't always where the most durable value accrues over a multi-year hold.

Founders building at the infrastructure layer of AI, energy, or deep tech are encouraged to apply to showcase at an upcoming GCN founder spotlight. Explore upcoming events to connect with investors actively deploying capital into this category.