Share with your CTO
The AI capex wave is hitting a physical ceiling, and three companies are positioned at that bottleneck. ERock (EROC), a modular natural gas power systems builder with a $1.7 billion contracted backlog and production sold out through 2027, sits where compute demand outruns grid capacity. Everpure (P), at $31 billion market cap, just landed its first EXA-scale hyperscale storage customer. Jabil (JBL) is projecting AI-related revenue growth from $9 billion in fiscal 2025 to $22 billion by 2027. The AI infrastructure power crunch analysis frames all three as proxies for the physical build layer underneath AI deployments.
What this means for your business
Power is now the binding constraint on AI deployment timelines, and that changes the conversation CTOs need to have internally. If your organization is planning large-scale model training or inference infrastructure over the next two to three years, the sold-out production capacity at companies like ERock signals that on-site power generation has moved from contingency planning into critical-path procurement. Organizations that haven’t started those conversations are already behind the queue.
The Jabil numbers deserve a closer read than the article gives them. Going from $9 billion to $22 billion in AI-related revenue across two fiscal years is an extraordinary growth claim, and the question isn’t whether the revenue comes, it’s what margin profile comes with it. Contract manufacturing at hyperscale volumes is a low-margin business by design, and “higher value Intelligent Infrastructure work” is the kind of language that tends to appear before margin disappointments, not after them. For CTOs evaluating Jabil as a supply chain partner rather than an investment, the more relevant signal is whether that hyperscale concentration makes Jabil a more or less stable vendor as program mix shifts.
The Everpure hyperscale win is the most consequential data point in this piece for infrastructure planners, even though it gets the least analysis. A first EXA-scale customer win, storage at the scale where a single deployment measures in exabytes, reshapes unit economics across an entire product line in ways that flow down to pricing and availability for everyone else in the queue. If you’re currently negotiating all-flash storage contracts for AI workloads, the competitive dynamic just changed, and the direction it changed probably isn’t in your favor as a buyer.
Concept deep-dive: On-site generation
On-site generation means a facility produces its own electricity rather than drawing exclusively from the utility grid, think of it as a power plant bolted to your data center instead of a wall outlet. It exists because grid interconnection queues in the US now run five to ten years in many markets, far too slow for AI infrastructure timelines. The business consequence is that power procurement has become a capital expenditure decision sitting inside the CTO’s scope, not just a facilities cost.
Based on reporting from 3 US AI Infrastructure Stocks Tied To The Data Center Power Crunch, originally published 2026-11-07 13:28:00.
