Cudo Compute finds infrastructure 71%
- CUDO Compute’s new “Land. Power. Compute.” report argues AI build-outs are now constrained less by GPUs than by land, power, cooling, and permits. - The headline number is 71%: respondents said infrastructure now drives most AI deployment cost, with power, networking, and energy inputs dominating decisions. - That matters because AI capex is still surging, but grid waits, site bottlenecks, and geopolitical risk are pushing compute strategy regional.
AI infrastructure used to sound like a chip story. Buy more GPUs, wire them up, scale the model. But the bottleneck has shifted. CUDO Compute’s new “Land. Power. Compute.” report says the hard part now is physical infrastructure — land, power, cooling, networking, and permits — and that is starting to reshape where AI gets built and who can build it at all. ### Why is infrastructure suddenly the story? Because AI data centers are no longer normal data centers with extra accelerators bolted on. They are power-dense industrial sites. CUDO’s report says infrastructure now drives most of the cost in AI deployment, and it calls out power, networking, and other energy-related inputs as the main pressure points. That changes the conversation from “who has chips?” to “who has a site that can actually run them?” (cudocompute.com) ### What does the 71% number really mean? It means builders are spending more of the AI budget on the physical stack around the compute than many people expected. Not just the servers, but the grid connection, substation work, cooling plant, fiber, switching, and the long lead-time hardware needed to make a cluster usable. The chip still matters. But the catch is that a GPU without enough power and networking is basically stranded inventory. (cudocompute.com) ### Why is power the hardest constraint? Because power moves on utility timelines, not software timelines. GPU platforms refresh roughly every year. Grid upgrades, interconnection, and permitting can take years. CUDO frames that mismatch as the real problem — all the pieces have to line up at once, and they rarely do. Deloitte’s 2025 survey lands in the same place: grid stress was the leading challenge for data-center development, and some interconnection requests face waits of about seven years. (cudocompute.com) ### Why does location matter more now? Because site selection is drifting away from “close to users” and toward “close to available power.” CUDO says regions with spare grid capacity and developable land are attracting investment, while established hubs are hitting congestion and delay. That is a big shift. It means AI geography starts to look less like classic cloud geography and more like heavy industry — you go where the inputs are. (cudocompute.com) ### Where does geopolitics come in? Once compute becomes critical infrastructure, country risk matters more. The report says geopolitics is shaping where AI data centers locate, which fits the broader pattern: governments and companies are getting more sensitive about energy security, regulatory stability, and dependence on foreign supply chains. The practical result is regional strategy — build where the grid is reliable, the rules are predictable, and the politics are less likely to scramble a 10-year asset. (cudocompute.com) ### Is this just a CUDO sales pitch? Partly, yes — CUDO sells infrastructure services, so its framing is not neutral. But the broader market is moving in the same direction. McKinsey put the global capital needed for data centers at nearly $7 trillion by 2030, with $5.2 trillion tied to AI-ready facilities. Even outside CUDO’s survey, the industry is telling the same story: the next AI race is as much about substations and transmission as semiconductors. (cudocompute.com) ### What does this change for builders? It rewards companies that can deliver “turnkey” capacity — not just racks of GPUs, but land, power contracts, cooling, networking, and operations in one package. It also favors hyperscalers and large developers that can spread risk across regions and lock in power earlier. Smaller buyers may still get compute, but they will increasingly buy into someone else’s prebuilt stack rather than assemble it piece by piece. That is the real shift hiding inside the report. (mckinsey.com) ### So what’s the bottom line? AI is becoming a grid-and-real-estate business wearing a software label. The companies that win the next phase will not just have the best models or the most GPUs. They will have the land, the megawatts, and the patience to get physical infrastructure built before demand outruns the map. (cudocompute.com)
