May 01, 2026

Microsoft just put numbers around what everyone in this market is feeling: demand is outrunning delivery. In its latest update, Azure growth ran at 40% YoY and the AI business hit a $37bn annual run rate — but the real tell was analysts pointing to delivery windows stretching from roughly six months to 18+ months because power, cooling, transformers, utility timelines, and physical capacity are all tight (Microsoft’s AI cloud demand outpaces data center delivery). That gap between what’s contracted and what can actually be built is quickly becoming the industry’s central plotline.

The Big Stories

Microsoft’s results underline that the “AI boom” isn’t constrained by customers or budgets — it’s constrained by infrastructure. Commercial RPO hit $627bn, a giant backlog of contracted demand, while build-and-energize timelines are being dragged out by the unglamorous bottlenecks: transformer and utility availability, liquid-cooling readiness, and available data center capacity. The competitive takeaway is simple: this isn’t just a race for GPUs; it’s a race for megawatts, water strategy, and supply chain execution.

Google took a very utility-forward step in Oklahoma, signing a long-term energy agreement with Oklahoma Gas and Electric to support its under-construction campuses in Muskogee and Stillwater — and committing to pay for associated infrastructure over the long term (Google signs long-term energy deal with OG&E in Oklahoma). This is the “pay for the wires” era of hyperscale: not just buying clean energy, but underwriting the grid work required to make large-load campuses real. Google’s earlier solar PPAs in Stephens and Muskogee Counties — with capacity purchase agreements to OG&E — also reads like a deliberate attempt to turn project-level renewables into utility-usable regional supply.

AirTrunk is doubling down on Johor with scale that’s hard to ignore: MYR 12bn (about $3bn) for two new data centres (JHB3 and JHB4) with a combined IT load over 280MW, pushing its Johor footprint to more than 700MW across four campuses (AirTrunk to invest MYR 12 billion in two Malaysian data centres). It’s also making the “AI-ready” design choices explicit — high-density support, energy efficiency, and 100% recycled-water cooling — and tying the buildout to local procurement (MYR 423m awarded so far, with MYR 5bn expected on completion). The message to customers (and regulators) is that this isn’t speculative: it’s industrial-scale delivery with a supply-chain plan.

In Europe, Pantheon Atlas went for the headline number: a planned 1GW hyperscaler campus in Topusko, Croatia, targeting commercial operations in Q1 2029 with an initial investment around €12bn (Pantheon AI plans 1GW renewable-powered data centre in Croatia). The energy concept is unusually concrete: 500MW of solar paired with 8,000MWh of BESS, with KKR-owned Greenvolt signing an LOI to build the behind-the-meter solar-plus-storage plant. The grid angle matters too — the project describes four 400kV transmission lines enabling 5.2GW of additional renewable integration, which is effectively positioning the campus as both a load and a grid-shaping node.

On the power pipeline itself, PJM’s new Cycle 1 interconnection intake shows just how crowded the runway is: 811 generation projects totaling about 220GW, led by 349 storage projects (67,465MW) and 157 natural gas projects (105,797MW) (PJM Cycle 1 draws 811 projects totaling ~220 GW capacity). The projects now go into validation ahead of a clustered study process that runs into 2028 — a reminder that “plenty of proposals” doesn’t equal “near-term electrons.” PJM also pulling in Tapestry/Google’s HyperQ AI tool for document review is a small but telling sign: even queue administration is now an AI-scale workflow problem.

Behind the Headlines

The industry’s operating model is shifting from “how big can we build?” to “how fast can we energize?” Developers are prioritizing “speed to power” with behind-the-meter generation, phased energization, and hybrid energy models to keep schedules from collapsing under grid constraints (Data centers prioritize ‘speed to power’ amid grid constraints). What’s new is the level of urgency: projections cited here range from energy consumption tripling by 2029 (IDC) to demand doubling by 2030 (IEA), which turns power-delivery timing into the gating factor for revenue, not a back-office detail. The mention of SMR pilots targeted as early as 2026 fits the same pattern — it’s less about “nuclear hype” and more about the market hunting for power sources with controllable timelines.

Oklo teaming up with Nvidia and Los Alamos National Laboratory is another signal that “AI infrastructure” is now being framed as an integrated power-and-compute product, not two separate procurement tracks (Oklo, Nvidia and LANL partner on nuclear-powered AI factories). The collaboration aims to accelerate plutonium-bearing fuel validation and design nuclear-powered AI factories, tying into Oklo’s Pluto and Aurora efforts and the DOE’s Genesis Mission (launched Nov 2025), with Oklo targeting commercial power by end-2027. The important subtext: hyperscale-era constraints are pushing AI leaders toward power R&D partnerships that would’ve looked exotic just a few years ago.

Local politics is starting to move from noise to actual project risk. Detroit City Council passed a resolution urging a two-year moratorium on data center permits while zoning and environmental rules are developed — with the decision now sitting with Mayor Mary Sheffield (Detroit residents push moratorium and zoning for data centers). The story also flags how quickly these fights spread: nearby Ypsilanti has a 12-month ban impacting a planned $1.2bn University of Michigan project, and DTE Energy has a $474m rate case tied to a Saline data center timeline. For investors, this is the reminder that “entitlements” and “social license” can become as schedule-critical as transformers — and that the next constraint after power may be permitting.