May 09, 2026
OpenAI just put a very loud marker down in the Midwest. Construction has started on a Michigan “Stargate” data center in Saline tied to OpenAI and Oracle — a $16bn project expected to draw roughly 1.4GW. That’s not “a big campus”; it’s a grid-planning event, and it lands as US regulators sharpen their stance on who pays for the power.
The Big Stories
The OpenAI/Oracle Stargate build in Saline, Michigan is the day’s gravity well: $16bn for a single facility and ~1.4GW of load, under an initiative that targets $500bn across the US over four years. The immediate read-through is that “AI scale” has escaped the coastal usual suspects and is now comfortably shopping for massive electrical headroom in secondary (but industrially well-connected) markets. The next fight will be less about land and more about interconnection timelines, firm capacity, and how much of the bill gets socialised.
Florida is trying to make that last part explicit. Governor DeSantis signed a law requiring regulated utilities to design rate structures so large-scale data centers bear their full electric service costs. Florida Power & Light says its existing large-load rate already pushes customers to pay 100% of new generation needed, backed by engineering reviews, financial collateral, minimum monthly payments, and long-term commitments. This is what “welcome the investment” looks like when the grid is tight: you can build, but you don’t get to hide the upgrade tab inside everyone else’s bills.
Louisiana appears to be headed down the same path, but with the debate still live. Regulators are weighing rules to decide whether data centers or existing ratepayers should fund the grid infrastructure required for power-hungry projects. Taken together with Florida, the signal is clear: in the Southeast—one of the hottest landing zones for new builds—policy is moving from case-by-case negotiation to formal cost-allocation frameworks. Developers should expect more standardised (and less negotiable) “you break it, you buy it” structures.
On the energy procurement front, Meta keeps stacking long-dated supply. Meta signed a long-term PPA with EDP Renewables North America for the 250MW Cypress Knee Solar project in southeast Arkansas, EDP’s third deal with Meta, taking their total to 545MW. The project is framed as part of Meta’s broader energy actions supporting its $27bn Hyperion data center. The point isn’t that solar alone “powers” an AI build; it’s that the big platforms are locking in volumes early and using repeat counterparties—because queues, equipment lead times, and local politics don’t reward improvisation.
Reliability also got its own reminder. AWS had a thermal-event-triggered power outage in US-EAST-1 (use1-az4) that impaired EC2 and EBS, forcing traffic shifts and recovery steps including snapshot guidance and staggered service restoration. Analysts highlighted physical-layer separation and regional concentration risk for mission-critical workloads. The uncomfortable subtext: as compute concentrates into fewer, denser zones, the blast radius of “ordinary” facility incidents expands—especially when customers assume availability zones equal immunity.
Finally, capital is still trying to monetise the sector’s growth narrative. Blackstone Digital Infrastructure Trust filed to list on the NYSE under “BXDC,” targeting a $1.75bn IPO with an expected listing date of 14 May 2026 and a target annual yield of 5.75%–7%. For investors, the filing is a temperature check: even with power constraints and interconnection friction, the market appetite is strong enough to float a yield story—so long as assets can demonstrate durable cash generation and not just headline megawatts.
Behind the Headlines
On-site power is moving from “edge case” to default workaround in constrained markets. Unison Energy is scaling on-site natural gas CHP and microgrid systems specifically to serve data centers stuck behind long interconnection waits, citing pipelines of hundreds of megawatts to gigawatts and phase-one deployments typically 50–100MW with roughly two-year contract-to-commissioning history. The investor implication is that a meaningful slice of near-term capacity growth may be gated less by server supply and more by who can industrialise behind-the-meter generation—permitting, gas contracting, emissions strategy, and all.
The US grid equipment bottleneck is getting treated like an industrial policy problem, not a utility procurement problem. RMI argues DOE and Congress should use newly appropriated $375m and Defense Production Act authorities to strengthen domestic manufacturing for grid gear, citing persistent shortfalls (domestic large power transformer production at ~20% of demand) and >$30bn of imports in 2024. They point to big commitments like Hitachi Energy’s $1bn investment and Siemens Energy’s planned US capacity by 2027, but the thrust is that “more load” is now colliding with “not enough hardware.” If you’re underwriting aggressive energisation dates for AI campuses, transformer and switchgear reality is becoming as important as land and tax incentives.
And storage keeps scaling fast enough to matter at system level, not just project level. BloombergNEF forecasts global non-pumped hydro energy storage rising from 112GW/307GWh in 2025 to 158GW/459GWh in 2026, with cumulative capacity reaching 2.9TW/10.5TWh by 2036; China and the US made up ~70% of 2025 deployments. For data centers, this is a double-edged trend: more batteries can help grids absorb new load, but the same supply chain (cells, power conversion, interconnect) is increasingly contested by utilities and IPPs. The practical takeaway is that “battery-backed” claims will get harder to execute cheaply unless buyers secure equipment slots early and align with local grid needs.