New York Froze 50MW+ Data Centers. The Math Shows Self-Hosted PaaS Was Never in That Weight Class

On July 14, 2026, Governor Kathy Hochul signed Executive Order No. 62, and New York became the first US state to freeze new hyperscale data center construction outright — not tax it, not zone around it, just stop issuing the permits. The threshold is 50 megawatts: any qualifying facility above that line can't get a discretionary state permit for up to a year. Every "what does this mean for the cloud" take that followed skipped the one number that actually answers the question: how big is a data center, in megawatts, compared to the ones this order targets?

The megawatt math nobody ran​

Here's that comparison, pulled from public facility specs rather than press-release framing:

The largest self-hosted-friendly colocation campuses in Europe — the kind of facility a dedicated-server provider like Hetzner, OVHcloud, or Scaleway runs, and the kind of facility a self-hosted PaaS fleet actually rents rack space in — top out somewhere between 10MW and 37MW. Every one of them sits under New York's enacted 50MW line. The hyperscale AI campuses the order was written for operate two orders of magnitude higher: Ohio's Prometheus supercluster and Texas's Stargate I both measure capacity in gigawatts, not megawatts, because they're built to power tens of thousands of GPUs at once.

That gap is the actual story. Self-hosted infrastructure doesn't dodge New York's moratorium through some clever structuring — it was never being built at the scale the moratorium was written to stop in the first place.

What Executive Order 62 actually does​

The order is narrower and more procedural than the "data center ban" headlines suggest, and the mechanics matter for anyone trying to reason about what it changes:

• Scope: it pauses the Department of Environmental Conservation's discretionary permits for new data center facilities capable of drawing 50MW or more. Projects already deemed complete before the order keep moving.
• Carve-outs: facilities used primarily for manufacturing, medical care, education, or specified research are excluded, regardless of power draw.
• Duration: up to one year, while the state drafts a Generic Environmental Impact Statement covering energy demand, water use, and air quality — the standardized review the state says didn't exist before.
• The threshold gap: New York's own legislature had already passed the Responsible Data Center Development Act in June 2026, five weeks earlier, setting the bar at 20MW — a much lower ceiling that would have caught mid-sized facilities too. Hochul's executive order raised that line to 50MW, a choice that, whether intentional or not, leaves the entire size class of facility a self-hosted fleet actually uses outside the freeze.

Hochul's stated rationale was direct: "data center development threatens to hike up utility bills, deplete our natural resources, and create uncertainty for New Yorkers." NYISO's own 2026 Power Trends report backs the concern with numbers — reliability margins are tightening as older plants retire faster than new generation gets added, and data center demand is a growing, harder-to-predict share of the load.

This isn't a New York problem — it's the 2026 pattern​

New York moved first, but it isn't moving alone. More than 300 data center bills were filed across 30-plus states in the first six weeks of 2026 alone, and the ratepayer math behind them is concrete, not speculative:

• In the PJM grid region spanning seven states, ratepayers were on the hook for $4.3 billion in electricity infrastructure projects approved in 2024 specifically to connect data centers to the grid.
• Ohio's data center tax exemption cost the state $1.6 billion in 2025 — about 12x what was originally projected — prompting Governor DeWine to pause new applications even as the state's own moratorium ballot measure (at a stricter 25MW threshold) failed to gather enough signatures to qualify.
• Virginia's HB 1515, a similar moratorium bill, got carried over to 2027 rather than dying outright.
• Industry-wide, 30-50% of the data center capacity that was supposed to come online globally in 2026 is now expected to be delayed or canceled, and the reason isn't capital — it's structural. Interconnection queues for new large loads now stretch five to seven years, against a one-to-three-year construction cycle for the buildings themselves. High-voltage transformer lead times have gone from 12-18 months to as long as 36-48 months, and community opposition and Chinese tariff impacts on electrical components are compounding both.
• More than 20 states now have some form of data center moratorium, ratepayer-protection, or special large-load rate class bill in play, according to tracking from Good Jobs First and MultiState — a shift the same trackers describe as moving state policy from "compete for the project with tax breaks" to "regulate the project's grid and ratepayer footprint," a reversal from most of the previous decade's incentive-driven posture.

None of this capacity is disappearing — it's relocating. Texas overtook Northern Virginia as the world's top data center market in May 2026, and Synergy Research projects inland states will host 53% of new hyperscale capacity going forward, chasing the grid headroom that Virginia (where data centers already draw 26% of the state's electricity) and the Pacific Northwest have run out of. The hyperscale build-out isn't stopping. It's moving to wherever a gigawatt-scale interconnection queue is still short enough to clear.

What this actually costs a hosted-PaaS vendor's roadmap​

Every cloud region a hosted PaaS runs on top of — Vercel, Render, Railway, Fly.io, and the hyperscalers underneath them — depends on that same interconnection queue clearing on schedule. When a state freezes permits or a queue stretches to five years, that risk doesn't stay contained to the hyperscaler; it propagates downstream to every platform whose next-region capacity plan assumed the grid would keep pace with demand. A PaaS vendor that priced its 2027 expansion around "AWS opens a new East Coast availability zone" now has to price in the chance that it doesn't, on the timeline it assumed — and that repricing shows up as either a capacity waitlist, a price increase to ration the region that did ship, or a quiet redirect of new customers to whichever region still has headroom, regardless of where those customers actually wanted their app to run.

That's not a hypothetical mechanism — it's exactly what's already happening one layer up the stack. Azure's internal capacity forecasts extend constraints into the first half of 2026 for specific regions, and hyperscalers are increasingly building on-site or hybrid power rather than waiting on grid interconnection at all, with off-grid and hybrid approaches now accounting for nearly half of newly announced AI capacity despite representing under 10% of total projects. A PaaS built on top of a hyperscaler region inherits whatever solution that hyperscaler picks — grid queue, private power plant, or relocation to Texas — with zero visibility into which one it'll be, or when.

A self-hosted fleet running on already-built colocation capacity sidesteps that specific risk, because it isn't asking a state to permit a new 50MW+ campus at all. Renting the next rack in an already-operating, already-permitted Hetzner or OVHcloud facility is a capacity-planning problem, not a regulatory one — the facility cleared its interconnection queue years ago, under a different demand curve entirely, and adding ten more dedicated servers to it doesn't trigger a new discretionary permit, a new Generic Environmental Impact Statement, or a new multi-year queue position. The operator planning that expansion is choosing between existing providers and existing regions today, not betting on a permit clearing on schedule two years from now.

The honest caveat: this is permitting risk, not price risk​

It would overstate the case to call self-hosted infrastructure immune to grid pressure generally — it isn't, and the distinction matters. Hetzner raised prices three separate times in 2026, in part passing through the same underlying cost pressure (rising electricity and memory prices tied to AI-driven demand) that's straining the grid everywhere else. What a self-hosted fleet on existing capacity avoids specifically is the construction and permitting risk: the multi-year interconnection queue, the discretionary-permit freeze, the year-long environmental review New York just imposed on anything crossing 50MW. Existing racks in an existing, already-connected facility don't need any of that to keep running. New ones being built from scratch do — and that's exactly the queue self-hosted infrastructure at this scale was never standing in.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on Hetzner and OVHcloud hardware you already have, not a new hyperscale campus waiting on a permit. See the project on GitHub.