Grid operators now take years to connect a large new load, and a frontier GPU generation stays commercially competitive for roughly two or three of them. Line those two clocks up and you can see why AI operators stopped waiting. A facility that sits in an interconnection queue until the end of the decade will energize hardware that is already behind. So the biggest compute builders have started acting like utilities. They put gas turbines on site, sign direct deals with nuclear plants, install fuel cells, and finance generation that has nothing to do with the public grid.
For engineers and executives who buy compute and don't build data centers, this is easy to file under "someone else's problem." That's a mistake. How your provider gets its power now feeds into what you pay per GPU-hour, whether the capacity in your roadmap will exist on time, and whether your rate holds if a regulator or permit board unwinds the arrangement underneath it.
Why the Grid Lost the Race to the GPU
The research behind this shift is consistent across sources. Grid capacity has replaced chips as the main limit on AI infrastructure. EnkiAI's analysis of grid limits describes operators moving from grid dependence to direct investment in dedicated generation, because public grids can't deliver enough reliable power on the timeline AI demand requires. datacenterHawk calls behind-the-meter solutions the industry's "new reality," not a niche workaround.
The mechanics are simple. A new multi-hundred-megawatt load usually needs transmission upgrades, substation work, and studies of how it affects everyone else on the system. Each step has its own queue. The EIA's electricity data shows commercial-sector demand rising after years of near-flat national load growth, and utilities that planned for flat demand don't have spare transmission waiting.
Grid planners and GPU buyers also run on different clocks:
| Clock | Typical horizon | Who controls it |
|---|---|---|
| GPU generation commercial relevance | A few years | Chip vendors |
| Data center shell construction | Roughly 1-2 years | Operator |
| On-site generation deployment | Months to a couple of years, permit-dependent | Operator and local regulators |
| Transmission upgrade and large-load interconnection | Often many years | Utility, RTO, state and federal regulators |
Horizons are illustrative ranges and vary heavily by region and project.
An operator can control the middle two rows. The last row is out of its hands. When the last row is longer than the first, the economics push operators to build around it. We covered the capacity-planning side in AI Data Center Power Constraints. This post covers what operators are doing about it and what that means for the people renting their capacity.
Four Ways Operators Are Buying Power
"Behind the meter" gets used loosely. The structures carry very different risks, so name them precisely before you assess anything.
| Structure | How it works | Speed | Main risk |
|---|---|---|---|
| On-site thermal generation | Gas turbines or reciprocating engines at the campus, sometimes as bridge power until the grid arrives | Fastest, if permits and fuel are available | Air permits, gas pipeline capacity, local opposition, fuel price exposure |
| Fuel cells | Solid-oxide or similar units, usually gas-fed, deployed in modules | Fast, modular | Higher cost per MWh, fuel supply, vendor concentration |
| Co-location at an existing plant | Data center connects directly to an existing generator, often nuclear, and takes output before it reaches the grid | Medium | Regulatory challenge over grid cost allocation and reliability |
| Grid-delivered PPA plus firming | Contract for generation elsewhere, delivered over the grid, with storage or backup | Slowest, needs interconnection | Interconnection delay, transmission charges, basis risk |
A lot of 2026 campuses mix several of these. A common pattern is on-site gas as bridge power, a PPA for long-term supply, and an agreement to convert to grid service once transmission is built. As this widely shared industry post on behind-the-meter traction notes, the trend is moving from experiment to default plan for large AI builds.
The structures that come online fastest also tend to be the most legally fragile. On-site gas depends on air permits that local communities can contest. Co-location depends on regulators accepting that a giant load attached to a generator doesn't owe much for the grid it still relies on for backup and stability. The grid-delivered PPA is the slowest but has the most settled legal footing.
The Regulatory Fight Over Co-Located Load
Co-location is where operator strategy and federal regulation collide. The argument is easy to state and hard to settle. A data center sitting next to a nuclear plant takes power that used to flow onto the grid. Other ratepayers lose that supply. The data center still benefits from the grid when the plant trips or goes down for refueling, and it may avoid paying transmission charges for that benefit.
The best-known test was the Talen Energy and Amazon arrangement at the Susquehanna nuclear plant in Pennsylvania. FERC rejected an expanded interconnection agreement for the co-located load in late 2024. The parties later restructured the deal as front-of-the-meter supply delivered through the grid. That sequence is worth remembering. A behind-the-meter structure built for speed and cost turned into a grid-connected structure with a different cost basis, and the regulator forced the change.
Regulatory pressure has grown since then. Reporting on FERC's 2026 grid order describes grid operators getting 60 days to address how they handle AI data center power. That's a short window by regulatory standards, and it shows the commission wants rules in place before the next wave of campuses connects. Docket activity and orders appear on FERC's news and events page. If your provider operates in an RTO with an open co-location proceeding, someone on your team should be reading it.
Three outcomes matter for compute buyers:
- Co-located loads pay new transmission or ancillary charges. The operator's cost per MWh rises, and the increase reaches you through pass-through clauses or at renewal.
- Co-location is restricted or capped. Some planned capacity disappears or gets delayed until a grid connection exists.
- Clear rules get approved. Uncertainty falls and co-located capacity becomes easier to finance and price.
None of these outcomes is guaranteed, and each one hits a buyer differently depending on how the provider is powered. That's why power structure belongs in compute procurement.
Power Structure Is Now an Input to Compute Pricing
The Carnegie Endowment's compute coalition report treats energy access as a strategic input to AI capacity at the national level. The same logic applies to a single compute contract. When you sign for reserved GPU capacity, you're also taking on the provider's power position, whether or not it's written into the contract.
A useful way to see this is to split any provider's megawatts into three buckets:
- Energized. Power is flowing today and the hardware is running. This is real capacity.
- Contracted. A signed interconnection agreement, PPA, or equipment and fuel contract exists, with permits in hand. It's probably coming, on a schedule that can slip.
- Speculative. It's in the plan, on the investor deck, or in a queue with no signed agreement. Treat it as an option and don't count it as capacity.
Many 2026 capex plans assume power the grid hasn't committed to. If a provider's growth story leans on the third bucket, your expansion clause is a bet on that bucket turning into the first.
Hypothetical example. A mid-size software company signs a three-year reserved-capacity contract with a GPU cloud. The contract includes an option to double capacity in year two. The provider's year-two capacity depends on a campus powered by on-site gas turbines as bridge power, with a grid connection expected in year three. In year two the air permit gets challenged and the turbine deployment slips. The expansion option is still in the contract, but the capacity behind it doesn't exist. The buyer's product roadmap, which assumed doubled inference throughput, slips with it. Nobody breached the contract. The risk just sat in a layer the buyer never asked about.
Pricing has a similar problem. Say a provider's low rate depends on co-located power that avoids transmission charges, and a FERC ruling adds those charges. The provider's cost goes up. Whether yours does depends on contract language most buyers never negotiate. For more on how reserved and on-demand GPU contracts are built, see Rent or Own: GPU-as-a-Service Pricing and Contract Economics.
Physical supply chains add another layer. Turbines, transformers, and switchgear all have long lead times of their own, which we covered in AI's Real Bottleneck Isn't Software. It's Materials. A provider promising fast on-site generation should be able to name its turbine and transformer orders and show delivery dates.
A Diligence Scorecard for Long-Term Capacity
Here's the scorecard we'd use before signing any multi-year compute commitment above a trivial size. Each question has a specific answer that gives the buyer protection.
| Question | Good answer | Red flag |
|---|---|---|
| What share of the capacity serving us is energized today? | A specific number for named sites | "Our total pipeline is X GW" |
| How is each site powered? | Named structure per site: grid, PPA, co-located, on-site thermal, fuel cells | "A diverse energy mix" |
| Which sites depend on co-location or behind-the-meter arrangements under regulatory review? | List with docket references | No awareness of the proceedings |
| What happens to our rate if a power arrangement is restructured? | Capped pass-through or fixed rate for the term | Uncapped "change in law" or energy cost pass-through |
| Can you curtail or relocate our workloads during power constraints? | Defined curtailment terms, notice periods, credits | Silent, or broad force majeure |
| Are expansion options tied to energized capacity? | Options backed by named, contracted sites | Options backed by "future capacity" |
| What is the fuel supply arrangement for on-site generation? | Firm pipeline capacity contracts | Interruptible gas service |
The curtailment row is easy to miss. Some operators are trading interruptibility for faster interconnection and agreeing to drop load during grid stress. That can be a reasonable deal for batch training jobs. It's a bad one for latency-sensitive inference. We covered the engineering side in Flexible Load Data Center Interconnection. A contract that lets the provider curtail without telling you which workloads get shed first moves the provider's grid deal onto your production traffic.
Contract language to ask for:
- Power-source disclosure by site, updated each year or whenever the structure changes.
- A repricing limit that caps how much energy or regulatory cost the provider can pass through during the term.
- Workload classes for curtailment, so training and batch jobs absorb interruptions before production inference does.
- Expansion options tied to named sites with energization milestones, and exit rights if those milestones slip.
- Substitute capacity commitments that name the region and latency profile, so a "replacement" in a distant region doesn't count as performance.
None of this is exotic. Large enterprises already negotiate these clauses for colocation. AI compute contracts have mostly been signed in a seller's market where buyers took what was offered. That's starting to change as more providers compete. Buyers who ask these questions now will get better terms than buyers who wait for an outage or a repricing notice.
Planning AI Workloads Around Power Risk
Most companies reading this won't sign a PPA or site a turbine. They'll run agents and inference workloads on capacity someone else powers. Power risk still affects how you should design those workloads.
Separate workloads by how much power interruption they can tolerate. Batch document processing, evaluation runs, and embedding backfills can move, pause, or run off-peak. Customer-facing agents can't. If your architecture treats them the same way, you'll pay premium rates for flexible work or carry interruption risk on critical work. A routing layer that sends batch jobs to cheaper, interruptible capacity and keeps production on firm capacity is a real cost lever.
Keep a second provider you can reach. Multi-provider routing is usually sold as a hedge against model quality or pricing. It's also a hedge against one provider's power problems. If one region's capacity slips because a permit fight delayed bridge power, a tested failover path decides whether you have a degraded week or an outage.
Plan capacity with ranges. If your roadmap assumes a specific throughput increase on a specific date, find out which of the provider's three megawatt buckets that increase comes from. Plan against energized and contracted capacity. Treat anything speculative as upside.
This is where implementation work matters. OpenNash builds production agent systems, and part of that work is designing the routing, fallback, and workload classification that keep agents running when a provider has a capacity problem. In an audit we map which workflows can tolerate delay and which can't. In design we put that classification into routing rules and approval gates. The client owns the resulting system, including the provider abstraction, so switching capacity sources doesn't mean a rewrite.
If you're about to sign a multi-year compute commitment or scale an agent workload that depends on one provider's expansion plans, OpenNash can review your capacity assumptions and failover design with you before the contract is signed. The questions in the scorecard above are the right place to start, whoever you work with.
The decision to make this quarter is concrete. Take your largest compute commitment, ask the provider for a per-site breakdown of energized, contracted, and speculative capacity, and read the change-in-law and pass-through clauses against the FERC proceedings in that provider's regions. If the provider can't answer, that tells you how to price the renewal.