The number that should be pinned to every energy investor's wall right now: 11.2×.
That's how much PJM's capacity auction price jumped from the 2024/2025 delivery year to 2028/2029 — from $28.92/MW-day to $325.00/MW-day. Not a gradual climb. Not a policy-driven nudge. A structural repricing of what it costs to keep the lights on across 13 states and Washington, D.C. And according to PJM's Independent Market Monitor, roughly 63% of total 2025/26 capacity costs — $9.33 billion of $14.69 billion — trace directly to data center load.
This is what civilizational-scale demand looks like when it hits a grid that wasn't designed for it. The old electricity market structure — built around predictable residential load curves, dispatchable baseload, and incremental industrial growth — is being stress-tested by a customer class that wants gigawatts, wants them now, and doesn't particularly want to be interruptible. The result is a fundamental renegotiation of how electricity markets price reliability, who bears the cost of new infrastructure, and whether demand response — the grid's traditional safety valve — can scale to meet a load profile that looks nothing like anything it was designed to handle.
The Old Demand Response Model Was Built for a Different Enemy
Demand response, in its classic form, was a clever hack for a specific problem: peak demand events that lasted hours, driven by air conditioning loads on hot summer afternoons. Utilities would pay large industrial customers — aluminum smelters, paper mills, cold storage facilities — to curtail consumption during those windows. Grid operators would aggregate these commitments into a virtual capacity resource. Everyone won: the utility avoided building a peaker plant that ran 40 hours a year, the industrial customer got paid for flexibility it already had, and residential customers avoided the worst of peak pricing.
The Congressional Research Service's September 2026 FAQ on data centers and the grid frames the emerging tension precisely: data centers have historically secured power from local utilities using existing transmission and distribution infrastructure, but the anticipated pace of U.S. electricity demand growth now "exceeds that of the last 20 years," raising questions about whether existing regulatory requirements are suited for this scale of development. That's the polite version. The impolite version is that the demand response toolkit was designed for loads that flex. AI inference clusters, by contrast, are designed to run flat-out.
A hyperscale data center running large language model inference doesn't have idle capacity to offer back to the grid on a Tuesday afternoon. Its load profile is determined by user query volume, which correlates with business hours and peaks when everyone is using AI assistants simultaneously. The flexibility that made industrial demand response work — the ability to defer a production run, delay a refrigeration cycle, pause a smelting process — doesn't map cleanly onto a facility whose entire value proposition is low-latency response to user requests.
This isn't a reason to abandon demand response as a concept. It's a reason to rebuild it from scratch for a fundamentally different load type.
PJM's IRAS Proposal Is the Market's First Real Answer
On August 13, 2026, PJM filed its Interim Resource Adequacy Service (IRAS) proposal with FERC — and it represents the most consequential rethinking of large-load market structure in a generation. Power Advisory's analysis of the filing lays out the core logic: PJM attributes 30 of the 32 gigawatts of forecast demand growth between 2024 and 2030 to data centers. The grid operator's response is to force a binary choice on new large loads — either bring your own capacity, or accept curtailment when supply gets tight.
The 50 MW threshold matters. Any end-use customer with combined peak demand of at least 50 MW at a single site gets classified as a Large Load, entered into a new registry, and required to either secure new generation capacity equal to or greater than their peak demand, or face curtailment before PJM deploys its standard Pre-Emergency Load Management resources. In plain terms: if you want to plug in a 200 MW data center and you haven't arranged for 200 MW of new generation to back it up, you go to the back of the reliability line.
This is a market mechanism, not a prohibition. PJM is not saying no to data centers — the board chair explicitly framed the question as "not a yes-or-no question" about whether to connect, but how to connect while preserving reliability. The "bring your own generation" option creates a direct financial incentive for data center operators to vertically integrate into power supply — signing long-term PPAs, co-locating generation, or investing directly in new capacity. That's not a burden. That's the market correctly pricing the externality that large loads have been imposing on everyone else.
The contrast with Ontario's approach is instructive. Power Advisory notes that Ontario's proposed Data Centre Playbook would have the government conduct strategic assessments to determine which data centers represent sufficiently valuable uses of scarce grid capacity — essentially ministerial project selection. PJM, operating under federal jurisdiction across 13 states, can't replicate that model. Instead it's using price signals and curtailment risk to achieve similar rationing outcomes through market mechanisms. Both approaches are trying to solve the same problem. The American version is messier and more litigious. It's also more likely to actually get built.
FERC's June 2026 Order Changed the Cost-Assignment Rules
The IRAS filing didn't happen in a vacuum. It followed FERC's June 18, 2026 show-cause orders to all six Regional Transmission Organizations and Independent System Operators — PJM, MISO, SPP, CAISO, ISO-NE, and NYISO — directing them to justify their current rules for large-load interconnection and tariffs or propose reforms. The FERC order, as analyzed by Abacus News citing Data Center Dynamics, set two response windows: 30 days to report on generation adequacy for existing and new large-load demand, and 60 days to justify tariffs or submit changes.
What FERC is actually doing here is forcing the grid's cost-assignment logic into the open. For decades, transmission upgrades triggered by new large loads have been socialized across the entire customer base — everyone's rates go up a little when a new industrial customer requires a substation upgrade. That made sense when large loads were rare and the upgrades were modest. It makes much less sense when PJM's Independent Market Monitor calculates that data center load alone drove a 174.3% increase in 2025/26 capacity auction revenues. At that scale, socialization becomes a subsidy — residential ratepayers in New Jersey and Maryland are effectively financing the grid buildout for hyperscale AI infrastructure.
FERC's order asks grid operators to address five categories of reform, including how firm large-load commitments must be, whether demand flexibility can be required as a condition of service, and who pays for network upgrades when a project changes or fails to materialize. The Abacus News analysis makes the key point: this doesn't make an inefficient AI cluster illegal, but it does change what questions a proposed facility must answer before receiving service. If interconnection rules attach more cost to peak demand, developers have a direct incentive to reduce stranded capacity and improve power scheduling. Indirect effects, but real ones.
Microsoft's 26 GW Announcement Tells You Why This Repricing Is Just Getting Started
Here's the civilizational scope of what we're actually talking about. Bloomberg reported on September 10, 2026 that Microsoft plans to more than triple its data center capacity, targeting over 38 gigawatts by 2032 — up from roughly 12 gigawatts today. Bloomberg notes that 38 GW would eclipse New York State's peak electricity consumption. One company. One buildout. More power than an entire major American state uses at its highest demand moment.
Microsoft is not alone. The DOE's resource page on clean energy for data centers reflects the federal government's recognition that this is now a national infrastructure priority, not a commercial real estate question. The CRS FAQ notes that Congress's 119th session has held hearings specifically on ensuring data centers don't raise costs or affect grid reliability for residential consumers — which tells you that the political pressure is already building on both sides of the equation.
The demand response market that emerges from this transition will look nothing like the one that existed five years ago. The old model was about paying existing loads to curtail occasionally. The new model is about structuring the terms under which enormous new loads get access to the grid at all — and making sure the cost of that access lands on the parties creating the demand, not on the households who were there first.
The Profitable Path Runs Through Flexibility Infrastructure
For investors and builders watching this space, the market signal is clear: the entities that figure out how to offer genuine flexibility from AI data center loads will capture enormous value. The capacity auction price collar that PJM filed with FERC in February 2026 — set with a cap of roughly $325/MW-day and a floor of roughly $175/MW-day — creates a predictable band within which flexibility has a calculable dollar value. A data center that can credibly offer 50 MW of demand reduction for four hours during a capacity shortage event is worth something specific and contractable in that market. That's not a conservation argument. That's an arbitrage opportunity.
The technology stack for this doesn't fully exist yet. Shifting AI inference workloads across geographic regions in response to real-time grid pricing requires coordination between grid operators, cloud schedulers, and network routing systems that currently operate on completely different timescales and with different optimization objectives. But the financial incentive to build that coordination layer is now enormous and growing. Australia's Bloomberg Economics analysis flagged the same dynamic from a different angle: the data center boom risks pushing capital expenditure above 2% of GDP in 2026-27, intensifying competition for construction capacity and drawing skilled trades away from renewable energy and housing. The grid stress is global, not just a PJM problem.
The civilizational bet here is that we build through this constraint rather than around it. The capacity auction repricing, the FERC show-cause orders, the IRAS filing — these are the market's immune response to a demand shock it wasn't designed for. They're painful and they're slow and they're going to generate
