PJM Data Center Study → Question 4

When Do Data Centers Leave the Grid?

Self-generation is already cheaper than grid power. Gas CCGT at $65/MWh is 36% below PJM's $102/MWh retail rate. But the cheap option is dirty, and the clean options barely pencil out. The regulatory structure determines whether departure helps or hurts everyone else.

Question 4

Self-Generation Cost Comparison

Q3 calculated what grid power costs with DC infrastructure investment. Q4 compares that to the cost of data centers generating their own power — using real transaction data from SEC filings, utility contracts, and industry benchmarks.

Levelized Cost of Self-Generation vs. PJM Grid Rate ($102/MWh)
$65
Gas CCGT (/MWh)
$80
Hybrid or Gas CT
$95
Nuclear PPA / Solar+Storage
$135
SMR (New Build)
Finding
Every option except SMRs already beats grid power. But the margins tell the real story: gas saves 36%, nuclear PPAs save only 7%. The cheap option doesn't meet carbon commitments. The clean options barely pencil out.

Sources: GridLab 2025 Gas Turbine Cost Report, Amazon-Talen SEC 8-K filing (~$85-95/MWh implied), Lazard LCOE v18, Meta El Paso $473M/366MW actual, Level Ten Energy PJM PPA data Q4 2024.

Grid Impact

Grid-Connected vs. Behind-the-Meter

The economics of self-generation are clear. But the regulatory structure determines whether DC departure improves or degrades grid reliability for everyone else. We modeled four self-generation scenarios at 30 GW of data center load.

Reliability Impact at 30 GW DC — by Self-Generation Structure
Scenario Hours Unserved Net Grid Effect
No self-gen (DC on grid only) 109 hrs +30 GW load, +0 GW gen
Grid-connected PPA (new gen) 0 hrs +30 GW load, +30 GW gen
BTM with new generation 0 hrs 0 net (both off-grid)
BTM taking existing nuclear 105 hrs −5 GW nuclear from grid

The regulatory structure matters more than the economics. Grid-connected PPAs (like the restructured Amazon-Talen deal) add new generation to the grid — everyone benefits. Behind-the-meter deals that take existing nuclear off the grid (like the original co-location attempt FERC denied twice) remove firm generation — everyone else pays the price. Same deal structure, opposite outcomes.

Hours Unserved by Scenario Across DC Load Levels

Model: hourly dispatch simulation with PJM capacity mix. Grid-connected PPAs add generation to system; BTM new gen removes both load and gen; BTM existing nuclear removes firm capacity without removing load.

Death Spiral

Do All Data Centers Eventually Leave?

We modeled the feedback loop: if grid rates exceed the cost of self-generation plus a reliability premium, rational DCs leave. Fixed costs spread over fewer customers. Rates rise. More leave.

Death Spiral Equilibrium — Starting DC Load vs. Final Grid-Connected DC
Finding
All data centers leave at every scenario tested. The equilibrium is zero DCs on grid — not because rates spiral catastrophically, but because self-generation was already cheaper before the spiral started. The "tipping point" question is moot. The economics already tipped.

But it's not catastrophic for residential rates. PJM's fixed costs spread across 815 TWh of non-DC load. Even if all DCs leave, residential rates rise modestly — a few percent from stranded infrastructure. The real cost is stranded infrastructure: CCGT plants built to serve data centers that become underutilized when those customers self-generate.

Model: iterative equilibrium — DCs depart when grid rate exceeds self-gen threshold ($65/MWh + reliability premium). Fixed costs redistributed each iteration until stable.

Reality Check

Why Are DCs Still on the Grid?

If self-generation is already cheaper, why haven't data centers left? The economics favor departure, but the barriers are real — especially for clean options.

  • — The cheap option is dirty. Gas CCGT at $65/MWh saves 36%, but Amazon, Microsoft, Google, and Meta all have carbon-neutral commitments. They can't build gas plants without reputational damage.
  • — The clean options barely pencil out. Nuclear PPAs at $95/MWh save only 7% — not enough margin to justify the complexity, especially with 2% annual escalators (Amazon-Talen deal).
  • — Interconnection and permitting take years. FERC denied the original Amazon-Talen co-location deal twice before it was restructured as a grid-connected PPA.
  • — Capital allocation. Meta's El Paso gas project cost $473M for 366 MW. That's capital diverted from core operations.
  • — Grid connection provides backup insurance. Dominion's new GS-5 rate class (25+ MW customers, 14-year contracts) is designed to keep DCs connected with standby power guarantees.
  • — FERC rules are still being written. The regulatory framework for large-scale co-location and behind-the-meter generation is actively being contested.

Gas self-gen is easy but dirty. Clean self-gen is expensive but mandatory for corporate commitments. That tension keeps data centers on the grid — for now. When clean options get cheaper, they'll leave.

← Prev Consumer Rates Study Back to PJM Study