Amazon's 7.
The scale problem nobody's talking about
Most coverage focuses on the "clean energy" commitments — Amazon's PPAs, their 2040 net-zero pledge, the renewable builds they're financing. But the physical reality of 7.65 GW continuous load changes the equation. That's not peak demand; that's baseload for training runs that don't stop. Solar and wind can't cover that profile without massive overbuild and storage that doesn't exist at this scale yet.
The PJM interconnection queue already shows the bottleneck: gas peakers and combined-cycle plants are the only resources that can come online fast enough to meet the 2027-2028 target dates Amazon's signaling. Renewables in the queue have 4-5 year lead times minimum. Nuclear? Forget it — the NRC review alone exceeds their timeline.
What 7.65 GW actually burns
Let's do the rough carbon accounting. A modern combined-cycle gas turbine runs ~50% efficient, emitting roughly 0.35 metric tons CO₂ per MWh. At 90% capacity factor (realistic for AI training clusters that run 24/7):
Annual generation: 7.65 GW × 8,760 hrs × 0.9 ≈ 60.3 TWh
Annual emissions: 60.3 TWh × 0.35 t/MWh ≈ 21 million metric tons CO₂
For comparison, the James H. Miller Jr. plant in Alabama — currently the US's top emitter — released ~20.7 MMT in 2022. Amazon's single campus would edge it out. And that's before transmission losses, before the diesel backup generators every hyperscaler specs for resilience, before the embodied carbon of the 2M+ GPUs this capacity implies.
The accounting sleight of hand
Here's where it gets frustrating. Amazon will report "100% renewable" for this campus via annual matching — buying enough wind/solar certificates globally to equal their annual consumption. But hourly matching? The grid doesn't work that way. When the Pennsylvania gas plants ramp at 3 AM to feed the H100 clusters, those electrons have a carbon intensity. The certificates don't change the physics.
Microsoft's 2024 sustainability report admitted as much: their Scope 2 market-based emissions looked great, location-based told a different story. Amazon's not even publishing location-based figures for this project yet.
What would actually move the needle
- Hourly CFE tracking — Google's 24/7 carbon-free energy approach, mandated contractually
- On-site storage mandates — 4-8 hour battery co-location as interconnection condition
- Demand response integration — training workloads that can shed 20-30% load during grid stress
- Nuclear PPA carve-outs — direct contracts with existing plants (like Talen's Susquehanna deal) rather than vague "future SMR" promises
The DOE's recent $1.5B grid resilience funding could help, but it's pennies against a 7.65 GW problem.
The uncomfortable question
We're building infrastructure that locks in fossil generation for 20+ years because the AI race demands capacity now. The hyperscalers have the capital to solve this — they're choosing speed over carbon. Until investors and regulators price the emissions externality at the project level, not the corporate portfolio level, every "net-zero by 2040" pledge is just marketing against a 2027 gas plant commissioning date.
The Pennsylvania DEP permit hearings start Q3. That's where this gets real.