The biggest bottleneck in AI right now isn't GPU supply or model architecture. It's power. And on September 16, 2026, three major industry players launched an alliance to address it.
NVIDIA, Google, and Emerald AI announced the AI Energy Management Alliance (AEMA) — a coalition of 18 companies spanning AI labs, utilities, and power producers. The core idea: make AI data centers flexible enough to adjust their energy use when demand is high, in exchange for faster and larger connections to the power network.
Anthropic is a member. National Grid is in. AES and NRG — two of America's largest power producers — joined as well. This is an effort to rewrite how AI infrastructure connects to the U.S. energy system.
The bottleneck
A new AI data center in a major U.S. market can wait five to seven years or more for a connection, per AEMA's analysis. Utilities must build capacity for peak demand — hot summer afternoons when cooling systems run full tilt — even though the network averages roughly 50% utilization year-round.
Traditional rules treat every data center as a fixed, always-on consumer. The Brattle Group estimates each 10% gain in utilization lowers rates by about 3.4%. AEMA cites a Duke University study finding that flexible data centers could unlock 100 gigawatts of capacity on the existing U.S. network — enough for roughly 75 million homes.
How it works
A flexible data center has several tools:
- Shift workloads — pause batch jobs, move inference to other regions, throttle non-urgent tasks
- Use on-site storage — batteries carry the facility through peak moments
- Tap paired generation — on-site renewable or backup generation steps in
- Pre-agreed response protocols — defined procedures for high-demand periods
NVIDIA and Emerald AI have completed six global demonstrations, including trials that reduced consumption by a third in under a minute. Later this year in Virginia, NVIDIA, Digital Realty, and Emerald AI plan to activate a nearly 100 megawatt facility designed as a test case for this approach.
Google already operates a demand-response portfolio of roughly 1 gigawatt across its data centers.
Key figures
| Metric | Figure | Source |
|---|---|---|
| Potential network capacity unlocked | 100 GW | Duke University Nicholas Institute (cited by AEMA) |
| Average U.S. network utilization | ~50% | AEMA |
| Rate reduction per 10% utilization gain | 3.4% | The Brattle Group |
| Avoided system costs per GW of flexible capacity | $733M | AEMA |
| Typical connection wait for new data centers | 5–7+ years | AEMA |
| Google's existing demand-response portfolio | ~1 GW | Emerald AI / Fortune (Sep 16, 2026) |
| Virginia test facility (late 2026) | ~100 MW | NVIDIA / Emerald AI |
Why it matters
The AI industry has treated energy as a logistics problem — build more facilities, sign more purchase agreements. AEMA offers a different framing: the network is a resource to collaborate with, not just consume from.
If regulators offer faster connections to flexible facilities, that 5–7 year wait could shrink significantly. That changes the economics of U.S. AI buildout and the competitive position of American labs.
Emerald AI CEO Varun Sivaram, writing in Fortune on September 16, 2026 at 08:58 (UTC-4), said: "America could unlock 100GW on our existing grid for flexible data centers." That's not marginal. It's potentially doubling available capacity without new generation.
Caveats
AEMA is technology-neutral and performance-based, but hasn't yet defined exactly how flexible a facility needs to be. Metrics like response speed and duration are still being worked out. The 100GW estimate comes from one study and assumes moderate flexibility during peak hours — real results may differ. The $733M per GW figure is AEMA's own estimate.
There's also a practical tension: frontier model training runs are not easily pausable. A job mid-checkpoint can't simply be throttled for hours. Demonstrations so far have focused on inference and batch workloads, which are more shiftable. Whether large-scale training can be flexible remains unproven.
What to watch
The Virginia 100MW facility is the first real test. If it delivers precise, controllable load response, other utilities will likely follow. Policy momentum is building: the Federal Energy Regulatory Commission in June 2026 directed six regional operators to accommodate flexible large customers. Texas is finalizing similar rules. Silicon Valley Power launched the first flexible-load connection program.
Prediction: within 18 months, at least two major U.S. utilities will offer faster connection timelines for flexible loads. When that happens, every AI facility developer faces a choice: build independently at high cost, or adapt to work with the network. AEMA is betting on the latter.
No comments yet