The Grid Becomes a Scheduler: Emerald AI's AEMA Turns 100 GW of Hyperscaler Power Into an Auctionable Resource

AEMA is the first cross-industry body to ask AI data centers to behave like grid resources rather than grid sinks. NVIDIA, Alphabet, and Emerald AI are the founding members; Anthropic, AES, Constellation, National Grid, NRG Energy, RWE, Fluence Energy, Voltus and a dozen others joined as launch partners the same day. The technical hook is Emerald Conductor, an orchestration platform that throttles GPU workloads, geo-shifts inference traffic, and dispatches onsite batteries in response to utility signals.

Goldman Sachs's research arm models US data-center demand more than doubling from 31 GW in 2025 to 66 GW in 2027, with the share of US summer peak demand attributable to data centers rising from 4.1% to 8.5% over the same window. The same research shows that capping data-center draw at 90% of nameplate for just a few hours a day would free 76 GW of connectable capacity, essentially the entire current deficit between announced projects and grid headroom.

PJM's December 2025 capacity auction for delivery year 2027/28 cleared at $333.44/MW-day, the federally capped Net Cost of New Entry for the third consecutive auction. The Independent Market Monitor attributes roughly 40% of the $16.4B clearing cost to data-center load, and the auction still came up 6,625 MW short of the 152,102 MW reliability target. Dominion Zone, the densest data-center cluster in North America, has revised its 2037 load forecast up more than threefold since 2022.

Demand response was already PJM's biggest non-generation resource at 7,299 MW of cleared UCAP after the ELCC factor was lifted from 69% to 92%. AEMA's standardized verifiability is designed to push that number much higher, and every incremental MW of credible data-center demand response is one fewer MW of new gas turbines PJM must procure at capped prices.

The hyperscalers Alphabet and NVIDIA are price-setters, not the marginal winners. They will lock in cheaper, faster capacity and amortize Emerald's software across their fleets. The marginal alpha accrues to the vendors who can deliver the flexibility they are now selling, and the supply chain splits into four layers.

The non-obvious angle is that AEMA turns PJM-style capacity markets from a bill for hyperscalers into a revenue line for utilities that can credibly bid flexibility bundles. Constellation Energy's $7.50B Q2 2026 revenue (+23% YoY), driven by its Calpine gas-and-nuclear portfolio, is the template: AEMA membership effectively monetizes that baseload into dispatchable credits. AES, NRG Energy, and National Grid join the alliance for the same reason. National Grid's UK trials with Emerald (200/200 targets hit) and the Freestone, Texas gas-backed data-center deal Constellation just lit up show the playbook in production.

Short-term (days to quarters), the catalysts are contractual: FERC's 60-day large-load tariff revisions land in late Q4, and PJM's June 2026 auction for delivery year 2028/29 will be the first to price AEMA-style demand response at scale. Watch for Microsoft and Amazon to either join AEMA or be forced to build a competing scheduler, a binary outcome for NVIDIA's DSX software margin.

Long-term (one to three years), the thesis is that flexibility becomes a software margin business with three layers: a scheduler (Emerald Conductor and its likely competitors), a measurement-and-verification utility (AEMA), and the equipment primes that monetize every new interconnection. Eaton's solid-state transformer pipeline and GE Vernova's 116 GW turbine backlog give them visibility on the equipment layer regardless of which scheduler wins. The risk is a hyperscaler-specific stack, OpenAI's recent 10 GW, $1.5B-NVIDIA-backed lease at SB Energy's Ohio campus is the prototype, pulling the scheduling logic back inside the buyer and pricing Emerald-style utilities out of the highest-margin sites.