$68 million in 16 months. Emerald AI, the software layer that tells a data center when to pull power, has gone from founding to a total haul bigger than most utilities spend on software in a year.
Six of the largest power producers in America signed on with NVIDIA. Roughly 100 GW of existing grid capacity could be unlocked without one new transmission line.
The shift rewards flexibility over raw size. Data centers that can flex their own load are starting to win the race for a grid connection.
Data centers have been treated as a burden on the grid for years. The grid was never designed for their power profiles, and the queues for new connections stretched toward a decade. The story is flipping. A new class of software is making the data center itself the tool that stabilizes the grid it draws from.
As we wrote in August, AI data centers are already buying power a decade early, locking up fusion and firm capacity years before it exists. This is the other half of the answer: not more power, but smarter timing of the power they already have.
The numbers behind the flexibility push
Emerald AI total funding
$25M expansion round led by Energy Impact Partners, with Eaton, GE Vernova, Siemens, Samsung Ventures and Salesforce Ventures among the investors. · Emerald AI, 2026
Existing capacity, no new lines
The Duke Nicholas Institute figure cited by the flexible-factory coalition for what power-flexible compute could unlock on the current US system. · Nicholas Institute, 2026
Power-flexible AI factory
NVIDIA, Digital Realty, EPRI and PJM are driving the Aurora facility in Manassas, Virginia as the first at commercial scale. · NVIDIA, 2026
What a grid-interactive data center actually does
Emerald AI's Conductor sits between the data center and the grid. It reads wholesale prices, weather, and grid stress signals, then shifts compute workloads, battery charge, and on-site generation to match. When the grid is tight, the facility draws less or pushes stored energy back. When power is cheap, it charges batteries and runs more work.
That is not demand response in the old sense of turning things off. The workload moves, not the business.
Five live demonstrations across Arizona, Illinois, Virginia, Oregon and London, run with EPRI and NVIDIA Blackwell Ultra GPUs, showed production workloads flexing without breaking latency. The first peer-reviewed evidence of the concept appeared in Nature Energy.
The growing side of the ledger
NVIDIA and Emerald AI announced a coalition at CERAWeek in March that reads like a who's who of American power: AES, Constellation, Invenergy, NextEra, Nscale and Vistra. The reference design, Vera Rubin DSX, treats energy and compute as one architecture. NVIDIA's Jensen Huang put it plainly: everything must be designed together, energy, compute, networking, cooling.
Adoption is spreading outside the US too. National Grid ran a UK-first trial at its Deeside innovation center. Silicon Valley Power launched a pilot in Santa Clara this spring. The pattern is the same everywhere: utilities that once fought data center connections are now competing for them.
The economics are doing the convincing. Retail electricity prices have risen 42% since 2019. In the PJM market, a six-gigawatt reliability shortfall is projected by 2027. A data center that can pause flexible work is worth more to the grid than one that cannot.
The falling assumption: constant load
The old model treated a data center like a factory running at full capacity forever. That assumption is breaking.
Grid planning assumed each new connection was a firm, uninterruptible block of load. That forced utilities to size transmission for peak draw that rarely happened. Flexible factories break the fiction. They commit to a range, not a ceiling.
This is why interconnection speed is the prize. 77% of data center executives say connection queues are delaying capacity deployment, per Capgemini's research. A facility that can shape its own load earns priority access, the same treatment battery storage gets for stabilizing the grid.
The new asset class: flexible AI factories
The label matters. An AI factory is not a data center with GPUs. It is a coordinated system where compute, storage, power conversion, and grid interface are designed as one.
Battery energy storage systems have become core to this design. NVIDIA published qualification guidelines for production-ready BESS in AI factories, treating storage as a grid-interactive control asset rather than backup power. Fast-changing AI loads, where racks can swing tens of megawatts in milliseconds, need buffers that the grid cannot provide.
Emerald AI's Conductor orchestrates computational flexibility alongside on-site generation and batteries, protecting priority workloads while meeting power targets. The 96 MW Aurora facility in Manassas is the test case for commercial scale.
Flexible factory vs. traditional data center
| Parameter | Traditional data center | Flexible AI factory |
|---|---|---|
| Load profile | ✔ Firm, constant | ◐ Flexible within range |
| Grid value | ✗ A burden to plan around | ✔ A dispatchable asset |
| Interconnection | ✗ Years-long queue | ✔ Priority, faster connection |
| On-site storage | ✗ Backup only | ✔ Grid-interactive buffer |
| Capacity unlocked | ✗ New build required | ✔ Uses existing capacity |
Where the returns land
For investors, the value sits in what the software lets a data center sell, not in the license itself.
Every hour, a flexible factory can choose between running compute and selling grid services. When reserves are tight, the price of firm capacity spikes, sometimes tenfold. A facility that can free 20% of its draw for an hour is sitting on a real source of revenue, not a footnote.
The same logic runs the other way. When power is cheap and renewable output is high, the data center charges its batteries and runs more training. In markets like Texas or California, where prices swing wildly within a day, that spread alone can shift the operating economics of a campus.
Analysts compare the position to batteries a few years ago. Storage was once an insurance purchase. Then markets started paying for speed, and it became a revenue asset. Flexible compute is following the same curve, only the installed base is already enormous.
The capital is not waiting
The list of investors behind Emerald AI reads like a map of who benefits. Energy Impact Partners is backed by utilities. GE Vernova makes the turbines. Eaton and Siemens make the grid equipment. Samsung Ventures and Salesforce Ventures bring the compute and software. NVIDIA's venture arm sits at the center.
That clustering matters. It means the technology is not being paid for by a single curious vertical. It is being funded by the whole value chain that needs the grid to survive AI demand.
For a principal evaluating private investment, the signal is the supply chain voting with real money while regulators are still debating the rules. Software that shapes demand has almost no capex penalty. It runs on hardware data centers already buy, and it pays for itself in avoided connection costs alone.
What could still break it
None of this works if the markets do not pay for flexibility. FERC and state regulators are still writing the rules for how large data centers interact with wholesale power markets. Until those rules are clear, flexible capacity is a promise without a price signal.
The regulatory paradox cuts both ways. Utilities struggle to build storage where it matters most, because rules block them from owning assets that trade in markets. Aggregators fill part of the gap, but only where economics make sense, not where the grid needs it most.
Software orchestration also assumes the data center can shed load without breaking the tenant's service. That requires contracts, not just algorithms. The first customers to sign those contracts will set the standard. Early movers are doing exactly that in Virginia and the UK, testing what a flex guarantee is worth in practice.
FERC and state rulemaking on data center demand response and wholesale participation
The Manassas Aurora facility reaching commercial operation on schedule
Whether interconnection queues for flexible sites shrink measurably versus firm loads
Power purchase agreements that price in flexibility as a service, not a discount
The bottom line
AI data centers will not stop being the grid's biggest problem. But they are starting to be part of the fix.
The money is following: $68 million into the category leader in 16 months, with the world's largest power generators and GPU maker behind it. The technical proof exists. What is left is the price signal, and that signal is arriving market by market. The rebalancing of expectations has already begun.
