WASHINGTON, 30 AUG 2026 — Emerald AI has raised US$150m at a US$1.05bn valuation to sell software that slows, pauses, caps or moves computing jobs when a utility needs demand reduced. Nvidia, Samsung Ventures, GE Vernova and Salesforce Ventures are among the investors.
What the company sells
The product treats a data centre as a controllable load rather than a fixed one. When the grid is short, the software curtails or shifts selected workloads, and it does so in a way the utility can verify.
That verification is the commercial point. A utility will not grant faster or larger grid access on a promise; it will do so against dispatchable flexibility it can measure and call on. Emerald has partnered with Silicon Valley Power on what is described as a first-in-the-nation Flexible Load Interconnection Program, which grants expanded grid access in exchange for exactly that.
The Series A was oversubscribed and co-led by Energize Capital and DCVC, bringing total funding above US$220m. In Manassas, Virginia, the company is working with Digital Realty and Nvidia on a nearly 100MW power-flexible facility due online later this year.
The 100GW figure is the company's own model
Emerald estimates that flexible operation could unlock more than 100 gigawatts of capacity on the existing US system, years before new construction could deliver it. The number is striking, and it needs context.
The figure is the vendor's own model of potential headroom, assuming operators accept curtailment. It is not observed capacity, and there is no guarantee operators will accept the terms at that scale.
The underlying physics is sound and unremarkable: grids are sized for peak, peaks are brief, and any load that can stand down during them frees capacity that already exists. Demand response has worked this way in industry for decades. The new part is applying it to data centres, a load class that has been treated as inflexible.
Whether AI workloads can actually flex
The funding announcement does not answer how much of that 100GW is achievable in practice.
Training runs are the least flexible thing in a data centre. A large run holds thousands of accelerators in a synchronised state for weeks, and pausing it is not free — checkpoints cost time and storage, and a job interrupted at the wrong moment loses work. An operator with a contractual delivery date for a model is not an eager participant in curtailment.
Inference is more tractable in principle, because requests can be routed elsewhere, but it is also the workload with the tightest latency expectations and the one closest to customer-facing revenue.
The flexible capacity is in batch processing, non-urgent fine-tuning, data preparation, and the idle or lightly loaded fraction of any cluster. That resource is real, and it is smaller than the total connected load, which explains the gap between the model and likely deployment.
Why this matters more here than in Virginia
Southeast Asia has the grid problem that flexibility is designed to solve, and almost none of the conversation.
We reported yesterday that Malaysia's utility describes a data centre pipeline of 8.3GW against roughly 1.05GW of load actually drawn. The Philippines' proposed Luzon hub would need 16 per cent of the island's grid. Singapore has been importing solar from Johor rather than building.
In each case the constraint is firm capacity at peak. To a regulator deciding who gets a connection, a data centre that can be curtailed during peaks is a different and much cheaper proposition than one requiring new generation, because the capacity already exists.
The obstacle here is institutional rather than technical. Flexible interconnection requires a utility able to write and enforce such a contract, a regulator willing to approve differentiated grid access, and metering that both parties trust. Thailand has already shown the appetite by using a grid-connection letter as the gate on investment incentives. Turning that gate into a flexibility contract is a smaller step than building a power station.
The alternative is making data centres pay
Flexibility is one of two answers to the same problem, and the other one is already being tried.
Regulators in several jurisdictions have moved towards making large loads carry the cost of the network they require, rather than socialising it across other customers. We have covered the US push to make data centres cover their own power, and FERC's sixty-day order to all six US grid operators on large-load interconnection rules.
Those approaches allocate a scarce resource by price. Flexibility increases the resource instead. They are complements rather than rivals, and the second is politically easier because nobody's bill goes up.
There is a third option that some operators have taken, which is to stop asking the grid at all. Anthropic paid US$9.1bn for electricity rather than computers by leasing capacity already connected for bitcoin mining. That works where such capacity exists and is not a general answer.
Read the investor list
Who put money in tells you what the round is for.
Nvidia sells the accelerators whose power draw created the problem, and has an obvious interest in any mechanism that lets more of them be connected. GE Vernova sells grid and generation equipment and is on the utility side of the same transaction. Samsung Ventures and Salesforce Ventures are strategic in the sense of being large consumers of compute.
The investor list is a coalition of parties who all benefit from getting data centres connected faster. That signals real demand for the product, and it also frames the 100GW estimate as an argument rather than a measurement.
The number to watch is not the valuation. It is whether a second utility signs a flexible interconnection programme, because one is a pilot and two is a category.
View author profile → · Editorial policy
