Energy · Deep dive

Emerald AI

Washington, DC startup from physicist and Taming the Sun author Varun Sivaram whose Emerald Conductor software makes AI data centers flexible grid assets — pausing, slowing and shifting GPU workloads during grid stress in exchange for faster interconnection — $68M raised in 16 months from Radical Ventures, NVIDIA's NVentures, Eaton, GE Vernova and IQT, with a 96MW flagship deployment opening in Manassas, Virginia in 2026.

emerging

The question that decides it: Emerald sells curtailment as currency: a data center gives back roughly 25% of its power for a few hundred hours a year and gets a grid connection years faster. Does that trade hold at commercial scale once someone has to pay for it — will operators of scarce GPU fleets, where an idle cluster burns far more in depreciation than it saves in electricity, accept live utility dispatch on revenue-bearing inference (not just batchable training, which Latitude Media notes is a shrinking share of load) — and does an independent orchestration layer survive NVIDIA folding flexibility into its DSX reference design, Google running gigawatt-scale demand response in-house, and utilities writing curtailment obligations directly into interconnection tariffs?

Snapshot

Emerald AI is the best-connected bet that the answer to AI’s power crunch is software. Its Emerald Conductor platform sits between the electric grid and a data center’s GPU clusters and, when the grid is stressed, pauses, slows or relocates AI workloads so the facility can shed 25% or more of its draw for a few hours — turning the fastest-growing load on the grid into a dispatchable asset, and in exchange getting new data centers connected years faster. Founded in late 2024 by physicist and former Kerry climate aide Varun Sivaram, the company raised $68 million in 16 months from a cap table that reads like an industrial alliance — NVIDIA, Eaton, GE Vernova, IQT, Samsung, Siemens, Energy Impact Partners — and its 96MW Aurora AI Factory in Manassas, Virginia, slated to open in the first half of 2026, is billed as the world’s first power-flexible AI facility. Revenue, so far, is undisclosed and likely minimal: the story to date is demonstrations, not contracts.

Founding story

Sivaram is a rare founder whose entire career is the pitch. A Stanford physicist and Rhodes Scholar with an Oxford DPhil on perovskite solar cells, he wrote Taming the Sun (2018), served as CTO of ReNew Power — India’s largest renewables producer — taught at Georgetown and Columbia, held a Council on Foreign Relations fellowship, then ran clean energy and innovation for John Kerry’s climate envoy office in the Biden administration, where he built the First Movers Coalition. In February 2023 Ørsted made him Group SVP of Strategy and Innovation. It was there, inside the world’s largest offshore wind developer, that he hit the founding insight: AI data-center demand was growing faster than anyone could permit and build generation, and interconnection queues were stretching toward a decade. His conclusion, as he has told it repeatedly: you cannot build your way out — you need intelligent demand (Fortune, March 2026).

He founded Emerald AI in Washington, DC in late 2024, as a solo founder but with an unusually deliberate bench: Boston University professor Ayse Coskun, arguably the leading academic on data-center demand response, as Chief Scientist, and ex-AWS compute leader Shayan Sengupta as Head of Engineering. The angel list — Jeff Dean, Fei-Fei Li, John Doerr, Kerry himself, former Australian PM Malcolm Turnbull — signals the same thing the strategic round later confirmed: this is a company built on Sivaram’s network at the intersection of energy policy, grid operators and AI, which is precisely the intersection the product must broker.

How it works

Emerald Conductor is a control loop between a grid signal and a GPU scheduler. A utility or system operator issues a dispatch — reduce to a given power target, at a given ramp rate, for a given duration. Conductor receives it and decides, job by job, how to comply. Workloads are pre-classified into flexibility tiers (reported as Flex 0–3): a checkpointed training run can be slowed or paused outright and resumed after the event; batch inference and fine-tuning can be throttled by capping GPU power or deferred; latency-sensitive inference is protected, or rerouted over fiber to a data center in an unstressed region within latency bounds; on-site batteries cover what compute cannot. A companion digital twin, the Emerald Simulator, predicts cluster power behavior — a peer-reviewed field paper reports 4.52% RMSE on power prediction — so the company can promise a utility a specific megawatt trajectory without breaching customer SLAs (arXiv, July 2025).

The proof points are real and unusually well documented. In a spring 2025 field demonstration in Phoenix with NVIDIA, Oracle Cloud, EPRI and utility Salt River Project, a 256-GPU cluster running representative AI jobs cut power 25% for three hours during a peak event via a 15-minute graceful ramp, then recovered without rebound above baseline. In December 2025, at Nebius’s London AI factory, a 96-GPU Blackwell Ultra cluster took 22 live dispatch events from National Grid Electricity Transmission and EPRI over five days — including a simulated “TV pickup” surge — and hit 100% compliance on 200+ power targets, with 30-40% reductions and sub-minute response (National Grid, March 2026). The strategic logic: Duke’s Nicholas Institute calculated (February 2025) that 76GW of new load could join the existing US grid if curtailed just 0.25% of hours; Emerald and NVIDIA’s version of the claim is that flexibility unlocks 100GW. Flexibility is thus a queue-jumping currency — utilities can connect a flexible data center years sooner because it stays off the peak.

Product and business overview

Three components carry the offering. Emerald Conductor, the orchestration layer, integrated with cluster schedulers and, since October 2025, with NVIDIA’s Omniverse DSX Flex reference design for gigawatt-scale AI factories. Emerald Simulator, the digital twin used to model a facility’s flexibility envelope — the tool that lets a developer show a utility, pre-interconnection, exactly what the site can shed. The reference design and certification standard announced with NVIDIA, EPRI, Digital Realty and PJM in October 2025, of which the 96MW Aurora AI Factory in Manassas — built by Digital Realty, described by Fortune as NVIDIA’s Vera Rubin AI Factory Research Center — is the first implementation. Around this sits a partner web unmatched for a seed-stage company: a March 2026 coalition with AES, Constellation, Invenergy, NextEra, Nscale and Vistra to co-develop grid-flexible AI factories, a Silicon Valley Power pilot in Santa Clara, and the National Grid work in the UK, where partners estimate flexible data centers could hand back over 2GW of the 6GW+ UK pipeline by 2030.

Business model and pricing

This is the page’s weakest section because it is the company’s least-formed layer. Emerald publishes no pricing; the model described in investor materials is software-as-a-service sold to data-center operators and developers, with value flowing from three places: interconnection speed (getting energized years early is worth enormous option value on a multi-billion-dollar facility), avoided peak-power and capacity costs, and demand-response or flexibility payments from utilities, some of which are only now inventing the tariffs that would pay for it. Salesforce Ventures adds energy arbitrage and ESG reporting to the list (July 2025). Nothing about per-MW fees, revenue share on flexibility payments, or contract structure has been disclosed, and through July 2026 no named paying customer has been announced — Aurora is a flagship deployment with partners who are also investors. The honest summary: the mechanism is proven, the monetization is still a hypothesis, and a reported effort to raise a further $100 million (BeBeez, July 2026) would be priced almost entirely on the partner roster.

Traction over time

Sixteen months from founding to $68 million with three governments’ worth of energy establishment on the cap table is extraordinary velocity. The bottom row is the caveat: every cell reads undisclosed.

Market analysis

The structural force is the largest demand shock the US grid has seen in a generation. LBNL’s DOE-commissioned report (December 2024) found data centers consumed 4.4% of US electricity in 2023 (176 TWh) and projects 6.7–12% by 2028 (325–580 TWh); interconnection and turbine lead times, not capital, are the binding constraint on AI buildout. Flexibility directly monetizes that scarcity: Duke’s 76GW-at-0.25%-curtailment finding (February 2025) and the Emerald/NVIDIA 100GW claim — roughly 20% of US peak capacity, framed as a decade of AI growth without new plants — define the prize, and FERC’s December 2025 direction to PJM on co-located and flexible load plus a wave of state flexible-tariff proposals are building the regulatory rails. The market being created is real but unpriced: nobody yet knows what a megawatt of verified data-center flexibility clears at, which means the TAM for the software layer that delivers it is a derivative of tariffs that mostly do not exist yet. The bear force is workload mix: flexibility is easiest on batchable training, and the industry’s revenue center of gravity is shifting to inference, which tolerates far less interruption (Latitude Media, 2026).

Competitive intel

The named set is in frontmatter; the shape matters more than the list. Emerald’s true competition is not another workload-orchestration startup — it effectively has that niche to itself at this scale — but three substitute paths to the same outcome. Batteries: Verrus and every developer pairing sites with storage deliver grid flexibility without touching compute, at a hardware cost that falls every year. Incumbent aggregators: Voltus, CPower and Enel X already sell curtailment into ISO markets with utility relationships Emerald lacks, just less surgically. In-house: Google has run carbon-intelligent, location-shifting compute since 2020-21 and signed utility demand-response deals in August 2025 — the strongest evidence both that the thesis is right and that the biggest fleets will not pay a third party for it. Emerald’s moat candidates are the NVIDIA DSX Flex integration, the certification standard it is co-authoring, and Sivaram’s unrivaled standing with utilities and regulators. The risk inside the moat: a reference design is by definition replicable, and its most important partner is also the industry’s most powerful potential commoditizer.

History and evolution

• Feb 2023 — Sivaram becomes Ørsted’s Group SVP of Strategy & Innovation; forms the intelligent-demand thesis watching interconnection queues.
• Late 2024 — Founds Emerald AI in Washington, DC; builds in stealth with Radical Ventures.
• Spring 2025 — Phoenix field demonstration with NVIDIA, Oracle, EPRI and Salt River Project: 25% power cut over three hours on a 256-GPU cluster.
• Jul 1, 2025 — Launches from stealth with $24.5M seed led by Radical Ventures; publishes the Phoenix results on arXiv.
• Oct 29, 2025 — Aurora 96MW AI Factory and power-flexible reference design announced with NVIDIA, EPRI, Digital Realty and PJM; 100GW unlock claim; DSX Flex integration.
• Oct 31, 2025 — $18M seed extension led by Lowercarbon Capital ($42.5M total).
• Dec 2025 — UK-first live grid trial at Nebius London with National Grid and EPRI.
• Feb 19, 2026 — Axios reports the strategic raise; SEC filing shows $22.7M from twenty investors.
• Mar 3, 2026 — UK results published: 100% compliance across 200+ dispatch targets.
• Mar 31, 2026 — $25M strategic round led by Energy Impact Partners with Eaton, GE Vernova, IQT, Samsung, Siemens ($68M total); NVIDIA-led coalition with AES, Constellation, Invenergy, NextEra, Nscale, Vistra.
• H1 2026 — Aurora slated to open; Silicon Valley Power pilot in Santa Clara.
• Jul 15, 2026 — Reported to be seeking a further $100M (BeBeez).

What people say

The case for. The technical record has drawn genuine third-party validation: National Grid publicly called the UK trial a success, citing sub-minute 30%+ reductions and perfect dispatch compliance (March 2026); EPRI and Salt River Project co-authored the Phoenix results; Duke’s Tyler Norris — the field’s most-cited researcher — has framed load flexibility as the fastest lever for integrating AI demand (February 2025). Lowercarbon’s investment memo dubbed the product an express lane for electrons (October 2025), and Utility Dive’s coverage describes utilities actively courting flexible interconnection as the political answer to data-center rate anger. The caliber of strategics — Eaton, GE Vernova, Siemens, IQT — is itself a reference check: the companies that sell the grid’s hardware bought into the software.

The complaints. The skepticism is structural, not personal. PJM’s independent market monitor called data-center flexibility a “regulatory fiction,” arguing that without binding curtailment authority there is no guarantee load actually sheds when it matters, and urged FERC to require matching generation instead (Latitude Media, 2026). Economists and operators note the brutal arithmetic of GPU scarcity: a cluster’s depreciation and opportunity cost per hour dwarf its electricity cost, so voluntarily idling compute is expensive precisely when tokens are most valuable — “it’s questionable whether they want to do it,” as one flexibility researcher put it (Utility Dive, 2026). Latitude Media’s reality check adds that the batchable training workloads flexibility depends on are a shrinking share of demand versus inference, which cannot be paused — potentially making software flexibility a niche for training-only sites. And the circularity is hard to miss: NVIDIA is investor, demo partner, reference-design co-author and Aurora anchor, meaning nearly every public proof point involves a shareholder. No customer has yet accepted curtailment with its own money at stake.

Outlook: the open question

Emerald works if, by roughly the end of 2027, curtailment has become a priced product with arms-length buyers: Aurora operating through a full summer of real PJM dispatches at 96MW scale, at least one utility tariff or ISO program paying verified data-center flexibility real money per megawatt, at least one customer who is not also an investor signing a multi-year Conductor contract, and the certification standard it co-wrote becoming a de facto interconnection requirement — making Emerald the toll collector on a lane regulators built. The tailwinds are real: interconnection scarcity is the binding constraint on the largest capex cycle in industrial history, the research consensus (Duke, LBNL, EPRI) supports the mechanism, and no competitor combines GPU-level orchestration with Emerald’s regulatory and utility standing. Emerald fails if flexibility gets delivered without it: batteries falling in cost until Verrus-style designs and simple on-site storage satisfy utilities with zero compute risk; hyperscalers following Google’s path and self-providing demand response; NVIDIA’s open reference design turning workload flexibility into a feature of every scheduler; or the inference shift shrinking the pausable share of load until 25%-for-three-hours is a training-farm party trick — in which world Emerald’s $68 million bought influential demos, a standards credit, and an acqui-hire by Eaton, GE Vernova or NVIDIA. The tells: whether Aurora’s opening slips past H1 2026, whether the reported $100 million raise closes with a disclosed valuation and any revenue narrative, whether any non-investor operator deploys Conductor, and what a megawatt of flexibility actually clears at when PJM stops treating it as fiction.

Sources and further reading

• Emerald AI Launches with $24.5M Seed Round to Transform AI Data Centers into Grid Allies (PR Newswire, July 1, 2025)
• Nvidia stakes new startup that flips script on data center power (Axios, July 1, 2025)
• Turning AI Data Centers into Grid-Interactive Assets: Results from a Field Demonstration in Phoenix, Arizona (arXiv, July 2025)
• Emerald AI Teams with NVIDIA and Partners to Develop Power-Flexible AI Factory and Reference Design to Unlock 100 GW of Grid Capacity (PR Newswire, October 29, 2025)
• Nvidia to deploy Emerald AI’s orchestration software at 96MW Aurora data center in Manassas, Virginia (Data Center Dynamics, October 2025)
• UK-first trial of AI Grid Technology Successfully Demonstrates the Ability of Data Centres to Adjust Power Needs (National Grid, March 2026)
• Emerald AI and Nvidia aim to offer a fast pass for data center grid connects, partnering with power producers and raising new funds (Fortune, March 31, 2026)
• A reality check on flexible data centers (Latitude Media, 2026)
• Duke Researchers: Grid Flexibility Key to Accommodate Load Growth — Rethinking Load Growth (POWER Magazine / Nicholas Institute, February 2025)
• 2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory, December 2024)
• Nvidia-backed Emerald AI seeking to raise a further $100m in funding – report (BeBeez International, July 15, 2026)