Energy / Data-Center Grid Flexibility Software · Deep dive

Emerald AI

The Washington DC unicorn — a 22-month-old software company that raised $150M Series A at a $1.05B valuation on August 25, 2026 to sell hyperscalers, colocation operators and utilities a workload-scheduling layer that lets AI data centers cut power 25% for hours on grid-stress days while keeping training and inference online, unlocking capacity that would otherwise sit stuck behind seven-year interconnection queues.

emerging

The question that decides it: Does Emerald Conductor become a required layer in the utility interconnection agreements every AI factory has to sign — before Google's own 1 GW internal demand-response stack, Microsoft's Sustainable AI group, Amazon's power-management team and Meta's DCFlex work collapse workload-level grid flexibility into a category hyperscaler platform teams build in-house and eat, and before Voltus, Enel X, CPower and Schneider-Autogrid bolt a data-center orchestration layer onto their existing utility contracts and re-take the aggregator margin?

My take ★★ Very interesting ★ Interesting − Maybe ✗ Not interesting

Snapshot

Emerald AI is a 22-month-old Washington DC software company that raised $150M in a Series A announced August 25, 2026 at a $1.05B post-money valuation, co-led by Energize Capital and DCVC. It sells Emerald Conductor — software that schedules AI workloads against onsite batteries and grid signals so data centers behave like flexible grid assets. Per LBNL’s January 2025 update, US data-center load could rise from 4.7% of electricity in 2024 to 9.5-15.3% by 2030, while 2,060 GW of generation sits stuck in interconnection queues running seven years in Northern Virginia. The cap table is the story: NVIDIA, Samsung Ventures, Siemens, GE Vernova, RWE, Aramco, JERA, Sabanci Climate, ADVentures, Salesforce Ventures and In-Q-Tel all wrote checks.

Founding story

Emerald is founder-first. Varun Sivaram alternated through clean-energy operating and policy roles at unusual seniority: McKinsey energy analyst, director of the energy and climate program at CFR, CTO at ReNew Power (India’s largest renewable-energy company), Group SVP for Strategy and Innovation at Ørsted from February 2023, senior advisor to John Kerry as US Special Presidential Envoy for Climate during Biden’s first two years (where he built the First Movers Coalition), and senior advisor to John Podesta at the White House. That rolodex matters because Emerald sells coordination across operators, utilities, regulators and national-security stakeholders — and Sivaram gets the call returned.

Emerald incorporated in DC in November 2024. The $24.5M seed announced July 2025 (Radical Ventures lead) came with an angel list — Jeff Dean, Kerry, Turnbull, Fei-Fei Li, Doerr — closer to a Time 100 dinner than a cap table. A $22.7M seed expansion in February 2026 came with power-producer partnerships (AES, Constellation, Invenergy, NextEra, Vistra). The Series A added Siemens, GE Vernova, RWE, Aramco, JERA, Samsung, Sabanci and In-Q-Tel — a set that only makes sense if Emerald is infrastructure diplomacy dressed as a software company.

How it works

Emerald Conductor sits above three things: the AI workload scheduler already inside the cluster (Kubernetes, Slurm, Ray or proprietary hyperscaler orchestrators), the onsite energy stack (UPS, batteries, generators, PPAs), and the utility or ISO/RTO dispatch interface. When a utility signals grid stress, Conductor decides which workloads can be paused, throttled or shifted without breaking latency SLAs (batchable training and offline inference qualify; live user-facing inference does not), how much of the drop is absorbed by onsite batteries, and how much of the site’s obligation can be monetized through capacity payments, demand-response revenue or peak-demand-charge reductions.

The May 3, 2026 Phoenix demonstration with Oracle Cloud Infrastructure, NVIDIA, Salt River Project and EPRI DCFlex is the proof point: a 256-GPU cluster ramped power down 25% over 15 minutes, held three hours during SRP’s system peak, then ramped back up without exceeding baseline. Five demonstrations — Arizona, Illinois, Virginia, Oregon, London — cover most US ISOs and one European market. The unanswered question is what Conductor does when grid stress collides with a hyperscaler’s own training-run SLA.

Product and business overview

Emerald Conductor is sold under three wrappers. The operator/hyperscaler wrapper is software installed at the site talking to the workload scheduler and the utility, priced as a service that unlocks additional capacity from an existing interconnection. The utility wrapper is a control-room deployment — Silicon Valley Power’s April 21, 2026 Santa Clara pilot — that dispatches flexible customers during peak periods. The AI-lab wrapper is a geo-shifting scheduler routing training compute across sites.

Business model and pricing

Emerald earns money on both sides of the transaction. On the utility side it takes a percentage of the additional power sales it enables — a utility that would otherwise slow-walk a new interconnection can now serve it. On the operator side it takes a percentage of value captured — peak-demand charge reductions, faster time-to-power, capacity-market revenue, demand-response payments. No per-megawatt rate is published; no ARR is disclosed. Emerald is selling a claim on the future flexibility premium of AI-driven grid capacity — which is directly what a hyperscaler general counsel would try to renegotiate at renewal.

Traction over time

No customer count, GW-under-management or ARR is disclosed. What is disclosed: five demonstrations in twelve months, one live utility pilot, one multi-megawatt commercial deployment.

Market analysis

LBNL’s January 2025 update projects US data centers at 649 TWh in 2030 — 11.8% of US electricity in the reference case, 9.5-15.3% across scenarios, up from 4.7% in 2024. Roughly 2,060 GW of generation sat in active US interconnection queues at end-2024 per LBNL, with waits of seven years in Northern Virginia and five in ERCOT. EPRI figures cited by IEEE Spectrum and APPA suggest flexibility could unlock 100+ GW of grid headroom without new generation — the number Emerald quotes. AI training is, in principle, exactly the interruptible load demand-response was built for.

Competitive intel

Rivals cluster in three layers. Hyperscaler self-build: Google announced in 2024 that its own workload-shifting stack targets ML jobs during grid stress; by mid-2025 it had signed ~1 GW of data-center demand response into long-term utility contracts. Microsoft’s Sustainable AI team, Meta’s DCFlex work and Amazon’s internal power-management team point at the same in-house build. This is the existential competitor. VPP / DERMS incumbents: Enel X (4.3 GW under management), CPower (>4 GW), Voltus (1.7 GW, Nasdaq-listed after its 2024 SPAC), Schneider-owned AutoGrid ($160M raised pre-acquisition), Octopus-owned Uplight, Bidgely, GridPoint and Sympower — each has utility relationships and ISO/RTO market infrastructure Emerald must rebuild. Adjacent entrants: GridBeyond ($90M raised) is running an AI-enabled data-center DR program with Constellation Energy in PJM — the closest direct competitor. If EPRI DCFlex produces a standard reference architecture, Emerald’s differentiation collapses into commodity implementation.

History and evolution

• November 2024 — Sivaram founds Emerald AI in Washington DC.
• July 2025 — Public launch, $24.5M seed round.
• October 2025 — Joins EPRI’s DCFlex initiative alongside Google, Meta, NVIDIA and 15+ utilities.
• February 2026 — $22.7M seed expansion; partnership with AES, Constellation, Invenergy, NextEra, Vistra.
• April 21, 2026 — Silicon Valley Power pilot announced in Santa Clara.
• May 3, 2026 — Phoenix demonstration with SRP, Oracle Cloud, NVIDIA and EPRI.
• August 25, 2026 — $150M Series A at $1.05B post-money.

No disclosed churned customer, no botched utility integration, no failed pilot. The absence of stumbles at 22 months is a data point, not a moat.

What people say

The case for. Bloomberg, Fortune, Latitude Media, SiliconAngle and NVIDIA’s own case study frame Emerald as the leading independent workload-flexibility software company, with the Phoenix demo cited as the most credible public proof of AI-workload DR to date. Energize (climate-tech growth) and DCVC (deep-tech) co-leading is a specialist-investor signal. The strategic-investor set shortcuts buyer discovery because every one of them either owns a data center, sells to one or regulates one.

The complaints. Buy-versus-build risk at hyperscaler scale: Google’s 1 GW internal program means the realistic addressable market is tier-2 colos, sovereign clouds, utilities and enterprise data centers — not the top three US hyperscalers. Two-sided revenue share caps margin at renewal. ScienceDirect and RMI research have flagged that shifting workloads can raise emissions where marginal generation is fossil. FERC and PUC flexible-load dockets can re-shape the wedge in one rate case; utility SLA risk-aversion (per EY’s 2025 report and the Northern Virginia backup-power incident) makes enterprise cycles run in years. $1.05B on undisclosed ARR bets Emerald wins the category outright.

Outlook: the open question

The open question is whether Emerald Conductor becomes a required layer of the utility interconnection agreement for every new AI factory — before Google’s 1 GW internal demand-response stack, Microsoft’s Sustainable AI team and Amazon’s own power-management work collapse workload-level grid flexibility into a category hyperscaler platform teams eat in-house, and before Voltus, Enel X, CPower or Schneider-owned AutoGrid bundle their existing utility contracts with a data-center orchestration SKU that carriers, not Emerald, control.

For the bull case, four things have to hold. A FERC-jurisdictional utility (Duke, Dominion, Georgia Power, ERCOT retailers) has to structure a rate or interconnection product that references Emerald Conductor to obtain faster or larger service. A non-hyperscaler AI compute buyer (Oracle, CoreWeave, Nebius, Crusoe, Lambda, a sovereign cloud) has to commit publicly to Emerald as its flexibility layer across a US fleet. Revenue-share pricing has to survive sophisticated renewal counterparties. And EPRI DCFlex must not publish a vendor-neutral reference architecture.

For the bear case: Google’s playbook becomes the reference Microsoft, AWS and Meta copy; addressable market shrinks before Emerald hits multi-hundred-million-dollar ARR. Voltus, Enel X or Schneider-AutoGrid ships a data-center wrapper leveraging existing ISO infrastructure and undercuts the revenue share. FERC or PUCs approve flexible-load tariffs that let utilities extract flexibility via rate design directly. Or a training-run interruption during a real grid event costs a customer a seven-figure loss and freezes hyperscaler enthusiasm for two years. The tells: the first Emerald contract naming a top-three US hyperscaler; the first utility rate case crediting an Emerald-enabled customer; whether a Series B above $2.5B post arrives before or after Google publishes its own orchestrator post-mortem.

How to attack it

Ship an open-source workload-flexibility scheduler and monetize the market-participation layer instead. Emerald’s structural weakness is that the workload-scheduling half of Conductor is not deep IP — it is a Kubernetes / Slurm / Ray operator with a market-signal input and a policy engine. A well-funded challenger can open-source that layer, drive standardization through the CNCF or Linux Foundation Energy, and get adopted by hyperscalers who will never install a proprietary agent on their training clusters. Monetization moves down-stack to the utility and ISO settlement layer — telemetry, forecasting, capacity qualification — where multi-year utility integration is a real specialist software problem. Voltus made the same move in C&I DR a decade ago.

Bundle grid flexibility with the compute contract. The durable wedge is not selling flexibility to the data center — it is selling compute that comes bundled with flexibility. A CoreWeave, Crusoe or Nebius challenger who signs a utility PPA with an embedded flexibility clause and passes the discount through to its GPU-hour price wins on cost per training run. Emerald cannot do this because it is the software vendor, not the compute seller.

The weaknesses. IP is thin — the Phoenix writeup and NVIDIA case study describe the mechanics in enough detail that a competent hyperscaler team can rebuild the pattern in a quarter. Pricing is transparent enough to invite renegotiation. Top-of-funnel is over-indexed to conference partnerships and DCFlex demos, not repeatable enterprise sales. The 12 strategic investors’ deliverable customers (Aramco, JERA, Sabanci) are not the US hyperscaler market. The founder-CEO is a policy operator, not a data-center operator.

Adjacent-segment play

Same playbook, different flexible load. The Conductor architecture — schedule interruptible load against grid signals and onsite storage, monetize flexibility to both sides — generalizes to every large flexible C&I load: crypto mining (Riot, Cipher, Marathon and Core Scientific already curtail on ERCOT signals), EV-charging depots (ChargePoint, EVgo), industrial hydrogen electrolyzers (Plug Power, Electric Hydrogen), aluminum smelters, ammonia plants and district-cooling chill-water. Each has a workload scheduler, an onsite storage option and a grid interconnection.

The buyer axis is more interesting. The utility wrapper is effectively a DERMS; a build-out into residential and small-commercial DR (Uplight, Bidgely territory) is a larger but lower-margin market. Sovereign clouds and defense-adjacent compute — the In-Q-Tel investment is the tell — are the natural up-market segment. Geographic adjacency is real: the JERA, Sabanci, ADVentures and Aramco investments point at Japan, Turkey, the UAE and Saudi Arabia as first-move international markets, where sovereign AI buildouts create clean-sheet buyers without US hyperscaler self-build competition. The likeliest non-AI adjacency is EV-charging depots — Emerald’s compute-scheduling IP maps to charge scheduling with modest rework.

Sources and further reading

• Emerald AI Raises $150 Million Series A at $1.05 Billion Valuation — Businesswire, August 25, 2026.
• Data center power startup Emerald AI raises $150M at $1.05B valuation — SiliconAngle, August 25, 2026.
• How Emerald AI Makes AI Factories Power-Flexible — NVIDIA case study, 2026.
• Emerald AI Launches with $24.5M Seed Round — PR Newswire, July 2025.
• Emerald AI and Nvidia aim to offer a fast pass for data center grid connects — Fortune, March 31, 2026.
• SRP Participates in AI Data Center Demonstration — American Public Power Association, 2026.
• Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers — LBNL, January 15, 2025.
• Google signed 1 GW of data center demand response — Google blog, 2025.
• EPRI launches data center flexibility initiative — Utility Dive, October 2024.
• Meet Varun Sivaram of Emerald AI — Axios, November 3, 2025.