Flexibility as capacity: How grid and AI compute interact, and what Emerald AI needs to justify $1.05B

Exec brief

Grid interconnection is the binding constraint on AI datacenter growth. Load-side flexibility is the umbrella term for the ways an AI datacenter converts a portion of its firm electrical demand into a contractually recognised grid resource. The economic product is a flexibility right: a contractual operating claim that allocates the right to change the customer's electrical operating range under defined conditions, in exchange for consideration (faster interconnection, lower firm-capacity charges, capacity-market receipts, or another form of payment).

Six architectural bets are visible in 2026.

• A software coordination layer that reduces AI training on grid signal (Emerald AI).
• A battery-integrated datacenter that shifts load in time (Verrus).
• A behind-the-meter gas microgrid that substitutes for grid power during stress (Enchanted Rock).
• A renewable and battery co-located datacenter (Prometheus + ENGIE).
• A flexibility-ready site development model (Cloverleaf, Rowan).
• A cooling-side thermal shift that appears in academic literature but has not produced a merchant vendor.

Different approaches sell different things. Cloverleaf sells interconnection access. Enchanted Rock sells firm on-site supply. Verrus sells energy stored for later dispatch. Emerald sells a coordination layer that decides which capability to use for which grid event. All of them turn some form of electrical flexibility into a contract a counterparty will pay for.

Because they operate at different points in the electrical stack, they layer on the same site more often than they compete.

The category anchor event is Emerald AI's Series A round announced in late August 2026, co-led by Energize Capital and DCVC with participation from NVIDIA, Radical Ventures, Samsung Ventures, Siemens and GE Vernova, plus Eaton in an earlier strategic round. The round was reportedly $150 million at approximately $1.05 billion post-money (Converge Digest). At the reported post-money valuation, the market has priced Emerald at $1.05B. That price reflects a mix of underwriting theses (hyperscaler curtailment at commercial scale, strategic acquisition optionality, category-standard adoption, interconnection acceleration value); individual investors may weight these differently. This piece unpacks what each thesis requires. Emerald and SRP have documented one commercial-scale curtailment trial, an assumption verified so far in one field trial documented in the Emerald / SRP arXiv paper (Colangelo et al., arXiv 2507.00909). This piece maps the architectures, works through where each creates a flexibility right, and lands on where the defensible margins sit.

01The problem and the product

A datacenter is only as real as the firm megawatts at its meter.

Across PJM, ERCOT, MISO and CAISO, large-load interconnection studies now drag between four and seven years for gigawatt-scale campuses. Generation queues trail further behind. PJM's Expedited Interconnection Track (FERC-approved 9 June 2026) opens narrow generation-side exceptions and does not change the load-side timeline.

Against silicon cycles refreshed every 18 months, that gap leaves full generations of accelerator hardware waiting on gravel lots.

Load-side flexibility is what the customer accepts in exchange for compressing that timeline.

Concretely: a 500 MW datacenter agrees to operate at 350 MW for the 200 hours per year when the grid is stressed. The utility approves an interconnection at 500 MW nameplate against a 350 MW firm commitment. No new capacity was created. Capacity risk moved from the utility to the customer, in exchange for MW that arrive in months instead of years.

Every approach in this piece monetises some form of electrical optionality at the grid interface.

The bankable flexibility right is the financial asset.

An economic flexibility right exists only after a counterparty recognises it, measures it, contracts for it, and pays for it. The scarce economic input is the portion of interconnection scarcity that a flexibility right can transfer from the grid to the customer, after accreditation, financing, enforcement, and value-sharing. Technical flexibility on its own establishes very little bankable investment value.

The product being sold is a contractual claim held by the grid, utility, or ISO to change the customer's permitted operating range under defined conditions. The underlying enabler may be software, a battery, a gas microgrid, a co-located renewable PPA, or an interconnection contract. The battery, generator, and scheduler are enabling assets. The right itself, and the accredited capacity attached to it, is what holds the financial value.

Every company in this piece hinges on two things. Who creates the flexibility right at the grid interface. Who owns the contract for it.

02Four economic functions: shed, shift, substitute, relocate

Load-side flexibility as a phrase covers four distinct economic functions. Every approach in this piece performs one or two of them, never all four. The industry calls all four "flexibility," but they are economically different products.

Four functions land under the flexibility umbrella: shed, shift, substitute, relocate. The visual below holds the definitions.

Older demand-response literature groups "firming" with load-flex. In practice firming clears through capacity-market plus interruptible-load, and that is where Enchanted Rock's gas microgrid economics belong. The layering point (that no single approach performs all four functions, so any real deployment stacks two or three) is developed.

03The stack, not the winner

These approaches operate at different points in the electrical stack. A single hyperscaler site typically layers three or four of them at once. The campus was developed by a Cloverleaf or Rowan (flexibility-ready land plus interconnection). The primary supply-substitution infrastructure is an Enchanted Rock gas microgrid or a Verrus-style battery microgrid. The renewables PPA feeding the site may come through an ENGIE-style structure. The software coordination layer that dispatches load-shed events on grid signal is Emerald AI or a hyperscaler-internal equivalent.

Three scarcity rents accrue at different layers of that stack, and each layer has a different margin structure.

• Interconnection scarcity. The MW at the meter, the queue position, the permitting. Captured by land and site developers. Real-estate developer margins, high single digits to low teens on invested capital. Substrate for everything else.
• Physical flexibility scarcity. The batteries, gas gensets, or renewable-plus-storage assets that create the operating range. Captured by infrastructure balance sheets under long-term contracts. Project-finance IRRs of 8 to 12 percent unlevered on the underlying assets, uplifted by contract structure.
• Coordination scarcity. The control plane between hyperscaler workload schedulers, the datacenter's physical flexibility assets, and the grid market interface. Captured by whichever software or protocol becomes the interoperability standard. SaaS-like margins if the standardisation happens, thin API-layer margins if it does not.

The rest of the piece walks the architectures in order, then returns to the stack to ask which layer captures which scarcity rent under which grid conditions.

04Problem-to-solution mapping

The matrix below maps each approach against the four economic functions.

The matrix confirms the layering established. It also makes visible a distinction the industry usually blurs: Enchanted Rock and Prometheus perform supply-side substitution and clear through capacity markets and PPAs, not through load-shed tariffs. Grouping them with software-coordinated load reduction obscures how differently they are priced.

Function × Approach · which architecture performs which economic functionPrimary = the natural tool. Strong = clear contribution. Marginal = contributes but not the fit. Empty = does not perform.Emerald AISoftwareVerrusBattery-DCEnchanted RockGas microgridPrometheusRenew co-locCloverleaf/RowanSite devCoolingThermal shiftSHEDReduce grid draw on cuePRIMARYSTRONGMARGINALMARGINALSHIFTMove consumption in timeMARGINALPRIMARYSTRONGMARGINALMARGINALSUBSTITUTEReplace grid with on-siteSTRONGPRIMARYPRIMARYRELOCATEMove compute where power is cheapSTRONGPRIMARYMARGINALPrimary toolStrong contributionMarginalDoes not perform

05Why hyperscalers curtail: the interconnection-hedge math over the energy-arbitrage math

Curtailment is often described as demand response against peak electricity prices. On unit economics, that framing does not survive.

An H100 or B200-class GPU-hour is worth $1 to $4 in 2026, whether measured by hyperscaler internal transfer prices or public cloud on-demand rates for the same silicon. Taking that GPU offline for an hour to duck a peak-power event costs the operator its full opportunity value. The peak-power event itself, even during a five-alarm wholesale spike that holds prices from $50 to $1,000 per MWh, saves at most $0.95 to $1.40 per GPU-hour on the throttled kilowatt.

Depreciation on the same silicon runs another $0.89 to $1.25 per GPU-hour (straight-line, four-year life). This is a sunk-capital anchor, silent on the incremental cash cost of an idle hour, but it confirms the direction the opportunity-cost number is already pointing.

Load-shed as energy arbitrage is at best break-even on marginal economics. Once you price in the training-deadline slippage from taking accelerators offline mid-run, it is structurally value-destructive. The framing is wrong.

So why do hyperscalers curtail?

Time-to-power. That is what curtailment buys.

Look upstream of the meter. Four to seven years of accelerator throughput waits on gravel lots while the utility works through its interconnection queue. If accepting 200 hours per year of throttled operation compresses that timeline by three years (from six years to three), the hyperscaler unlocks three years of otherwise-unsold GPU capacity. On a 100 MW AI training cluster (~100,000 GPUs at 1 kW each including overhead), at $1 to $4 per GPU-hour of opportunity cost, that pull-forward runs into the billions per site.

That is the economic engine large enough to support Emerald AI's $1.05B valuation. A probabilistic underwriting model for the architectures is published at adikumar.co/flex-model/. It also explains why the platform sells to the datacenter operator and not to the utility. The value accrues to the party waiting for MW.

Introducing the kill-shot ratio. The kill-shot ratio (also called the conditional acceleration ratio) tells the operator how much interconnection pull-forward capacity each dollar of curtailment cost buys, in present-value terms over the contract term. It is scale-invariant in cluster MW and $/GPU-hour (both cancel out) and moves only with four operational assumptions: acceleration years unlocked, throttle depth, curtailment hours per year, and contract term. Formally: (accel × 8760) ÷ (throttle × hours × years). A ratio of X:1 says every $1 of curtailment cost the operator handles unlocks $X of accelerated capacity value.

The conditional acceleration asymmetry. Model it explicitly on a 100 MW cluster. Assume a $1.50 per GPU-hour net contribution after opex and depreciation. That is the conservative midpoint of the opportunity-cost range.

Collapsing the interconnection queue by three years (26,280 hours) pulls forward:

• 100,000 GPUs × $1.50/hour × 26,280 hours ≈ $3.9B of gross contribution capacity pulled forward

The $3.9B is a throughput-value upper bound, not NPV, revenue, or incremental cash flow. It reflects gross contribution capacity that earlier energisation makes accessible, subject to actual utilisation and monetisation.

The cost of the curtailment commitment that unlocks it, at 200 hours per year of 30% throttling across a 10-year contract:

• 100,000 × 0.30 × 200 × 10 × $1.50 ≈ $90M cumulative opportunity loss

Ratio: ~44:1. The exposed inputs are the four operational assumptions. The $/GPU-hour is a red herring. The ratio simplifies to (acceleration × 8760) ÷ (throttle depth × curtailment hours/yr × contract years) = (3 × 8760) ÷ (0.30 × 200 × 10) = 26,280 ÷ 600 = 43.8. Both the $/GPU-hour figure and the GPU count divide out. The ratio holds at 44× whether the cluster is 10,000 GPUs at $0.50/GPU-hour or one million GPUs at $5. The absolute dollar figures ($3.9B pull-forward, $90M cost) scale with $/GPU-hour and GPU count; the ratio does not. The ratio depends on the structure of the trade; dollar scale drives a different metric.

Sensitivity of the ratio to the four operational drivers. Each row varies one variable, holding the others at base case.

Ratio caveats before you cite the 44×. The pull-forward calculation assumes 100% GPU utilisation across the recovered window; realistic utilisation is lower. Curtailment cost values throttled hours at the average $/GPU-hour, but curtailment correlates with grid peaks, which correlate with high-value compute, so real curtailment cost runs above the mean. The $3.9B is undiscounted; NPV would apply a discount rate to future contribution. And 100,000 GPUs on 100 MW assumes 1 kW/GPU including overhead; at PUE 1.3 for H100 / B200-class hardware, a realistic 100 MW facility holds 65,000-80,000 GPUs, not 100,000. Adjust for these and the absolute dollar figures shrink materially. The ratio holds provided the four operational assumptions hold.

Queue compression is a thesis-level assumption. If ISO reforms shrink interconnection queues from 4-7 years down to 2-3 years across major markets, every architecture in this essay loses part of its scarcity premium. Cloverleaf and Rowan take the direct hit on land inventory value. Emerald's 44× also compresses because a smaller acceleration window shrinks the numerator. Queue duration is the exposed macro variable across the whole flexibility-right thesis.

The 44× describes one economic regime. The ratio is the output of one specific scarcity regime: three years of acceleration, 30% throttling, 200 curtailed hours per year, and a ten-year commitment. Under a one-year acceleration and heavier curtailment, the ratio falls to 7.3×. Under queue reform that compresses interconnection scarcity broadly, the numerator shrinks before any technology moves. The more durable insight is that the value of flexibility is convex to interconnection scarcity. That convexity is what makes the trade attractive today. It is also why queue reform is the common-mode bear case across every architecture in this essay.

In interconnection-constrained regions, load-flex becomes a P&L input for hyperscalers. Flexibility is one of the few ways to work around a scarce queue.

So what. Any investor sizing the load-flex opportunity should benchmark against the interconnection-acceleration pull-forward, and stress-test the acceleration assumption first. Avoided energy cost is a rounding error against the pull-forward number.

UNDERWRITING LEDGERPublic · NVIDIA Blackwell server ASP ranges from public reseller pricing (~$250k-350k for 8-GPU HGX B200 platforms in 2026). NVIDIA H100 SXM5 draws 700 W; Blackwell B200 draws 1,000-1,200 W per package.Derived · GPU-hour depreciation and marginal power cost derived from the ASP + 4-year straight-line + per-GPU-kW draw at 2026 wholesale price bands.Interpretation · The load-shed transaction only clears when it is priced as an interconnection-speed hedge. Energy-arbitrage framings can understate the economic value by orders of magnitude when the flexibility right accelerates time-to-power.

06Approach 1: Software workload coordination · Emerald AI

Emerald AI is the software layer that reads a grid stress signal, translates it into a curtailment request, and coordinates AI training workloads across a datacenter so that compute output drops on cue without breaking the training job. The product name is Conductor. The bet is that AI training is elastic enough at the workload-scheduler level to shed load for hours at a time, and that hyperscalers will accept the throughput hit in exchange for faster grid interconnection and lower firm-capacity charges.

Named deployments.

• Phoenix field trial (published July 2025). 256-GPU NVIDIA A100 cluster orchestrated through Databricks MosaicML. 33 experiments, 212 individual training/inference/fine-tuning jobs across SRP, APS and CAISO sample grid events. Headline: 25% power reduction sustained for three hours on a peak-stress day (Colangelo et al., arXiv 2507.00909). Partners: Salt River Project, Arizona Public Service, NVIDIA, EPRI DCFlex programme.
• Aurora AI Factory (H1 2026). 96 MW facility in Manassas, Virginia. Digital Realty. Running against PJM signals.
• Vera Rubin AI Research Factory (H2 2026). ~100 MW. EPRI, Dominion Energy, PJM Interconnection (Latitude Media).

Emerald has multiple announced pilots and integration partnerships. None yet at cross-hyperscaler production scale.

Capital and investor base. Emerald closed its Series A in late August 2026, co-led by Energize Capital and DCVC. The round was reportedly $150 million at approximately $1.05 billion post-money (VentureBeat). Participants include NVIDIA, Radical Ventures, Samsung Ventures, Siemens and GE Vernova. An earlier strategic round in March 2026 added Eaton, reportedly at approximately $25 million (Emerald blog). Twelve Fortune Global 500 companies are now on the cap table. Total capital raised is over $220 million across seed, strategic round and Series A per Emerald AI's own disclosure.

Eaton, Siemens and GE Vernova are the three largest incumbents in grid infrastructure and power conversion. Their strategic checks say the utility side has decided load-flex has commercial substance. NVIDIA's participation says the compute-side counterparty is aligned. Samsung Ventures is the memory and packaging counterparty. The absence of a hyperscaler direct investor (Google, Meta, Microsoft, AWS) is a notable gap. Hyperscalers view cluster scheduling (Borg, Apollo, and their internal equivalents) as core proprietary infrastructure, and letting a third-party control plane inject checkpoint-and-throttle primitives across proprietary interconnect topologies crosses a boundary they have not chosen to cross yet. Emerald's addressable market in the near term is therefore neo-clouds (CoreWeave, Lambda, Crusoe) and wholesale colocation landlords (Digital Realty and comparable REITs) that lack proprietary scheduler stacks, and this piece returns to the question.

Business model. The commercial contract has not been fully disclosed. Public statements describe a platform-fee model paid by the datacenter operator, with revenue sharing tied to capacity-market receipts or interruptible-load tariff credits generated by the curtailment behaviour. The Conductor platform is deployed as a software layer between the utility SCADA signal and the hyperscaler workload scheduler.

UNDERWRITING LEDGERPublic · Series A round disclosed late August 2026 (VentureBeat, Emerald blog). Round size reportedly $150M. Post-money valuation approximately $1.05B. SRP Phoenix demo results published July 2025 by EPRI. Two named 2026 deployments (Aurora, Vera Rubin) with Digital Realty and EPRI respectively.Derived · The $1.05B post-money prices Emerald at a ~5× multiple of cumulative paid-in capital (~$200M). With platform revenues running well under $10M ARR across early sites, this is an option-value multiple on category creation (not an operating SaaS multiple). The forward valuation clears only if commercial-scale hyperscaler adoption arrives within 24 months.Assumption · Platform take of 15-30% of stack flexibility value at reference deployments. Hyperscaler curtailment repeatability across 20+ sites within 24 months. No hyperscaler-internal coordination layer standardises first.Interpretation · The valuation is defensible if the SRP demo behaviour transfers to production Nvidia training loads at commercial scale. The absence of a hyperscaler cap-table check is the risk signal.

Emerald AI Conductor · software coordination architectureGrid signal in, workload throttle out. Emerald owns no physical infrastructure.UTILITY / ISOGrid stress signalSCADA / DR event /wholesale priceEMERALD CONDUCTORSoftware platformTranslates grid state intocurtailment scheduleHYPERSCALER APIWorkload schedulerKubernetes / Slurm /custom orchestratorTRAINING JOBLoad throttle100% → 75% for3 hours (SRP demo)Measured curtailment reported back for capacity-market settlementWHAT EMERALD OWNSThe software layer between utility signal and workload scheduler. No physical infrastructure.Revenue model: platform fee + revenue share on capacity-market receipts. Deployed at Aurora AI Factory (96 MW, Manassas, H1 2026) + Vera Rubin AI Research Factory (~100 MW, Dominion territory, H2 2026).

07Approach 2: Battery-integrated datacenter · Verrus

Verrus takes the opposite architectural bet from Emerald. Instead of coordinating curtailment through software, Verrus designs the datacenter as a battery-first microgrid so that load can be shifted physically without touching the compute schedule. The workload never sees the grid signal. The battery handles the arbitrage or peak-shave window and the training job runs at nameplate throughout.

Technical design. Each Verrus datacenter building holds 50 MW of compute, 20 MW of other electrical demand (HVAC, lighting, controls) and a co-located utility-scale battery sized at 70 MW nameplate for four hours of discharge, giving 280 MWh of storage capacity per building (TechCrunch). The battery is the primary buffer between the grid tie and the compute load. Software allocates energy to specific tasks in real time. The datacenter can run at nameplate compute for hours while drawing zero net grid power, provided the battery is charged.

Named deployments.

• Stealth exit March 2024 as a Sidewalk Infrastructure Partners venture (Alphabet-affiliated spinout).
• First three sites: Arizona, California, Massachusetts. Target ops late 2026 / early 2027.
• Customers: Not publicly named.

Capital. Verrus is backed by SIP's balance sheet (not by a marked venture round). Exact capital-in has not been publicly disclosed. Site development at the announced scale requires nine-figure equity plus project debt, so the actual capital commitment is much larger than the announced Series total for Emerald.

Business model. Verrus builds, owns and operates the datacenter. Revenue arrives as long-term datacenter lease payments from hyperscaler and enterprise tenants, with an implicit uplift for the flexibility features. The battery generates a second revenue stream from wholesale energy arbitrage plus grid ancillary services participation in the ISO market covering each site. The economics rely on hyperscaler willingness to pay a rent premium for a flexibility-optimised building versus a conventional one.

08Approach 3: Behind-the-meter gas microgrid · Enchanted Rock

Enchanted Rock is the incumbent in behind-the-meter natural gas microgrids for large commercial and industrial loads. The company deploys modular gas-fired gensets at customer sites, operates them as a microgrid that runs during grid outages or grid-stress events, and participates in interruptible-load and capacity-market programmes on behalf of the customer. The bet is old and proven in traditional C&I: what is new is that hyperscalers have now adopted the model at gigawatt scale for AI datacenters.

Named deployments.

• Meta El Paso. El Paso Electric to contract 813 modular gensets. ~366 MW of behind-the-meter firm power (El Paso Matters).
• Microsoft San Jose. Renewable-natural-gas microgrid with Enchanted Rock and U.S. Energy. Covers grid outages and California BIP events (U.S. Energy).

Both are contracted deployments at named production sites.

Business model. Enchanted Rock operates on an Energy-as-a-Service contract. The customer pays a fixed monthly capacity fee plus a variable fee tied to actual dispatch events. Enchanted Rock owns and operates the gensets, procures the fuel, and monetises capacity-market participation on top of the customer contract. The economic case strengthens as annual dispatch hours rise, and the required utilisation depends on the tariff structure, fuel cost, capacity payment and the customer's avoided grid cost. Contract-specific fee structures stay undisclosed.

Capital. Enchanted Rock is a growth-stage private company with disclosed multiple hyperscaler contracts and undisclosed revenue. Its balance sheet finances gensets against long-term customer contracts; the underlying financial profile resembles a specialised infrastructure operator more than a venture-backed software company.

09Approach 4: Renewable and battery co-location · Prometheus Hyperscale

Prometheus Hyperscale builds datacenters that sit physically next to renewable generation and battery storage assets, tying the compute load to the on-site energy source as the primary supply and treating the grid connection as a firming and export path; the primary feed is on-site renewables. ENGIE North America is the co-development partner and provides the renewables and storage assets under a shared platform (Barchart).

Named deployments.

• Texas I-35 corridor. First sites targeted 2026, expansion through 2027.
• Compute capacity per site: Not publicly disclosed.
• Commercial structure: Long-term PPA with ENGIE-operated renewables + datacenter lease from Prometheus.

Model and economics. A typical site combines solar plus wind plus battery storage at a scale that covers 60-80 percent of the datacenter's annual energy demand on a matched basis, with grid draw covering the remainder. The battery smooths renewable intermittency and captures ancillary-services revenue in ERCOT. During grid stress, the site dispatches battery capacity to the grid (not drawing from it). Commercial structure: ENGIE takes renewables IPP economics, Prometheus books the datacenter lease, flex value accrues to whichever entity holds the grid interface. The siting thesis is to co-locate at renewable generation nodes to shorten effective interconnection time and reduce the LCOE of the compute load.

10Approach 5: Interconnection optionality · Cloverleaf and Rowan

Cloverleaf Infrastructure and Rowan Digital Infrastructure do not sell flexibility. They sell the option to become flexible later. Both companies develop hyperscale datacenter campuses on land that is pre-permitted for high-MW grid interconnection and pre-designed to accept future flexibility retrofits (battery microgrids, gas gensets, PPA hookups) that a tenant hyperscaler may want to add. The primary product is a shovel-ready site with power infrastructure in place. Flexibility features enter the site plan as optionality; Cloverleaf and Rowan do not run them as operating layers.

The economic distinction matters. Cloverleaf and Rowan operate outside the load-flex value chain that Emerald, Verrus, Enchanted Rock and Prometheus occupy. They are one layer beneath: the interconnection-ready substrate on which the flexibility infrastructure of the other four approaches gets built. Land inventory, permitting, substation capacity, and interconnection-queue position are what these companies own. As the ISO queues stretch through the 2026-28 window, that substrate becomes progressively scarcer and its unit economics compress upward.

Named deployments.

• Cloverleaf × NVIDIA (21 Aug 2026). Strategic partnership on US AI datacenter development, including utility and energy-provider integration (DataM Intelligence). Multiple gigawatt-scale projects in development.
• Rowan Digital Infrastructure. Operating across 20 US states with build-to-suit hyperscale campuses. Quinbrook-backed.
• Rowan Cinco (Texas). 300 MW campus under construction (rowan.digital).

Business model. Cloverleaf and Rowan operate as datacenter real-estate developers with a power-infrastructure investment layer. Revenue arrives as long-term hyperscaler lease payments plus a return on the power infrastructure. Flexibility features (space for a future battery, access to a gas lateral, sub-station capacity for a future PPA hookup) show up as options (direct revenue stays unrecognised until deployed).

Technical model. The differentiator versus a conventional datacenter developer is the deliberate pre-provisioning of headroom: land for future storage, electrical layout that can accept a microgrid retrofit, tenant contracts that permit third-party flexibility overlays. In practice this is a real-estate underwriting call; energy-system underwriting applies differently.

11Cooling-as-flex: the embedded option

Cooling accounts for 15-30 percent of total datacenter power draw at hyperscale AI facilities, depending on climate and cooling architecture. In principle, the thermal mass of the cooling system offers a flexibility knob independent of compute. Pre-cool ahead of a grid peak, let the datacenter coast at slightly higher inlet temperature during the peak, then run cooling harder afterwards to catch up. The compute workload never sees the flexibility event. The physics works and has been studied in academic literature. The commercial model has not appeared.

Physics and commerce. The physics is well-behaved. Liquid-cooled 130 kW AI racks have thermal time constants of several minutes, whole-hall chilled-water buffers extend that to hours, and campus-scale thermal storage extends it further. The theoretical flex range is a few percent of the datacenter's total electrical load, sustainable for tens of minutes to a few hours, at essentially zero compute impact. Commerce has been the problem. Hyperscaler thermal SLAs are tighter than the cooling-flex window would require, the addressable revenue per site is small relative to compute-side flex, and no merchant vendor has emerged to productise the capability. The functionality is embedded inside DCIM incumbents (Vertiv, Schneider, ABB) as a background control feature.

Where it becomes viable. Cooling-flex works as a layered add-on. A Verrus-style battery-integrated site can dispatch cooling-side flex in parallel with battery cycling. An Emerald AI curtailment event can be paired with a coordinated cooling ramp. In both cases, value accrues to whichever operator owns the site.

UNDERWRITING LEDGERPublic · Cooling as a fraction of datacenter power draw is documented across published PUE studies (typical 1.2-1.5 PUE implies 15-30 percent non-compute).Derived · For illustration, assume cooling represents roughly 75 percent of the 30 MW facility overhead implied by a PUE of 1.3 at a 100 MW IT load. That implies about 23 MW of cooling load, though the actual split depends on architecture, climate, and heat-rejection topology. A cooling-flex event that shifts half that load for one hour yields roughly 12 MWh, worth $500-1,500 at typical wholesale peak prices. That is energy-arbitrage value only. Contractual capacity or reliability value under a bilateral tariff would price differently.Interpretation · The addressable flex value per event is small enough that a standalone cooling-flex platform cannot support venture-scale economics. The idea is worth noodling as a complement to other approaches. It stays conceptual, not yet a company.

Cooling-as-flex · thermal shift concept (physics works, commerce has not)Pre-cool the datacenter ahead of a grid peak. Coast at higher inlet temperature during the peak. Recover afterwards.T-2hT-0 (peak start)Peak (2-4 hours)T+peak endRecoverPRE-COOLChiller ramps upCool below setpointto build thermal buffer+15% cooling loadGRID PEAKChiller ramps downDatacenter coasts on thermal buffer;inlet temperature rises within SLA-50% cooling load (roughly)RECOVERChiller catches upReturn to setpoint+10% cooling loadWHY NO MERCHANT VENDORHyperscaler thermal SLAs restrict extended excursion. Flex range is small (~1-3% of DC total load) versus compute flex (~25%).Capability is embedded in DCIM incumbents (Vertiv, Schneider, ABB) as a background feature, not sold as a standalone product.

12Unit economics: what each approach costs

Each approach has a distinct capital intensity and a distinct revenue mechanism. The economics that decide whether a bet clears reduce to four numbers: capex per MW of flex delivered, annual opex, the revenue stack per MW-year, and the number of dispatch hours per year the approach captures. The rest of this section works those four numbers for each approach at 2026 market prices. Every number below is an author estimate against public sources where sources exist. Each approach's Underwriting Ledger flags which numbers are Public, Derived, or Assumption.

Software coordination (Emerald AI). Capex per MW of flex is near zero at the platform vendor level; the underlying compute and grid infrastructure is already owned by the hyperscaler and utility. Emerald's own opex is R&D and integration engineering, funded from the $220M+ total raise. On the customer side, Emerald's reported model is a platform fee plus a revenue share on capacity-market receipts. Author estimate of the revenue stack: $20-50 per kW-year of flex delivered (composed of PJM capacity-market clearing prices, interruptible tariff credits, and avoided demand charges), of which Emerald captures 15-30 percent as platform revenue. At 100 MW of coordinated flex, the arithmetic runs: 100 MW × $20-50/kW-year = $2-5M stack flex value, of which Emerald captures 15-30% = $300k-$1.5M annual per-site Emerald revenue. Dispatch hours are grid-driven, typically 50-200 hours per year in PJM.

Battery-integrated DC (Verrus). Capex per MW of flex is dominated by the battery. At 2026 installed prices ($150-200/kWh), a 70 MW / 4-hour battery costs $42-56 million per building. The datacenter shell adds $8-12 million per MW of compute capacity, standard for large-load AI-ready construction. Opex is dominated by battery cycling degradation (roughly 3-5 percent of installed capex per year at heavy cycling) and site O&M. Revenue: datacenter lease at $100-140/kW-month (comparable to Cloverleaf, Rowan) plus battery arbitrage of $30-80 per kW-year (ERCOT ancillary and wholesale energy arbitrage at 2026 prices). The combined per-building total site revenue (datacenter lease plus battery arbitrage) is on the order of $70-100 million annually before financing. Only a portion of that number is attributable to the flexibility layer itself; the payback question is whether the flexibility premium in the lease rate covers battery capex, and neither the premium nor the capex is publicly disclosed.

Gas microgrid (Enchanted Rock). Capex per MW of firm capacity is $700-900/kW installed, plus fuel infrastructure and site work. Opex is variable and dominated by gas fuel cost during dispatch (roughly $80-140/MWh at 2026 Henry Hub prices plus RNG premium where applicable), balanced against per-event dispatch fees paid by the customer. The revenue stack is the cleanest of the six approaches: a fixed monthly capacity fee ($8-15/kW-month equivalent, contract-dependent), plus per-event dispatch fees, plus capacity-market receipts on top. Author estimate: at contracted volumes, the Meta El Paso installation could imply annual EaaS revenue in the tens of millions. The commercial contract economics stay undisclosed.

Renewable co-location (Prometheus + ENGIE). Capex is the heaviest of the six. Solar plus wind plus battery at the scale required to match 60-80 percent of a 100 MW datacenter's energy demand runs $400-600 million per site. That capital is deployed through project finance and long-term PPAs (venture equity is the alternative wrapper), so the equity check per site is smaller than gross capex, but the return horizon is 15-25 years. Revenue splits across three parties. ENGIE takes the renewables IPP economics (project IRR of 8-12 percent unlevered at 2026 wholesale prices in ERCOT). Prometheus books datacenter operator economics at conventional AI-DC lease rates. Flex value is a small overlay on top and accrues to whichever entity owns the grid interface.

Flex-ready site development (Cloverleaf, Rowan). Capex per site is land acquisition plus power infrastructure (substation, medium-voltage distribution, interconnection agreements, permitting, site work), typically $500k-$2M per MW of interconnection capacity depending on region and how much utility-side network upgrade the tenant handles. This is a developer model, so returns compound through leasing the pads to hyperscalers at $100-140/kW-month over 10-20 year terms. Flex-ready design adds perhaps 3-8 percent to construction cost for pre-provisioning. Flex-optionality does not directly generate revenue for the developer; it accrues to whichever tenant retrofits and monetises it.

Cooling-as-flex. No merchant capex or revenue model exists to compute. The value accrues to whichever operator owns the site. Illustrative revenue: 12 MWh per event at wholesale peak prices yields $500-1,500 per event, or roughly $50k-150k annually at 100 events per year on a single 100 MW site. That does not support venture-scale economics.

Quick reference · unit economics per approach (2026):

Capex composition per approach ($ per MW of flex delivered)Where the capital is positioned determines what kind of investor prices each approach.Software (near-zero)BatteryBuilding shellLand + interconnectGas gensets + fuelRenewables (solar/wind)Embedded featureSoftware coordinationEmerald~$0/MWCooling-flexembeddedNoneGas microgridEnchanted RockGas~$800k/MWFlex-ready siteCloverleaf, RowanShellLand + power infra $500k-2M/MW interconn$500k-2M/MWBattery-integrated DCVerrusBattery $42-56M / buildingShellInfra$600-800k/MWRenewable co-locationPrometheus + ENGIESolar + wind (200-300 MW)BatteryDC$4-6M/MW DC

13Cross-cutting comparison

The six approaches compared on the dimensions that decide investability: capital intensity, revenue mechanism, defensibility, dependency on external counterparties, and the specific grid conditions each one requires. The economics cells represent indicative economic value associated with the relevant layer (not directly comparable company revenues). Read the comparison as capture potential across layers, not like-for-like revenue multiples. Cells use a single denominator: $/MW-year of bankable revenue flowing to whichever entity holds the flexibility right. Absolute magnitudes vary widely across rows because different businesses monetise different rights (substrate rent, arbitrage, capacity payments, platform take). Read the table as a directional underwriting screen.

The table reads signals immediately. First, capital-at-risk spans roughly four orders of magnitude across the six approaches, although the denominators differ; Emerald is the only obvious venture-equity-native business model; everything else depends primarily on project finance, infrastructure or real-estate balance sheets. Second, no single approach hits the intersection of low friction, low capex, and high defensibility; every bet trades one of the three for the other two. Third, the software layer (Emerald) has the sharpest defensibility risk because its value is bracketed above by hyperscaler willingness to curtail and below by utility willingness to pay.

14Where value flows: layer economics

Value in a stacked architecture accrues by layer, and the layers have different margin structures.

• Land and interconnection (Cloverleaf, Rowan): the substrate. Margin in this layer comes from real-estate leases plus power-infrastructure returns. Defensibility comes from the queue position and the permitting that took two to five years to assemble. Overall economics track datacenter REIT unit economics, high single digits to low teens on invested capital.
• Physical flexibility infrastructure (Verrus, Enchanted Rock, Prometheus + ENGIE): the flex chassis. Value accrues to the balance sheet that owns the batteries, gensets or renewables. Margins are infrastructure-project-style, low double digits IRR, and the defensibility comes from long-term customer contracts.
• Software coordination (Emerald AI): the flex control plane. Value here is SaaS-like at the platform level but is capped above by hyperscaler willingness to curtail and capped below by utility willingness to pay for load-flex through capacity payments. Defensibility depends on becoming the interoperable standard between hyperscaler schedulers and grid operators.
• Cooling-side flex: embedded feature, not a layer. Value accrues to the operator that owns the DCIM.

The reading matters for two reasons. First, treating Emerald and Verrus as competitors is a mistake; they solve different sub-problems and are more likely to be layered together than to displace each other. Second, the venture-market pricing of $1.05B for Emerald bakes in the assumption that the coordination layer captures the largest share of stack economics, and the empirically obvious outcome differs. Enchanted Rock at the physical-infrastructure layer has more revenue evidence today than any other approach in the shortlist, and it is a private company without a comparable valuation event.

Third, the marginal value of any layer depends on what already exists at the site. Software coordination plus battery inertia yields more than software alone. Software plus a co-located gas microgrid yields more than either independently. Behind-the-meter generation plus workload orchestration yields more than either in isolation. The stack composes layered value; substitution between layers stays limited. Underwriting any one in isolation misses what the layered site earns. The strategic question becomes who owns the coordination between the layers, and who captures the rent from that coordination.

15Two hard underwriting questions

Questions decide whether the category is investable. Every valuation event to date, including Emerald's $1.05B post-money, prices in a specific answer to both, and the evidence remains thin.

Q1. Will hyperscalers curtail production training at commercial scale?

The Salt River Project demonstration in May 2025 showed a 25 percent load reduction sustained for three hours on a working AI datacenter cluster. That remains the only publicly documented field deployment at commercial datacenter scale. Beyond it, the record is a set of controlled tests and hyperscaler policy statements.

The reason the question is hard: a curtailment event on a live training job either interrupts a checkpoint (small compute waste, easy recovery), throttles the gradient step (larger loss, harder recovery), or forces a partial reschedule of the run across GPUs (worst case, potentially days of retraining). Which one happens depends on the specific curtailment protocol negotiated between Emerald's Conductor and the hyperscaler's workload scheduler. Hyperscaler CTOs have published willingness-to-curtail commitments; those commitments have not been repeatedly tested against production workloads in the way that the SRP demo tested one site on one day.

Repeatability across 20+ deployments decides it: different hyperscalers, different workload types (training vs inference), different grid-stress patterns (heat waves, winter storms, forced generator outages). If Emerald's Aurora and Vera Rubin deployments in 2026 produce that repeatability, the $1.05B post-money is defensible. If not, the software layer margin compresses toward zero, because the platform ends up as an SRP-demo lookalike; a hyperscaler-wide standard would be a different outcome.

The inference transition could materially change the economics. Training clusters have massive sunk capex and tightly coupled all-reduce communications across InfiniBand or NVLink, which makes ad-hoc throttling expensive and operationally risky. Inference workloads are stateless, geographically routable, and decoupled at the request level. If inference becomes a larger share of datacenter MW through 2026-2028, Emerald's coordination problem shifts from complex scheduler checkpointing to token routing and batch-latency relaxation. Curtailment on inference is orders of magnitude easier to operationalise than curtailment on foundation-model pre-training. The valuation case gets easier as the workload mix shifts.

The technical hurdle runs deeper than checkpoint-and-resume. Modern large-model training uses 3D parallelism (data, tensor, pipeline) across hundreds or thousands of nodes with tightly coupled all-reduce communications over InfiniBand or NVLink. A curtailment event that pauses a subset of nodes risks stranding the interconnect fabric on the remaining ones, or forces a partial reschedule that costs hours of retraining. That is different from throttling GPU frequency, which Conductor's Phoenix demo used at A100 scale. Production training with 3D parallelism at Blackwell scale has a higher failure surface than the demo covered.

The physical plant also imposes a ramp-rate ceiling. Transformer, UPS, cooling-loop and accelerator thermal dynamics mean the commercially useful response window is likely measured in minutes (seconds-level response overstates the physics). The sub-minute cadence a naive software model assumes overestimates what the physical plant can cycle. Repeatable throttling frequencies live behind silicon reliability curves that Nvidia and hyperscaler reliability teams have not publicly quantified for this use case, and the ceiling is architecture- and reliability-dependent.

Q2. Who captures the flexibility value between platform, hyperscaler, and utility?

Flexibility value at a 100 MW site plausibly clears at $2-15 million per year depending on grid conditions and dispatch frequency. No published tariff yet exists for AI training curtailment as a resource category. Three parties want that value.

The hyperscaler contributes the compute asset and takes the throughput hit during curtailment. The utility receives the grid-side benefit (avoided peaking, deferred transmission upgrades) and controls the price signal. The software platform (Emerald) mediates between them and enables the transaction. In every deployment structure disclosed to date, the value split has been bespoke. There is no published tariff for "AI training curtailment" analogous to PJM's capacity market clearing price for conventional demand response.

The venture-market bet on Emerald prices in a scenario where the platform captures 15-30 percent of the total flex value across a stack of deployments. That 15-30 percent share is an underwriting assumption. Observed evidence remains thin. That share is possible if Emerald becomes the interoperability standard between hyperscaler schedulers and grid operators. It is also compressible to 5 percent or less if hyperscalers build the coordination layer internally, or if utilities standardise a direct-to-hyperscaler API that skips the platform.

The hyperscaler build-vs-buy scenario tree. Hyperscalers do not outsource cluster scheduling. Google runs Borg; Microsoft runs Apollo variants; AWS runs proprietary orchestration. Extending Borg to ingest OpenADR or utility SCADA signals is a normal engineering-sprint effort for a hyperscaler platform team. The historical precedent is instructive: startups tried to sell software-defined WAN routing across hyperscaler backbones, and each hyperscaler built its own (Google B4, Microsoft SWAN, AWS Global Backbone). Global control-plane software of any kind rarely survives inside hyperscaler proprietary stacks.

Emerald has two realistic terminal outcomes. First, Emerald becomes the merchant operating system for tier-2 GPU clouds (CoreWeave, Lambda, Crusoe) and wholesale colocation landlords (Digital Realty, Equinix, CyrusOne) that lack proprietary scheduler stacks. Terminal state: independent SaaS company at $1-3B enterprise value depending on standardisation. Second, strategic acquisition by a grid-infrastructure incumbent (Schneider, Eaton, Siemens, GE Vernova) that bundles Conductor into DCIM and switchgear product lines. Terminal state: acquisition at $500M-$2B depending on adoption evidence.

Neither terminal path runs inside Google or Amazon. Underwriting Emerald as a category-standard platform across all hyperscalers overstates the addressable market; the defensible frame prices it as the merchant/tier-2 standard plus a strategic-acquisition candidate.

16Regulatory dependency by approach

The regulatory question about load-side flexibility is no longer whether the grid recognises flexible large loads. FERC materially advanced that question on 18 June 2026 with six Section 206 show cause orders directing PJM, MISO, SPP, CAISO, ISO New England and NYISO to justify or reform their tariffs for large-load interconnection, with specific focus areas including transmission services for flexible large loads and co-located behind-the-meter generation (McGuireWoods summary; Utility Dive). The unresolved question is narrower and more consequential: how much accredited capacity does a particular AI load receive, under what measurement and performance rules, and who captures the resulting value.

The specific instruments that will hold the answer include Conditional Firm Interconnection Agreements (letting a load interconnect at a higher nameplate against a lower firm commitment), ERCOT's Large Flexible Load protocols under evolution in 2026, PJM demand-response participation frameworks that pre-date the show cause orders, and CAISO Rule 24 on baseline construction for participating demand response. The tariff filings that respond to the FERC orders through H2 2026 and H1 2027 will decide how each of these instruments translates a technical flexibility claim into a bankable capacity credit.

Technical flexibility versus bankable flexibility

Technical flexibility is what the machine can do. Bankable flexibility is what the grid contract recognises, measures and pays for. They measure different quantities, and the gap between them is where most of the underwriting risk concentrates.

Emerald AI's Phoenix field trial demonstrated a 25 percent load reduction sustained for three hours on a 256-GPU AI cluster running production training and inference workloads. That is technical flexibility. Whether the same site earns 25% x 3h of accredited capacity in PJM or CAISO depends on separate measurements: baseline construction, telemetry standards, response time, minimum dispatch duration, availability windows, performance penalties, and the specific tariff or auction the resource participates in. None of these are settled for AI training curtailment as a resource category. FERC's June 2026 show cause orders put those questions into formal tariff proceedings across the six ISOs.

The bridge between technical and bankable flexibility is the underwriting variable most likely to compress or expand valuations in this space over 2026-2028.

There is a political dimension the accreditation debate does not always surface. PUC proceedings are increasingly contested by consumer-advocate groups arguing that flexible large loads should not be permitted to jump the queue if consumer reliability could degrade during forced curtailment, or if generation costs get socialised across ratepayers. Texas has already paused new large-load connections pending an audit. The bankability of flexibility depends on PUC willingness to grant preferential treatment for large flexible loads. That willingness is politically contingent.

The breaker-trip test. For an underwriter, the critical enforcement question is: what happens if the customer does not perform? I call this the breaker-trip test. The answer need not be a literal breaker trip; it can be automated supervisory control, telemetry-backed enforcement, collateral, or financial penalties calibrated above the customer's completion incentive. The resource cannot be bankable if the promised response is economically optional at the moment of dispatch. Utilities preparing to accredit AI training curtailment as capacity are asking the same question in mechanical terms: what physically forces the load off if the software says no. A hyperscaler mid-run on a foundation model 99 percent complete has a strong incentive to eat a $50,000 per MWh liquidated-damages penalty; letting the run crash is the alternative outcome. Toothless penalties mean the ISO cannot count the flexibility as capacity; punitive penalties mean the hyperscaler CFO will not sign the interconnection agreement. Bankable flexibility requires an enforceable mechanism the grid operator can rely on under failure conditions. In some architectures that may be a physical relay-actuated breaker trip; in others, automated supervisory controls, contractual penalties, and telemetry may be sufficient. The breaker trip is my proposed floor test. Regulatory doctrine has not settled here.

So what. Software-only architectures without an enforceable fallback face a valuation ceiling in the accreditation process regardless of technical demonstration quality. The specific mechanism (relay trip, automated controls, financial penalties calibrated above the hyperscaler completion incentive) is architecture-dependent, but the requirement for enforceability stays constant.

The specific accreditation and market participation rules vary across US ISOs, EU markets, and other regulatory regimes.

Two accreditation outcomes restructure the space through 2027. First, ISO-specific tariff filings responding to the June 2026 FERC show cause orders will define measurement, baseline construction and dispatch performance rules for flexible large loads. Those definitions determine how much of Emerald's technical 25 percent load-shed converts into accredited capacity in each ISO. Second, tightening of state-level emissions rules on aggregated gas gensets (in Texas or Virginia specifically) would compress Enchanted Rock's addressable market for new deployments. In parallel, PJM's Expedited Interconnection Track approved by FERC on 9 June 2026 (10 large generation projects per year, 10-month interconnection agreement target, 3-year commercial-operation target) reforms the generation side of the queue but does not directly speed up large-load connections (PJM Inside Lines).

17What would falsify each bet

Every approach has a small number of observable events that would falsify the investment case. Watching for these is more useful than tracking category-level enthusiasm.

• Emerald AI: hyperscaler-internal coordination layer disclosed publicly (Google, Meta, Microsoft, AWS), OR SRP-style demo results that fail to reproduce at Aurora / Vera Rubin, OR utility direct-to-hyperscaler API standardisation that bypasses the platform. Any one would compress the $1.05B post-money significantly.
• Verrus: battery cell prices reverse and rise back above $200/kWh installed, OR ERCOT / CAISO wholesale price volatility flattens enough to eliminate arbitrage, OR a Verrus-designed site opens without a signed hyperscaler tenant at the flex-premium lease rate. The economic case unwinds quickly on any of the three.
• Enchanted Rock: hyperscaler carbon commitments (Google, Meta, Microsoft) accelerate to exclude gas-fired backup at new sites, OR federal or state emissions rules on aggregated gas gensets tighten to make the modular fleet economically infeasible. The Meta El Paso design would remain in operation but new deployments would slow.
• Prometheus + ENGIE: hyperscaler PPA structures shift to require 24/7 renewables matching, which the co-location model cannot satisfy without oversized storage. That would push the model toward the fully-firmed variants that Microsoft and Google have already publicly favoured.
• Cloverleaf / Rowan: ISO interconnection queue reforms materially reduce time-to-interconnection (from 4-7 years down to 2-3 years). Land inventory value compresses in proportion. Queue reforms are being discussed in every major ISO but none is scheduled to take effect before 2027.
• Cooling-flex: already effectively falsified as a standalone investment case; no merchant vendor has emerged and the addressable value per site is too small to support one. Remains a feature-layer add for whichever operator owns the site.

18Emerald at $1.05B: the contrarian view

The technical proof exists. Five documented demonstrations, up to 40% power reduction in under a minute across 200 simulated events, commercial deployment underway at Digital Realty's Manassas site and NVIDIA's planned Vera Rubin facility. The technical mechanism has been demonstrated.

The economic proof is a different question. Emerald discloses deployment counts but not revenue, ARR, contracted MW, gross margin, or backlog. Deployed sites and monetised sites measure different quantities, and MW under Conductor's coordination during a partnership differ from MW under a paying contract.

What contracted MW does $1.05B require?

Start from operations, not from multiples. Emerald earns platform revenue proportional to the MW under Conductor coordination and the share of stack flex value it captures. At a 30% platform take on the essay's $20-50/kW-year flex value, that is $6-15k in Emerald revenue per contracted MW-year. At a 5% take, it collapses to $1-2.5k per MW-year.

Work backward. For $1.05B to look defensible at a 10-15× revenue multiple, Emerald needs $70-105M in annualised recurring revenue. Divide that by the per-MW revenue and you get the contracted MW required:

Emerald's currently announced footprint is Aurora (96 MW) plus Vera Rubin (~100 MW), roughly 200 MW total. The favorable case (30% take, high revenue per MW) requires 25-60× the current footprint under contract by 2029. The unfavorable case (5% take) fails the $1.05B underwriting hurdle under any reasonable multiple, regardless of MW growth. The take rate and the MW scale multiply. Both have to hold.

Cross-check: what ARR does that imply?

The MW frame converts to standard software multiples cleanly:

Even at 25× revenue (bubble-level for infrastructure software), Emerald needs roughly $42M in annualised recurring revenue for the mark to hold. There is no public evidence Emerald is anywhere close to those numbers today.

Emerald's solution

Conductor is a lightweight orchestration layer. The central Python service runs in Emerald's cloud and sends throttle commands into the customer's datacenter. Lightweight agents inside the customer environment execute the commands by calling the customer's own workload scheduler (MosaicML, Slurm, Kubernetes, or proprietary orchestrators) via public API. Conductor works alongside those schedulers.

Hyperscalers routinely adopt third-party SaaS in early stages (Datadog, Splunk, PagerDuty, Databricks), then build internal replacements once a function becomes core to operations. For the $1.05B mark to clear, Emerald has to remain the orchestration layer at gigawatt scale, over many years. That is the underwriting problem.

One caveat sharpens the risk without eliminating it. The Borg/Apollo precedent is about cluster scheduling. Emerald's actual capability includes ISO tariff arbitrage, utility settlement mechanics, and measurement/verification against grid baselines. Those are different competencies from what hyperscalers built internally for job scheduling. Building the utility-facing rights layer is a different problem from building Borg. That is one reason the category-standard scenario probability is above zero (zero probability would overstate the certainty).

The four terminal scenarios

The scenarios in text. Chart-anchored numbers converted to prose for reader auditability:

My personal call. $1.05B is roughly fair as an acquisition-modal bet. The mark makes sense if you believe grid-infrastructure incumbents (Schneider, Eaton, Siemens, GE Vernova, all on the cap table) will pay $500M-$2B to bundle Conductor into their DCIM offerings within 3-4 years. It does not make sense as a pure standalone SaaS bet unless you rate the hyperscaler-standardisation scenario at over 25%. My personal weight on that scenario is 15%. The category-standard upside is genuine but low-probability. The base case is that Emerald gets acquired, and the mark reflects that base case adequately.

The bull case

Emerald's bull case is coordination. Software becomes the layer that dispatches across batteries, generators, and interconnection infrastructure, without needing to replace any of them. That is the case for a real tollbooth: the coordination layer sees every form of flexibility on a site and decides which one to spend for which grid event. Verrus's battery, Enchanted Rock's gas microgrid, Prometheus's renewables buffer, cooling-flex, and Emerald's own curtailment API can all be conducted by one platform if a platform earns the right to conduct them.

There is also a time-arbitrage argument that cuts against the bear. In the 2026-2028 window, securing 100-300 MW of AI capacity 24 months earlier is worth billions in terminal enterprise value to a hyperscaler racing frontier compute to market. Even if the coordination function gets internalised eventually, hyperscalers have every incentive to pay a software tollbooth now to unlock that acceleration. Emerald's near-term revenue window may hold 3-5 years even in the bear terminal case. That changes the return math on the Series A.

Being the coordination layer and owning the coordination rent are different things. Whether Emerald keeps the rent depends on whether it becomes the standardised interface for the industry, or whether hyperscalers, utilities, NVIDIA, and switchgear OEMs claim it first.

System value, site value, and vendor equity are three different things. A $100 million system flexibility benefit does not imply a $100 million software TAM. The benefit splits across the hyperscaler occupying the site, the utility avoiding grid capacity spend, the infrastructure owner holding the asset, and the flexibility provider running the response. Financing captures its own share on top. The investment question is which of those parties owns the contractual claim that captures the specific slice.

Who else can capture the value

Even if flexibility becomes a large market, Emerald has to own the tollbooth. Four parties can plausibly capture the value instead.

• NVIDIA. Emerald's Conductor is already integrated with NVIDIA DSX Flex. NVIDIA owns the compute layer of AI infrastructure. If flex becomes an essential AI-factory feature, NVIDIA has strong incentives to productise the coordination layer directly inside DSX. That would turn Emerald into a component supplier of NVIDIA's stack rather than the category owner.
• Utilities. ERCOT's flexible-load protocols and PJM's participation frameworks are being written to accept flexible loads directly. A utility can standardise an OpenADR API to hyperscalers and skip the platform vendor layer entirely.
• Battery operators. A 100 MW / 4-hour battery produces the same firm flexibility a utility wants, and does it without asking the hyperscaler to sacrifice compute. Contractually cleaner than an agreement that requires training runs to pause. Emerald competes for the same flexibility wallet as physical infrastructure that already has financing precedent.
• Hyperscalers themselves. Once the concept is proven, hyperscalers can integrate a curtailment scheduler into their own orchestration. The tech is not especially hard once someone has shown it works at commercial scale. That is the free-option problem: big players benefit from Emerald proving the concept, then decide whether to build, acquire, or partner.
• Switchgear OEMs. Schneider Electric, Eaton, Siemens, and GE Vernova are baking grid-orchestration telemetry directly into substation switchgear and battery control units at the hardware level. That threatens to reduce a software coordination layer to a translation shim between hardware-native orchestration and workload schedulers, compressing its margin structure toward hardware-vendor economics.
• Hardware-native power fabrics. DG Matrix's Intelligent Power Fabric targets the same outcome as Emerald (cheapest MW to highest-revenue load), at a different layer of the stack: physical routing hardware with embedded orchestration software, housed inside the datacenter's power distribution fabric. Hardware fabrics can compete on temporal flexibility (throttling power to a rack) and resource flexibility (routing between grid, generation, storage). They cannot do spatial flexibility (moving AI workloads across geographically distributed datacenters over fiber). Electrons cannot move across regions, but jobs can. Spatial flexibility is inherently a software layer. If temporal and resource dominate the flex economy, hardware-native players win the coordination rent. If spatial becomes a material share of the flex mix (regional grid stress varies, inference routes to cheapest power), software-native players like Emerald have an exclusive lane. But that lane narrows if AI datacenters increasingly go behind-the-meter with onsite generation (Enchanted Rock at Meta El Paso, Microsoft's San Jose RNG microgrid, eventually SMR plus storage). A site with stable on-site power has less reason to chase cheap grid MW across regions. In that world, resource flexibility (hardware fabrics) dominates the flex economy and spatial flexibility shrinks to a niche.

Four gates for the Emerald bull case

Before the software-moat debate, questions have to clear.

Gate 1: Technical. Can Conductor repeatedly reduce load without damaging compute economics. Answered partly by the SRP pilot and Nvidia integration work, still open at frontier training scale.

Gate 2: Bankability. Will an ISO or utility accredit the response as firm capacity. Answer varies by ISO and by whether the tariff process recognises software-only flex. Emerald is materially exposed to this per-ISO variance.

Gate 3: Interconnection transfer. Does the accredited response allow the site to secure materially more MW, or secure them materially faster, than it would otherwise. This is the queue-avoidance question the model isolates. Without genuine transfer, the flexibility right captures very little value regardless of software quality.

Gate 4: Value capture. Can Emerald retain 15 to 30% of that value; the hyperscaler, utility, NVIDIA, or site scheduler captures the rest in the loss case, utility, NVIDIA, or the site's internal scheduler. This is the coordination-rent question, and it is where the Series A cap-table structure becomes relevant.

Only after all four does the $1.05 billion mark become defensible. The current mark is exposed to any one of them failing.

The workload-mix bet inside Gate 3

The training-versus-inference distinction matters more than the essay has treated it, but not the way the standard framing suggests. Foundation model training is highly batchable: a run can pause for three hours on a hot afternoon and resume from a checkpoint. Real-time user-facing inference is the opposite. A ChatGPT query, a customer service agent response, or a latency-bound API call degrades the moment throttling hits.

If AI datacenter demand tilts toward instant consumer inference over the next 3-5 years, the addressable flexible load shrinks proportionally. Emerald's bull case depends on the workload mix staying weighted toward training plus batch and agentic inference. Overnight code refactoring, protein modelling, deep analytical workflows, embeddings generation. Those bucket into low-priority tiers and throttle without user-visible impact. Real-time consumer inference does not fit that profile.

The workload-mix bet belongs inside Gate 3 (interconnection transfer). The transfer only happens if the load being served can be curtailed. Emerald's addressable market stays below "all AI compute." It is "AI compute the operator is willing to schedule around grid stress." That share is high today because training dominates. If real-time inference becomes the majority, the share compresses regardless of software quality.

What information would raise confidence in the valuation

None of the numbers below are currently disclosed. All of them would be standard for a Series B or later round with the same mark.

• Contracted ARR: annualised recurring revenue under contract. Pipeline counts, customer counts, and deployment announcements do not answer this.
• MW under contractual control: how many MW of AI compute are contractually committed to Conductor coordination, distinct from demonstrated or potential MW.
• Revenue per controlled MW: shows whether the 100 GW TAM headline translates into a real business or stays a marketing figure.
• Customer economics: for a 100 MW AI facility, how much annual value does Emerald create for the customer, and what percentage does Emerald capture?
• Renewal and expansion: does a customer go pilot → 10 MW → 100 MW → 1 GW? Or pilot → thank you → internalise?

And one strategic question that decides whether it becomes a category standard

Does any of Google, Meta, Microsoft, or AWS publicly or contractually adopt Emerald Conductor for a production training cluster? Or does one of them publish its own version (an internal scheduler that reads OpenADR grid signals for curtailment)?

Emerald already has commercial deployments (Digital Realty, planned Vera Rubin with NVIDIA). The specific question is whether any hyperscaler runs Conductor as its production coordination layer, or whether one of them announces its own. Commercial deployment is a different claim from hyperscaler-standard coordination layer.

When we know

Aurora comes online in late 2026. Vera Rubin follows in Q1 2027. Between them, we get the first look at production curtailment data on Nvidia Blackwell training loads. FERC and the ISOs will file their tariff responses through the same window. The hyperscaler decision (adopt Conductor or build in-house) can come at any point. It is the largest single mover in the tree. By mid-2027, we will know which of the four paths Emerald is on.

Testing the thesis with a model

We built a probabilistic engine (adikumar.co/flex-model/) to see whether the framework survives real assumptions. It runs 5,000 Monte Carlo draws per architecture across four scenarios, with correlated inputs, time-varying GPU rates, a full project-finance waterfall (MACRS, storage ITC, debt-equity, DSCR, salvage), and per-ISO bankability. Benchmarked against Meta El Paso, SRP Phoenix, Aurora, and Vera Rubin, capex within ~6% of disclosed values.

Four scenarios cover the plausible 2026-2028 macro. Base case holds queues at 12-30 months with moderate curtailment. Grid-constrained boom stretches queues to 24-54 months with capacity payments surging. Capex winter collapses GPU rates to $1.50 per hour, shrinks queues, and drops curtailment. Fast ISO reform accelerates tariff recognition and lifts bankability.

The Emerald label in the model represents the software-shed architecture class. Emerald AI is the modelled representative because its deal is public and its economics are more disclosed than any competing software vendor. The same class rent is contested by EnergyHub, AutoGrid, Voltus, and hyperscaler-internal control planes. If OpenADR, IEEE, and hyperscaler control-plane APIs converge on a standard, the entire class loses rent regardless of which vendor currently holds the market share. Read "software wins" in the results below as an architecture-class outcome.

Enchanted Rock wins 96% of Base Case draws, 100% of Grid Boom, and 99% of Fast ISO Reform. Emerald wins 4% of Base Case and 1% of ISO Reform. Then Capex Winter arrives, and under that scenario's assumptions Emerald wins 99.7% while Enchanted Rock and Verrus both post negative median NPVs. That figure is a Monte Carlo frequency inside the model. Read it as a model output. Verrus never wins a single scenario outright. Storage ITC of about $36 million on a 100 MW deal cannot rescue battery firming, because it captures neither the queue-avoidance premium nor the capital-efficiency downside protection.

Two inputs drive most of the model output. Enchanted Rock's queue-avoidance share is the first: if BTM gas bypasses the queue completely, gas wins 96%+ everywhere. Drop that share to 80%, the winner flips to Emerald. Drop to 70%, Emerald wins 84%. Emerald's software bankability is the second, and it barely moves the ranking: sweep it from 30% to 95% and outcomes shift by 10 percentage points at most. Interconnection policy holds more of the model outcome than software bankability or tariff-recognition assumptions.

Which reframes the whole flex-architecture debate. It is usually cast as software versus hardware, coordination versus generation. Both frames miss the actual pivot: how much of the interconnection queue can behind-the-meter gas bypass. If regulators treat BTM gas as fully islanded, it bypasses 100% and Enchanted Rock wins in three scenarios out of four. If regulators require grid-parallel treatment because BTM sites still need standby import, backfeed rights, or export capability, gas bypasses 70 to 80% at most and software wins the coin flip. The threshold lands at 80%. Above it, gas wins 96% of Base Case draws. Below it, software wins.

That is a jurisdictional policy question, testable in every ISO's specific rules. The trade an investor is making is a bet on where interconnection policy lands in the specific ISO the deal is written for. And in the specific scenario where the AI capex cycle rolls over, Emerald is the only architecture that survives, because it has no capital at risk. That reframes the $1.05 billion valuation as a put option on the AI capex cycle. Whether the price is fair depends on how likely you think the cycle rolls over.

The call

Bottom-up bear math. If Emerald reaches $50M sustainable revenue with 8× multiple, that clears $400M. If it reaches $100M revenue with 10× multiple, that clears $1B. Getting to $1.05B requires Emerald to reach ~$100M ARR and hold a 10× multiple. Both conditions have to hold.

A harsher read: $20-40M sustainable revenue at 5-8× multiple gets you $100-320M. That's the outcome if Emerald ends up as a component inside NVIDIA's stack or a strategic acquisition target, instead of a standalone SaaS platform.

The Series A cap table includes twelve Fortune 500 strategics (NVIDIA, Siemens, GE Vernova, Eaton, plus eight others) and zero hyperscalers. The industrial and grid ecosystem cares about the category. No hyperscaler has yet backed Emerald as a production coordination layer. That is what the current mark is exposed to.

The valuation is most sensitive to two linked unknowns: flexibility value per MW-year, and Emerald's share of that value. The $20-50/kW-year range and the assumed 15-30% take rate together hold more of the $1.05B defence than any single item elsewhere in this section. Attack those first if the valuation looks aggressive.

19The verdict

The underwriting question is the same for all six approaches. Batteries, schedulers, and permit-ready sites create flexibility, but flexibility on its own does not pay. The money comes from someone paying for the option: interconnection concessions, capacity payments, avoided infrastructure, energy-market revenue, resilience premiums, or accelerated time-to-power. Contract owners earn the money.

Against that frame: Enchanted Rock has real revenue but no headline valuation number like Emerald's, and the gap comes from the market pricing infrastructure revenue differently. It is because the market prices infrastructure operators (like Enchanted Rock) more conservatively than software platforms (like Emerald), no matter how much revenue evidence exists. Aurora and Vera Rubin are necessary validation events. They stop short of sufficient valuation events. If they fail to reproduce SRP-style behaviour, the $1.05B mark is difficult to defend. If they succeed, the next question becomes whether that technical proof converts into contracted MW, accredited capacity, and durable platform take. The failure path compresses the mark toward the $200-400M range. Cloverleaf and Rowan sell the interconnection substrate every other approach layers on top of, and that substrate keeps hardening as the queue stretches.

The bottleneck keeps shifting. Today it is physical MW at the meter, which is why site developers and interconnection queues matter. When queues start clearing, the constraint moves to accredited MW: whoever can sell contracted flexibility a utility will pay for becomes the scarce party. That is the window for coordination software vendors like Emerald. But coordination software follows the standard commoditisation path. Once OpenADR, IEEE, and hyperscaler control-plane APIs converge, the coordination layer stops earning rent. The durable positions after that convergence are the ones holding physical scarcity that cannot be commoditised: permitted sites in load-heavy regions, gas and water rights near datacenter clusters, transmission entry points that cannot be replicated. Each phase of the migration compresses the last.

Scarcity can migrate, or it can disappear

Physical MW is scarce while interconnection is constrained. Bankable MW becomes scarce once the grid begins recognising flexibility but before accreditation becomes standardised. Coordination rights become valuable only if the control plane captures a substantial share of that accredited value.

The final risk is that queue reform compresses all three scarcity rents faster than the market expects. Cloverleaf and Rowan take the direct hit on land-inventory value. Enchanted Rock loses part of the queue-bypass premium if BTM permitting gets harder or if grid connection times drop. Emerald's coordination rent shrinks if the accreditation process becomes standardised across ISOs and hyperscalers write directly to that API.

Under that risk, the flexibility architecture that captures the largest share of interconnection scarcity after accreditation, financing, enforcement, and value-sharing is the durable investment. Which architecture that turns out to be depends on how much of the queue each approach removes, and how much of the removed scarcity each approach's vendor can keep.