Related works: The Authorization Market — Standardized Bargaining, Rationed Power, and the Competition to Build America’s AI Infrastructure, AI Data Center Authorization Bargaining Power, AI Datacenter Edge Computing: Ship the Workload Not the Power
Critical references: The Model AI Infrastructure Authorization Code · The Data Center Authorization Market: A 50-State Regulatory Atlas
Thesis. Match each class of AI computation to the electricity and jurisdiction it can run on, and clean power carries far more of the compute the United States has reason to keep.
Why now. PJM’s 2026 large-load framework, FERC’s June 2026 large-load orders promoting flexible transmission service, and recent nuclear and geothermal offtake put firmness on the table this planning cycle. Every audience faces the placement decision now: policymakers · operators · counsel · investors.
Executive Summary
The clean-power debate for Artificial Intelligence (AI) has focused heavily on supply. Build more generation, more transmission, more storage. A second, underused lever sits on the demand side. Match each class of computation to the electricity it can run on, and the same clean generation carries far more of the compute the United States has reason to keep.
One effect governs the argument, and the paper names it the firmness inversion. Moving flexible computation toward abundant energy lowers total domestic load. The workloads left behind grow more dependent on firm power, because latency and security hold them in place. Firm clean generation becomes more valuable as flexible compute leaves, not less.
The rule the paper defends compresses to one line. Move the compute that can move, and firm the compute that cannot.
The thesis follows. Through 2040, the binding infrastructure question shifts from where to site a data center to which computation belongs where. Operators who separate workloads by firmness and mobility will run more of their anchored compute on clean power than operators who keep optimizing campus scale.
MindCast builds Cognitive Digital Twins (CDTs) of the operators, utilities, and jurisdictions in the contest. The twins simulate their decisions under pressure.
Behavioral economics supplies the decision rules, including the bigger-is-better heuristic that keeps capital flowing to campuses the grid cannot firm. Game theory supplies the payoff structure and the equilibrium the bargain reaches. Predictive Behavioral Economics + Dynamic Game Theory produces the forecast.
The paper evaluates why concentration collides with the grid, then separates compute by its power and geography needs. It locates the domestic floor the United States must firm, then defines the bargain operators and jurisdictions strike through 2040.
The formal simulation releases six MindCast Simulation Predictions, one primary and five secondary. The retained tier buys firmness at 76-88%, the primary claim. Workload differentiation becomes public strategy at 82-90%, and capacity-backed service governs constrained conditions at 84-92%.
🏛️ Policymakers should write firmness and mobility into authorization terms, and value accredited capacity rather than raw megawatts.
💼 Executives should move from campus-scale optimization to compute-portfolio optimization, and place each workload where its power lives.
⚖️ Counsel should treat offshore placement as a move inside the export-control perimeter, mapped to current destination controls rather than around them.
📊 Investors should evaluate the firmness profile of retained load, because the retained tier drives firm-clean demand more than announced gigawatts do.
I. The Gigawatt Concentration Problem
Scale built the current data center. Larger campuses lowered the cost of construction, networking, and operations. The industry made the multi-gigawatt campus its default, and the grid did not cooperate.
A multi-gigawatt load concentrates generation, transmission, and reliability demand at one point. Grid headroom is finite at any point. The larger the single load, the more of a region’s remaining capacity it consumes, and the harder its consent becomes to win.
Concentration also collides with a rule most siting analysis skips. Connection is not firmness. On July 27, 2026, the PJM Board directed development of an interim resource-adequacy framework. Under the framework, qualifying large loads without sufficient capacity backing could face earlier reductions during shortage conditions. PJM is the largest United States regional transmission organization (RTO).
A campus can clear every state gate, energize, and still face earlier curtailment if its service lacks sufficient capacity backing. Giant campuses will keep being built where the power supports them. The open question is whether the giant campus stays the default destination for every incremental workload, and the grid is making that assumption progressively harder to defend.
II. The Method: Predictive Behavioral Economics + Dynamic Game Theory
MindCast is a simulation firm built on Predictive Behavioral Economics + Dynamic Game Theory. MindCast builds Cognitive Digital Twins of the operators, utilities, and jurisdictions in a contest. The twins run forward under pressure.
Behavioral economics supplies the decision rules. Bounded rationality, salience, and loss aversion set them. An operator over-weights campus scale and under-weights firmness, and a community resists a cost that lands far from its benefit.
Dynamic Game Theory supplies the payoff structure and the equilibrium the contest reaches. The game itself changes while the actors play. Rules, rivals, and available strategies shift mid-contest. Neither discipline predicts alone, and the combination produces MindCast Foresight Simulation Predictions, each with a band and a falsifier.
The simulation runs opposing twins against each other. An operator plays a utility, and a developer plays a regulator, so the forecast reflects the countermoves each actor provokes rather than one actor’s plan in isolation.
III. Not All Compute Needs The Same Geography
Compute is not one product, and the grid treats its varieties differently. Workloads differ by latency, interruptibility, and security. Power intensity separates them further.
Production inference answers users in real time. Latency binds it near the people it serves. Training runs over days rather than in the moment, so it follows energy more readily than latency-bound inference, within the limits utilization and synchronization set. The largest frontier runs are the exception, because tight low-latency fabric across tens of thousands of accelerators keeps them anchored in the retained tier. Batch inference and preprocessing fall between the two.
Inference and flexible compute split cleanly. Inference tends to move toward users. Flexible compute can move toward energy. A firm that sorts its workloads this way gains a placement it can defend to a utility, a regulator, and a community.
Workload class should become a primary determinant of where the next megawatt of AI load lands.
IV. Connected Is Not Firm
Firmness is the property the clean-power debate keeps collapsing. A facility can hold interconnection, buy annual clean energy, and still fail to deliver power when the servers run. Three products hide inside the word firm, and the weakest one governs.
The firmness equation states it plainly.
Firm Service = minimum ( Interconnection, Deliverable Energy, Accredited Capacity )
Interconnection is permission to attach. Deliverable energy is power that reaches the site over available transmission. Accredited capacity is supply the grid counts on during a shortage. A site missing any one of the three stays exposed, whatever the other two promise.
Clean supply divides on the same line. Annual matching proves procurement over a year. Hourly matching aligns purchased generation with consumption hour by hour, and still does not prove nodal deliverability or capacity.
Firm or highly dependable clean resources can provide stronger capacity contribution than standalone variable generation, depending on configuration and accreditation. Variable clean resources earn partial capacity credit. Storage and complementary supply raise portfolio capacity contribution, and transmission improves deliverability.
A clean mandate written without the capacity layer can produce sites that are green on paper yet exposed to earlier curtailment under constrained conditions. Policymakers hold that risk directly.
V. The Domestic Compute Floor
Some computation carries enough value in American geography to stay regardless of where power runs cleanest. The paper calls that set the domestic compute floor. The floor holds the workloads the United States has economic or strategic reason to keep onshore.
Production inference leads the set, held by latency to the users it serves. Four further categories stand as candidates the formal run resolves. Defense and critical-infrastructure compute qualify on security, and regulated data qualifies on law. Sensitive model operations and selected frontier training qualify on control and strategic value.
The floor changes what the United States must generate. Move flexible compute toward abundant energy, and total domestic load falls against the counterfactual. The load that remains is the load hardest to move, and hardest to run on intermittent supply.
The retention boundary is contested, and the paper holds it open rather than fixing it in advance. Define the floor first, because the floor sets the firm-power bill, not the announced gigawatt total.
VI. The Firmness Inversion
The retained floor produces the paper’s central and least intuitive finding. Compartmentalization does not relieve the American clean-firm problem. Compartmentalization concentrates it.
The firmness inversion works in two steps. First, flexible workloads leave for energy-advantaged sites at home or abroad, and aggregate domestic load falls. Second, the workloads that remain are latency-bound and security-bound, so their firmness requirement per megawatt rises. Total load falls, and the remaining load runs firmer.
Firmness rises on two axes at once. Retained inference needs firm service to run around the clock. Retained inference also needs a capacity-backed service arrangement where the governing market differentiates shortage treatment by capacity support. A site can contract clean power and still fail the capacity test.
The inversion favors firm clean resources. Nuclear and hydro and geothermal gain strategic value precisely as flexible compute departs. Variable clean generation gains a complementary role, paired with the flexible workloads that bend to it. Moving flexible compute away raises the firm-clean intensity of what stays, so specialization makes firm clean more valuable, not less.
VII. Matching Workloads To Power
Placement becomes an allocation problem once workloads and resources are sorted. Each compute class carries a firmness need and a geography need. Each resource carries a delivery shape. Good placement pairs them.
Persistent, high-utilization workloads suit firm resources, so nuclear and hydro and geothermal anchor them where available. Flexible workloads suit variable resources, so wind and solar with storage carry training that follows the curve. Transmission opens geographic arbitrage and adds its own constrained layer.
Compute can also provide grid value rather than sit as a passive load. A training cluster that ramps with wind, or pauses during a shortage, delivers demand response the grid would otherwise procure. Flexibility earns capacity-equivalent value only where the market accredits it, and Duke research estimates roughly 98 GW of new load could be integrated across US balancing authorities if flexible loads accept an average annual curtailment rate of 0.5%. Interruptibility converts a liability into a service.
The right site pairs a workload’s firmness need with a resource’s delivery shape, and lets flexible compute earn capacity-equivalent grid value.
VIII. Energy-Advantaged And Offshore Placement
Flexible compute follows energy, and the first move is domestic. Pacific Northwest hydro, wind-rich interior grids, and geothermal basins offer clean power away from the metros that host inference demand. A training campus does not have to leave the country to leave a congested grid.
Offshore placement extends the same logic across a border, and complements domestic capacity rather than replacing it. Allied clean-power jurisdictions can host energy-intensive training with limited user-proximity requirements. Canadian hydro and Nordic firm renewables lead the set. Offshore capacity works as a release valve that preserves scarce domestic firm power for the anchored floor.
Offshore access opens only inside a legal boundary, and Counsel should treat the boundary as the design constraint. United States export-control law gates advanced computing hardware and advanced AI model weights by destination, and separates allied jurisdictions from adversary ones. License Exception Artificial Intelligence Authorization and Supplement No. 5 to Part 740 name the eligible allied jurisdictions, and both remain current law amended through 2026 as enforcement posture shifts.
Four factors decide offshore eligibility together, not a frontier line. Workload mobility and jurisdiction eligibility set it, and security classification and infrastructure economics complete it. Geographically mobile workloads are the strongest initial candidates, and the permitted set can widen or narrow on a rule change no operator controls.
IX. The Grid-Anchored Bargain Through 2040
Authorization, energy, and compute architecture converge into one bargaining system. Utilities and jurisdictions hold generation, transmission, and authorization. Operators hold flexible load, curtailment capability, and generation investment. Each side wants what the other controls.
Game theory sets the payoff structure. Jurisdictions weigh consent, firm supply, and ratepayer protection. Operators weigh speed, certainty, and firm service. Standardized terms turn the contest into a competition for scarce authorization, and the operator who brings flexibility and firm-clean investment clears terms the announce-first operator cannot.
Behavioral economics explains the friction. The bigger-is-better heuristic keeps capital flowing toward campuses the grid cannot firm, because scale is salient and firmness is not. Communities weigh concentrated local benefit against diffuse ratepayer cost, and loss aversion sharpens resistance when the bill lands far from the jobs. Executives who read those decision rules bargain better than executives who model only megawatts.
The simulation resolves the bargain toward a portfolio of specialized sites anchored to different power systems, not a single optimized campus. The long-horizon question follows. Does the industry optimize the next generation of AI infrastructure by campus, or by a portfolio of specialized compute anchored to the grids that firm each workload?
X. MindCast Simulation Predictions
The paper turns on one question. Can workload architecture increase the amount of strategically important AI that clean power can firm? The simulation releases six Simulation Predictions, one Primary (P) and five Secondary (S). Each carries a band, a falsifier, and a public source.
Major operator is frozen to seven firms for validation: Microsoft · Google · Amazon · Meta · OpenAI · Anthropic · xAI. Primary marks causal weight, not the highest band.
Competing hypothesis. Supply-side and market reform could preserve the economic advantage of heterogeneous mega-campuses long enough that workload-by-geography separation stays marginal through 2029. Contrary-route test: fewer than two major operators adopt explicit workload-by-geography strategies while new large campuses keep absorbing heterogeneous workload classes without differentiated-service constraints.
P1 · The retained tier buys firmness (76-88%): Two or more major operators pair anchored United States compute with firm-clean or capacity-backed supply, distinct from their flexible workloads. Validated against power purchase agreement (PPA) and interconnection disclosures by December 31, 2029. Falsifier: fewer than two major operators publicly pair anchored compute with firm-clean or capacity-backed supply by the window.
S1 · Workload differentiation becomes public strategy (82-90%): Two or more major operators publicly separate workloads by firmness and geographic mobility as infrastructure strategy. Validated against operator infrastructure disclosures and earnings materials by December 31, 2029. Falsifier: operators keep guiding toward undifferentiated campus scale with no public placement strategy.
S2 · Flexible compute relocates toward energy, domestic first (72-81%): At least three major new flexible AI compute sites locate in energy-advantaged domestic regions rather than the demand metros. Validated against site announcements and interconnection filings by December 31, 2029. Falsifier: fewer than three such sites locate outside the demand metros by the window.
S3 · Firm clean gains strategic value (78-88%): Cumulative disclosed firm-clean nameplate capacity contracted by AI operators, in megawatts, grows faster than cumulative disclosed variable-only nameplate capacity contracted by those operators. Validated against procurement disclosures by December 31, 2029. Falsifier: firm-clean nameplate contracted stays flat or declines against variable nameplate.
S4 · Offshore compute complements domestic capacity inside the permitted perimeter (60-73%): At least one major operator places a substantial energy-intensive, geographically mobile AI workload in an eligible allied jurisdiction under United States export-control compliance. Validated against operator announcements and regulatory filings by December 31, 2030. Falsifier: no such placement, or placement outside the eligible perimeter. Prediction-break: a federal rule change that materially widens or narrows the eligible set.
S5 · Capacity-backed service governs constrained conditions (84-92%): At least one major RTO or large-load regime operationalizes service or curtailment differentiation by capacity backing, firm supply, or enforceable flexibility. Validated against RTO filings and operative tariffs by December 31, 2028. Falsifier: no such differentiation is operational in any major market by the window.
Two interpretive findings sit outside the event pool, and do not pool with the bands above.
Firmness inversion (85-92% directional): Moving mobile compute away lowers aggregate domestic demand while raising the firmness intensity of the load that remains. The mechanism is the paper’s governing contribution.
Portfolio advantage (directional, held moderate-high): Operators able to allocate workloads across differentiated sites gain greater clean-power optionality than operators tied to concentrated load. The advantage is structural, not exclusive, because a concentrated operator can still specialize over time or contract firm supply.
The primary prediction carries the paper’s weight, and the firmness inversion is the mechanism beneath every entry.
XI. Risk Mitigation
Each prediction converts into a stakeholder exposure and a set of unilateral actions. Per prediction the layer names the loss on inaction, the unilateral moves that reduce it, and the residual that survives full mitigation. Actions are analytic options rather than recommendations, and none assumes advance knowledge of the outcome.
P1 · The retained tier buys firmness (76-88%).
💼 Executive exposure: an operator that anchors latency-bound compute without firm clean or accredited capacity can face higher power costs, curtailment exposure, and delayed time-to-compute on the retained fleet. Actions: energy procurement contracts firm clean or accredited capacity for anchored sites by the 2029 checkpoint; operations secures qualifying capacity-backed arrangements ahead of the resource-adequacy window; portfolio planning classifies the fleet by firmness need now. Residual: firm-clean lead times exceed the contracting window, so a share of anchored load runs unfirmed through the buildout.
S1 · Workload differentiation becomes public strategy (82-90%).
💼 Executive exposure: an operator optimizing undifferentiated campus scale while rivals differentiate loses the marginal siting decision and strands interconnection deposits at congested sites. Actions: strategy formalizes a workload-placement architecture separating classes by firmness and mobility by the 2029 checkpoint; siting redirects the next incremental campus to the differentiated rule. Residual: sunk mega-campus commitments cannot be re-differentiated, so legacy concentration keeps its firmness cost.
S2 · Flexible compute relocates toward energy, domestic first (72-81%).
🏛️ Policymaker exposure: an energy-advantaged state that does not ready interconnection and firm supply forgoes the construction spend and tax base a relocation wave brings, measured against the session calendar. Actions: the legislature writes a large-flexible-load authorization track with firm-supply and interconnection terms by the next session; the utility commission pre-clears a flexible-load tariff so terms sit ready. Residual: transmission buildout exceeds a session cycle, so a state can write terms and still lack deliverable headroom.
💼 Executive exposure: an operator anchoring flexible training in a demand metro can pay for locational firmness beyond the workload’s geographic requirement, measured in power cost per megawatt-hour above an energy-advantaged site. Actions: siting routes new flexible workloads to energy-advantaged regions; network engineering provisions backhaul so relocation holds the workload together. Residual: some flexible workloads carry data-gravity ties to a metro that relocation cannot fully sever.
S3 · Firm clean gains strategic value (78-88%).
💼 Executive exposure: an operator without secured firm-clean offtake competes for a shrinking pool as rivals contract it, measured in the premium paid for late procurement. Actions: procurement weighs earlier firm-clean contracting or equity participation where it reduces supply risk; corporate development evaluates development partnerships with advanced nuclear or geothermal projects. Residual: firm-clean deployment timelines cap how fast contracted supply becomes delivered supply.
S4 · Offshore compute complements domestic capacity inside the permitted perimeter (60-73%).
⚖️ Counsel exposure: an operator placing compute offshore without mapping current export-control destination rules risks a violation, measured in licensing delay and enforcement exposure. Actions: trade compliance classifies each offshore-candidate workload against current Export Administration Regulations destination controls and license requirements before deployment; legal builds a jurisdiction-eligibility gate on offshore siting. Residual: export-control policy can shift within the window, so a compliant placement today can fall outside the eligible set faster than an operator can relocate the load. Constraint: counsel cannot pre-clear an export with the Bureau unilaterally, so the action is internal classification and a license application, not an assurance of approval.
💼 Executive exposure: an operator treating offshore as a substitute over-commits abroad and under-firms the retained domestic tier, measured in quarters of retained-tier firmness deferred by the offshore bet. Actions: infrastructure strategy sizes offshore to the mobile tier only; portfolio planning holds domestic firm-clean procurement on its own track. Residual: allied-host transmission and interconnect limits cap how much mobile load offshore can absorb.
S5 · Capacity-backed service governs constrained conditions (84-92%).
🏛️ Policymaker exposure: a market without capacity-backing and flexibility terms before the next shortage faces contested curtailment events it cannot adjudicate cleanly, measured in disputes per constrained season. Actions: an RTO files capacity-accreditation and flexibility-service rules with FERC ahead of the constrained season; a state commission acts through its own authority to set large-load service terms. Residual: FERC review timing sits outside the filer’s control, so a filed rule can lag the shortage it governs. Constraint: an RTO rule requires FERC action to take effect, so the unilateral move is the filing and the interim tariff, not the approval.
💼 Executive exposure: an operator whose large load lacks capacity backing can face earlier curtailment under applicable capacity-linked service rules, measured in hours of forced curtailment per constrained season. Actions: operations secures accredited capacity or enrolls enforceable flexibility for constrained-condition service; engineering builds curtailment-tolerant scheduling so a shortage lowers throughput rather than halting it. Residual: a severe multi-day shortage exceeds what enrolled flexibility can absorb, so residual curtailment survives on the retained tier.
Several actions here, firm-clean procurement and workload differentiation among them, also reduce exposure in the buy-side bargaining companion. The linkage holds, and neither analysis claims the action twice.
XII. What To Watch
Three observables decide the direction before the formal windows close. PJM’s Interim Resource Adequacy Service reaches an operative tariff or stalls at FERC. Hyperscaler firm-clean offtake, in nuclear and geothermal, either accelerates or plateaus through 2027. New flexible-compute sites announce in energy-advantaged regions or concentrate in the demand metros.
The dominant fork is federal. Export-control posture on allied-jurisdiction compute can widen or narrow the offshore option, and the administration’s move away from a single replacement rule leaves the eligible set open to amendment. Watch the rulemaking that follows the 2026 amendments, because it sets whether the mobile tier can leave.
XIII. Conclusion
Scale built the data center, and firmness is unbuilding the mega-campus as the universal default. Grid rules increasingly differentiate how large loads can be served, and the workloads that cannot move inherit the firmest, scarcest power. Operators who match computation to electricity and jurisdiction will run more of their anchored compute on clean supply than operators who keep optimizing campus scale.
The paper releases six MindCast Simulation Predictions on that mechanism, led by the retained tier buying firmness at 76-88%. Move the compute that can move, and firm the compute that cannot.
Appendix: MindCast Corpus
The paper builds on prior MindCast work. Each entry carries its role in one line.
The AI Data Center Authorization Price: A 50-State Baseline (2026). The sell-side baseline this paper’s demand architecture bargains against.
AI Data Center Authorization Bargaining Power (2026). The buy-side operator ratings whose clean-portfolio axis this paper deepens.
The Authorization Market (2026). The series umbrella and the four-layer authorization stack this paper sits inside.
The Model Data Center Authorization Code (2026). The measurement instrument behind the firmness and clean-energy provisions.
The Power Stack (2025). The resource hierarchy this paper extends into workload placement.
AI Datacenter Edge Computing: Ship the Workload Not the Power (2025). The origin of the ship-the-workload principle carried here into the firmness frame.
VRFB’s Role in AI Energy Infrastructure (2025). The long-duration storage layer that lets variable clean generation firm flexible compute.
The Two-Ledger Data Center Bargain (2026). The local net-benefit framework behind community consent.
Three Competing Governance Equilibria (2026). Conditional acceleration as the coordination equilibrium this paper assumes.
The Federal-State AI Infrastructure Collision (2026). The FERC and RTO layer the connected-is-not-firm argument draws on.
AI Data Center Credit Risk (2026). The credit lens behind the investor exposures in Risk Mitigation.
The MindCast AI Data Center Record (2026). The corpus index for the full data-center workstream.
Sources
External references.
PJM Board of Managers, direction on an interim resource-adequacy framework (July 27, 2026).
United States export controls on advanced computing and AI model weights, 15 CFR Part 740.
License Exception Artificial Intelligence Authorization and Supplement No. 5 to Part 740, current as of 2026.
Nicholas Institute, Duke University, Rethinking Load Growth (2025).
Lawrence Berkeley National Laboratory, 2024 Report on U.S. Data Center Energy Use, for the Department of Energy.
XIV. Working With MindCast
MindCast turns each forecast into a commissioned engagement. The offers key to the predictions.
MindCast turns the analysis into three commissioned engagements. Each one converts the paper’s mechanism into a decision a specific team can act on.
💼 Executives can commission a Fleet Firmness Map. The engagement sorts every site by whether its workload can move or must stay, then matches each anchored site to a firm-clean or capacity-backed supply plan and each flexible site to an energy-advantaged region. Operators leave knowing which megawatts to firm and which to relocate.
🏛️ Policymakers and RTOs can commission a Large-Load Service Design. The engagement drafts the tariff and interconnection terms that differentiate shortage treatment by capacity backing and enforceable flexibility. A state gains a way to attract data-center investment without shifting reliability cost onto existing ratepayers.
⚖️ Counsel can commission an Offshore-Placement Decision Map. The engagement checks each candidate workload against the current export-control structure and names which jurisdictions and workload classes qualify. A firm gains a defensible path to place mobile compute abroad without stumbling into a violation.
Reach MindCast at mcai@mindcast-ai.com. Visit our Corporate Site.
