Parent installment: The Grid-Anchored Clean Power Bargain (2026). The firmness inversion that paper proved becomes the mechanism beneath the match modeled here; 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. AI compute and clean generation clear as one two-sided market, sorting into matched pairs by power firmness and compute mobility, and the load that cannot clear the pairing stays on fossil generation.
Why now. Firm-clean offtake is already attaching to anchored compute this planning cycle, through nuclear restarts and geothermal contracts, while variable procurement chases flexible load. The pairing is forming in public ahead of the market framing that explains it.
Executive Summary
The electricity market for Artificial Intelligence (AI) does not clear by generation technology alone. The market clears jointly across two scarce attributes, power firmness and compute mobility. Firmness measures whether power arrives when the servers run, and mobility measures whether a workload can move or reschedule around energy. The pairing either forms or fails.
The thesis follows. Compute and clean generation sort into stable matched pairs. Variable clean carries the compute that can move, firm clean carries the compute that cannot, and the load that clears neither pairing stays on fossil supply.
MindCast reaches the result by simulation rather than trend extrapolation. The firm builds Cognitive Digital Twins (CDTs) of operators, developers, utilities and jurisdictions. Opposing CDTs then run against each other under pressure, and Predictive Behavioral Economics + Dynamic Game Theory produces the forecast.
The paper names the clearing mechanism and sorts the compute classes against the power resources. The formal simulation tests where the largest training runs cross from movable to anchored and estimates the fossil residual the resulting match leaves behind. The simulation releases a set of MindCast Foresight Simulation Predictions, one primary and five secondary. Each prediction carries a band, a falsifier and a public source.
The formal run places the coupled clean-compute match at 78-88%, with firm clean concentrating on anchored compute at 85-93%.
🏛️ Policymakers should treat access to firm clean as an authorization problem, because the residual forms where mid-scale operators cannot clear the interconnection and capacity gates.
💼 Executives should sort the fleet by mobility before contracting power, because a workload’s firmness need decides which clean resource it can pair with.
⚖️ Counsel should read export-control posture as a matching variable, because a rule change relocates frontier compute and reshuffles which clean supply it can reach.
📊 Investors should evaluate firm-clean scarcity as the binding constraint, because two high-value compute classes now compete for the same nuclear and geothermal supply.
I. AI Compute And Clean Power Clear As One Market
The clean-power debate treats AI compute growth and clean generation growth as two parallel stories. Both stories share one market. Compute buyers now contract generation directly, so a compute class and a power resource form a pair or fail to.
Firm-clean procurement already attaches to anchored compute. Google contracted Fervo geothermal and Microsoft anchored the Three Mile Island restart. Amazon and Meta signed nuclear and small-modular deals, and roughly 9.8 GW of nuclear now sits committed across the four largest operators.
Procurement is also concentrating. United States corporate clean buying reached a record 29.5 GW in 2025 while the number of buyers fell from about 67 to 33. Fewer, larger buyers now shape where scarce supply lands.
The market clears when a workload’s power need meets a resource’s delivery shape. Matching, not forecasting, governs the outcome. The binding question is not how much clean power the United States builds, but which compute that power can pair with.
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, developers, utilities and jurisdictions in the contest. The CDTs then run forward under pressure.
Behavioral economics supplies the decision rules the simulation tests. Bounded rationality, salience and switching burden set them. The rules include campus-scale over-weighting, a developer’s tilt to the creditworthy counterparty, and a mid-scale buyer’s default to the fastest available power.
Dynamic Game Theory supplies the payoff structure and the equilibrium the contest reaches. The simulation runs opposing CDTs against each other, so an operator plays a developer and a developer plays a utility. The combination produces MindCast Foresight Simulation Predictions, each with a band and a falsifier.
III. Two Scarce Attributes Clear The Market
The match sorts on two scarce attributes at once, power firmness and compute mobility. Firmness is the reliability of supply when the load runs. Mobility is the freedom of a workload to shift in time or place.
The two attributes complement each other, which drives the sort. Firm power carries the most value where compute cannot move, and variable power carries the most value where compute can. Assortative matching follows, so the scarcest firm supply pairs with the least movable load.
The sorting rule compresses to one line. Pair the firmest power with the compute that cannot move, and the most variable power with the compute that can. A pairing holds when no alternative gives both sides more value.
IV. Movable Compute Pairs With Variable Clean
Movable compute ramps and relocates. Batch inference, fine-tuning and mid-scale training run over hours rather than in the moment. Each follows energy more readily than latency holds it.
Variable clean carries the movable set. Wind and solar with storage supply the workloads that bend to the curve, and flexibility converts a grid liability into a service. Duke University’s Nicholas Institute estimates that roughly 98 GW of new load could integrate across United States balancing authorities if flexible loads accept an average annual curtailment rate of 0.5 percent.
The pairing gives a movable operator a placement it can defend to a utility and a regulator. Move the compute that can move, and pair it with the power that cannot firm itself.
V. Anchored Compute Captures Firm Clean
Anchored compute stays put. Real-time inference sits near the users it serves, held by latency, and runs around the clock. Persistent, high-utilization load needs power that delivers during a shortage.
Firm clean carries the anchored set. Nuclear, geothermal and hydro provide continuous output. Scarce continuous output flows to the load that most needs it, and the offtake already visible attaches firm supply to persistent-load campuses.
Firm clean gains strategic value precisely because the anchored load cannot move to cheaper power. Anchor the compute that cannot move, and pair it with the power that can firm it.
VI. The Frontier Boundary: When Training Leaves The Movable Set
Training moves until scale anchors it. A model treating all training as movable misassigns the largest runs, because a single frontier run could reach 4 to 16 GW of peak power by 2030. Tight low-latency fabric across tens of thousands of graphics processing units (GPUs) holds the largest runs in one place.
The largest frontier runs may cross from the movable pair to the anchored pair. Movable training and batch stay with variable clean, while gigawatt-scale frontier training may migrate toward firm clean.
The crossing puts two high-value claimants on the same scarce supply. Real-time inference and frontier training now compete for nuclear and geothermal output. The boundary between movable and anchored training is the market’s central open question, and the formal run places the crossing in the gigawatt range, roughly 0.5 to 2 GW of contiguous synchronized power.
VII. Why The Cross-Pairs Fail To Scale
The model predicts that cross-pairs remain non-material, which is what makes the market an equilibrium rather than a correlation. Firm clean is too scarce and too valuable to spend on freely movable batch. Variable clean cannot hold real-time inference at scale, because intermittency breaks a latency-bound load.
The failures couple the two successful matches. Firm supply captured by anchored compute leaves the pool, which pushes movable compute toward variable supply. One match strengthens the other by removing supply from the alternative.
A matching model earns its distinction by explaining who does not transact. Watch the combinations that stay non-material, because their scarcity supports the sort.
VIII. The Fossil Residual Is Built In
The model treats the residual as endogenous to the mechanism, not a temporary failure to build clean generation. Fragmented mid-scale inference and speed-first flagship siting can clear on fossil supply. Neither clean type fits their scale or their timeline.
Authorization cost helps drive the residual. A mid-scale operator would take firm clean. Where the operator cannot clear the interconnection queue, the capacity-accreditation timeline or the minimum-scale power purchase agreement (PPA)that firm developers require, gas can win on speed and the load stays on grid mix. The simulation tests whether that exclusion persists enough to hold the residual.
The two-claimant squeeze sharpens the outcome. Mid-scale inference may be outbid twice, by real-time inference and by frontier training, so the firm-clean pool can clear before the smaller buyer reaches it. The simulation treats the residual as a market-design result that can persist even as aggregate clean procurement rises.
IX. The Firmness Premium
A matching equilibrium can leave an economic signature, a premium on the scarce attribute. Assortative matching on firmness raises the value of firm supply relative to variable supply as the pairing tightens.
Firm-clean terms may widen against variable-clean terms through the window. Energy analysts holding the correlation see the pairing but can miss the differential, because a correlation carries no rent while an equilibrium does. The widening gap corroborates the match rather than proving it.
Read the matching evidence first, then the differential. A persistent firmness differential paired with workload-specific sorting gives a signature consistent with the equilibrium, because a premium alone could arise from independent scarcity.
X. The Federal Fork
Government can change the matching rules, and export control is the dominant lever. United States export-control law gates advanced computing hardware and AI model weights by destination. 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.
A posture change can relocate frontier compute. A tighter rule concentrates the largest runs in specific domestic jurisdictions, and a looser rule opens allied hosts abroad. Either move reshuffles which clean supply the anchored load can reach.
A federal rule change can redraw the geographic market. Watch the rulemaking that follows the 2026 amendments, because it reshapes whether the mobile tier can leave and which clean supply the anchored tier reaches.
XI. MindCast Foresight Simulation Predictions
The paper turns on one question. Does the AI-power market sort compute and clean generation into a stable matched equilibrium, or does the apparent pairing dissolve once actors respond to scarcity and each other? The formal simulation releases six Simulation Predictions, one Primary (P) and five Secondary (S).
Each prediction carries a band, a falsifier and a public source. Major operator is held to seven firms for validation: Microsoft · Google · Amazon · Meta · OpenAI · Anthropic · xAI. Primary marks causal weight, not the highest band.
Competing hypothesis. Long-duration storage could firm variable generation enough to collapse the firm-and-variable distinction, so the market pairs on cost alone. Contrary-route test: variable-plus-storage contracts anchor real-time inference at material scale while firm-clean offtake flattens.
P1 · A coupled clean-compute match forms (78-88%). Major operators show workload-differentiated procurement, with anchored compute attaching to firm or capacity-backed supply and movable compute attaching to variable clean through flexible siting. Validated against PPA, interconnection and siting disclosures by December 31, 2029. Falsifier: operators procure power without differentiating by workload firmness and mobility.
S1 · Firm clean attaches to anchored compute (85-93%). Two or more major operators pair anchored compute, meaning real-time inference or gigawatt-scale frontier training, with firm-clean supply distinct from their movable workloads. Validated against PPA and interconnection disclosures by December 31, 2029. Falsifier: firm-clean offtake flows materially toward freely movable batch.
S2 · Movable compute attaches to variable clean (66-80%). Two or more major operators pair batch, fine-tuning, or mid-scale training with variable-clean supply through flexible siting or scheduling. Validated against site announcements and procurement disclosures by December 31, 2029. Falsifier: movable compute anchors in demand metros regardless of energy, and mobility stays theoretical.
S3 · Frontier training crosses toward the anchored side (60-72%). At least one frontier run above the run’s contiguous-synchronized-power boundary of roughly 0.5 to 2 GW pairs with firm-clean or capacity-backed supply rather than variable-only supply. Validated against site and procurement disclosures by December 31, 2030. Falsifier: frontier training above that boundary stays siting-flexible and pairs with variable-only supply.
S4 · Cross-pairs remain non-material (70-82%). Firm-clean-to-movable-batch and variable-clean-to-anchored-inference pairings stay below the run’s material threshold, 10 percent of AI-tied contracted clean capacity in the mismatched class. Validated against procurement and interconnection disclosures by December 31, 2029. Falsifier: either cross-pair reaches that threshold.
S5 · A fossil residual persists among mid-scale anchored inference (62-75%). Mid-scale inference load holds gas or grid-mix supply while firm-clean procurement concentrates among the largest operators. Validated against generation-share data and procurement disclosures by December 31, 2029. Falsifier: aggregation or utility structures give mid-scale operators firm-clean access and the tail decarbonizes.
Two findings sit outside the event pool and do not pool with the bands above.
Supporting economic signature (directional). The firmness differential strengthens if scarcity and matching tighten. Project tenor, tax treatment, geography and rates also move the differential. A flat differential does not by itself falsify the match.
The residual is endogenous (directional). The fossil tail follows from authorization cost and firm-clean scarcity, not from insufficient clean generation. The finding reframes the residual as a market-design feature.
XII. Risk Mitigation
Each prediction converts into a stakeholder exposure and a set of unilateral actions. Actions are analytic options rather than recommendations, and none depends on knowing the outcome first. Exposure severity and probability run as separate axes, so a lower-band entry can still warrant early spend where the loss is large.
P1 · A coupled clean-compute match forms (78-88%)
💼 Executives. An operator that contracts power by campus scale rather than by workload pays a locational firmness premium on load that could have moved, measured in dollars per megawatt-hour above an energy-advantaged pairing. Actions: portfolio planning classifies the fleet by firmness and mobility before the next procurement cycle; energy procurement pairs anchored sites with firm-clean supply and movable sites with variable-clean regions by the 2029 checkpoint. Residual: data-gravity ties hold some movable load in a metro that pairing cannot fully relocate.
S1 · Firm clean concentrates on anchored compute (85-93%)
📊 Investors. An investor weighing clean exposure on nameplate alone misreads the firm-clean squeeze, measured in multiple compression as two compute classes bid for the same scarce supply. Actions: investment analysis weights firm-clean access and match quality alongside announced capacity by the next diligence cycle; corporate development evaluates equity participation in advanced nuclear or geothermal where it lowers supply risk. Residual: a late entrant can still contract firm supply over time, so the exposure narrows without closing.
S2 · Movable compute uses variable clean (66-80%)
🏛️ Policymakers. An energy-advantaged state that leaves interconnection and firm supply unready forgoes the construction spend and tax base a relocation wave brings, measured against the session calendar. Actions: the legislature opens a large-flexible-load authorization track with interconnection terms by the next session; the utility commission pre-clears a flexible-load tariff so terms sit ready. Residual: transmission buildout can exceed a session cycle, so a state can set terms and still lack deliverable headroom.
S3 · Frontier training crosses toward the anchored side (60-72%)
⚖️ Counsel. An operator siting a frontier campus without mapping destination and access controls risks a plan that later conflicts with export-control restrictions, measured in licensing delay and enforcement exposure. Actions: trade compliance classifies each frontier-candidate workload against current Export Administration Regulations controls before siting; legal builds a jurisdiction-eligibility gate into the siting process by the 2030 checkpoint. Residual: posture can shift inside the window, and counsel cannot pre-clear an export, so the move stays internal classification and a license application.
S4 · Cross-pairs remain non-material (70-82%)
📊 Investors. An investor underwriting a cross-pair play backs a combination the model puts below the material threshold, measured in stranded capital if the pairing fails to scale. Actions: investment analysis tests every clean-compute thesis against the firmness-mobility match before commitment; diligence weights the exit risk of a mismatched pairing ahead of the 2029 window. Residual: a policy or storage shift could lift one cross-pair, so the exposure narrows without closing.
S5 · A fossil residual persists among mid-scale anchored inference (62-75%)
🏛️ Policymakers. A market that leaves firm clean reachable only at hyperscaler scale strands mid-scale anchored load on fossil supply, measured in the clean-conversion shortfall across the mid-tier. Actions: the legislature opens an aggregation or standardized-offtake track that lowers the minimum scale for firm-clean access by the next session; the utility commission revises accreditation so smaller anchored loads can reach firm supply. Residual: interconnection and accreditation timelines can outlast a session cycle, so terms can pass and headroom still lag.
Several actions here, firm-clean procurement and workload differentiation among them, also lower exposure in the buy-side bargaining companion. The linkage holds, and neither analysis claims the action twice.
XIII. What To Watch
Three observables decide the direction before the formal windows close. Hyperscaler firm-clean offtake in nuclear and geothermal either accelerates or plateaus through 2027. New movable-compute sites announce in energy-advantaged regions or concentrate in demand metros. The firm-over-variable procurement gap widens or holds flat.
The dominant fork is federal. Export-control posture on allied-jurisdiction compute can widen or narrow where frontier training sits, and the rulemaking that follows the 2026 amendments materially reshapes whether the mobile tier can leave.
XIV. Conclusion
AI compute and clean power clear as one market, and the market sorts on firmness and mobility. Variable clean carries the compute that moves, firm clean carries the compute that stays, and the load that clears neither pairing holds on fossil supply.
The fossil residual can be built into the mechanism, not a gap that more clean generation closes. Operators who sort their workloads by mobility before contracting power will pair more of their anchored compute with clean supply than operators who contract by campus scale. Pair the power that firms with the compute that stays, and pair the power that varies with the compute that moves.
Appendix: MindCast Corpus
The paper builds on prior MindCast work. Each entry carries its role in one line.
The Grid-Anchored Clean Power Bargain (2026). The firmness-inversion installment this paper extends into a two-sided match.
The Two-Ledger Data Center Bargain (2026). The local net-benefit framework behind community consent for sited load.
The Data Center Authorization Price: A 50-State Baseline (2026). The sell-side baseline the demand architecture bargains against.
The Data Center Authorization Market: A 50-State Regulatory Atlas (2026). The jurisdiction-by-jurisdiction map behind the authorization-cost residual.
Sources
Wide-Area Power System Oscillations from Large-Scale AI Workloads. arXiv, 2025. 4 to 16 GW single frontier training run, peak power density.
AI Data Centers Energy Consumption in 2024 to 2026. TTMS, 2026. 80 to 90 percent inference share of total AI energy, fleet energy.
Nicholas Institute at Duke University. Rethinking Load Growth, 2025. Roughly 98 GW integrable at 0.5 percent flexible-load curtailment.
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.
Corporate clean-energy procurement volume and buyer count, 2025. 29.5 GW contracted across roughly 33 buyers.
Working With MindCast
MindCast turns the analysis into commissioned engagements. Each one converts the paper’s mechanism into a decision a specific team can act on.
💼 Executives can commission a Compute-Clean Match Map. The engagement sorts every site by mobility, then pairs each anchored site with a firm-clean plan and each movable site with a variable-clean region.
🏛️ Policymakers and regional transmission organizations (RTOs) can commission a Firm-Clean Access Design. The engagement drafts the aggregation and accreditation terms that lower the minimum scale for firm-clean access and shrink the fossil residual.
⚖️ Counsel can commission an Export-Control Placement Map. The engagement checks each frontier-candidate workload against current destination controls and names the eligible jurisdictions.
Reach MindCast at mcai@mindcast-ai.com. Visit www.mindcast-ai-simulation.com.
