A Causal Inference Guide to the AI and Energy Feedback Loop

What the evidence says, why the studies disagree, and simulations that show where each method breaks

Everything here is reproducible. Paid subscribers to Vector1 Research get the complete code, the simulated AI and energy models, and all figures at https://brianjamescurry.substack.com/.

In July 2026, PJM, the grid operator serving 13 states and Washington, D.C., ran its capacity auction for the 2028/2029 delivery year. The price came in at the FERC-approved cap of $325/MW-day across the entire footprint, and total capacity cleared was 6,831 MW short of the reliability requirement. Two years earlier the same market was paying a fraction of that: $28.92/MW-day in most zones for 2024/2025, then $269.92 for 2025/2026, then $329.17 for 2026/2027. One analysis estimated that without the cap, the 2028/29 auction would have cleared at $555/MW-day.

PJM’s independent market monitor puts the blame squarely on data centers. In the last four base capacity auctions, it attributes $29.4 billion of charges to data centers, which is 46% of the $63.6 billion total. For the first half of 2026, it estimates the total wholesale cost per MWh rose 9.7 percent, from $103.40 without data center load impacts to $114.50 with them.

Notice what kind of claim that is. “Without data centers, the price would have been $103.40” is a counterfactual. It is a statement about a world nobody observed. One trade publication made the point directly: the attribution comes without a published method, and splitting a capacity charge across customer classes depends on which load forecast is used and how much of the increase is assigned to demand growth versus supply that failed to show up.

That is the problem this article is about. The most consequential questions in AI and energy right now are causal. What does a new data center do to local prices? Does cheap power attract AI load, or does AI load make power expensive? Does grid investment relieve prices, or does it pull in more demand? Forecasting models cannot answer these questions. Causal methods can, but only when their assumptions hold, and the published studies do not agree with each other.

Below, I cover the evidence for each link in the AI and energy loop, the methods that can estimate those links, why the recent studies reach opposite conclusions, and a set of simulations with known ground truth that show exactly where each method works and where it fails.

The loop, link by link

Start with scale. Lawrence Berkeley National Laboratory estimates that U.S. data centers consumed about 4.4% of total electricity in 2023 and could reach 6.7% to 12% by 2028, rising from 176 TWh in 2023 to between 325 and 580 TWh. The IEA’s global view is similar: data centers used around 415 TWh in 2024, about 1.5% of world electricity, and are projected to reach around 945 TWh by 2030, with data centers accounting for nearly half of U.S. electricity demand growth over that period. Averages hide the real pressure, though. Nearly half of U.S. data center capacity sits in five regional clusters. In Ireland, data centers used 7,663 GWh in 2025, 23% of all metered electricity.

The more important point is that electricity isn’t just an input to AI growth. It is also shaped by it. Every arrow in the loop now has real-world evidence behind it.

AI load affects prices. The PJM auctions above are the most visible example. The size of the effect is contested, as the next section shows.

Prices affect where AI load goes. A June 2026 working paper by Calvin Zhang at UCLA models where AI data centers choose to locate. It finds that a 1 cent/kWh increase in commercial electricity prices is associated with approximately 44% lower odds of receiving a new operational AI data center, and that counties with a pre-2023 data center have approximately 147 times higher odds of receiving a new one. That selection effect will matter a great deal later.

AI demand affects grid investment. Heiss, Sautner and Schmid use the public release of ChatGPT as a plausibly exogenous shock and find that power plant investment rose substantially more in regions with data centers, with most new capacity coming from renewables and battery storage, driven largely by power purchase agreements.

Policy reshapes the load itself. Texas’s Senate Bill 6 applies to loads of 75 MW or more and requires sites coming online after December 31, 2025 to include systems that let ERCOT remotely disconnect them during load-shedding events. In Ireland, after an effective pause on new connections, regulators said new data centers could be built if they met at least 80% of annual demand with new renewable electricity.

Flexibility changes how much grid is needed. A Duke study found that the 22 largest U.S. balancing authorities could accommodate 76 GW to 126 GW of new load if that load could be curtailed for 0.25% to 1% of its maximum uptime. Critics note the authors call it a first-order estimate, did not model transmission constraints, and say it is unclear whether data centers can actually deliver that flexibility. Still, flexibility is a policy lever that changes the size of every other effect in the loop.

Physics couples everything. On July 10, 2024, a 230 kV line fault in the Eastern Interconnection coincided with about 1,500 MW of load reduction that operators had not anticipated. None of it was disconnected by utility equipment. Customer-side protection and controls tripped it. A data center’s behavior at one bus can move voltages and frequency far beyond its own site.

Why forecasting is not enough

Most of the machine learning in the power sector is predictive: load forecasts, price forecasts, renewable output forecasts. Predictive accuracy and causal validity are different properties, and energy research has learned this the hard way.

The clearest example predates AI. Fowlie, Greenstone and Wolfram ran a randomized evaluation of the federal Weatherization Assistance Program across about 30,000 Michigan households. They found upfront costs were about twice the actual energy savings, and the model-projected savings were more than three times the actual savings. The engineering models were confident and wrong, and only an experimental design revealed it.

A second example shows how much analyst choices matter even with good data. Burlig, Knittel, Rapson, Reguant and Wolfram studied efficiency upgrades in California schools using smart-meter data. A standard panel fixed-effects approach found the upgrades delivered anywhere from 12% to 86% of expected savings, depending on specification and outlier handling. Using machine learning to guide specification narrowed the range to 52% to 98%, with a central estimate of 60%. Same data, same question, and the answer swung by a factor of seven depending on choices most readers never see.

A new peer-reviewed review in Energy and AI (Jiang, De Vos, Wang, Deconinck and Kazmi, 2026) takes stock of how causal methods are used across power and energy research more broadly. It groups the literature into five application areas: forecasting, counterfactual generation, energy market analysis, fault detection, and energy and human interaction, and it separates causal discovery (learning which variables drive which) from causal effect estimation (measuring how much).

The review’s findings on evaluation are sobering. Most papers in forecasting and counterfactual generation validate their causal claims with standard accuracy metrics like MAE and MSE, and very few use placebo tests, sensitivity analyses for unobserved confounders, or similar causal checks. Its recommendations are modest and practical: report treatment effect estimates alongside accuracy metrics, run at least one placebo or refutation test, and perform a basic sensitivity analysis. Keep that bar in mind as we look at the data-center evidence.

The evidence disagrees, and that’s the point

Here is what the recent literature says about data centers and electricity prices.

The number that dominated headlines came from Bloomberg’s 2025 analysis, which found that wholesale electricity near significant data center activity cost as much as 267% more for a single month than five years earlier. That is a descriptive maximum at individual price nodes, not an average effect. A PolitiFact review pointed out that wholesale nodal prices only affect the supply part of a bill, which makes up roughly 30% to 50% of what consumers pay.

Pointing the other way, LBNL and Brattle found that over the last five years, states with the highest load growth generally saw average prices decline in inflation-adjusted terms, while states where load contracted generally saw increases. Their explanation is fixed-cost spreading, with an important caveat: a higher-growth future can raise retail prices if new supply and delivery infrastructure is constrained and costly, as it currently appears to be in some or many states.

Correlations can’t settle this. The causal studies that followed, most of them from 2026, try to:

• EPRI (June 2026) explicitly addresses selection. Its authors note that data centers may have chosen to locate in low-price areas, which could mask a price-increasing effect. After controlling for that, they find a doubling of data center capacity caused residential retail prices to fall by roughly 4%, and the average effect of growth from 2019 to 2024 was a 6% decrease.
• Zhang (UCLA, June 2026) looks at wholesale prices instead. Using a staggered difference-in-differences design, he finds that a county’s first AI data center raises wholesale prices by about $5.3/MWh two years after entry in his preferred Virginia-area PJM sample, roughly 11 percent of the sample mean. In the western market he studies, treated counties show no significant price change, possibly because data centers there are built alongside dedicated generation.
• Other 2026 working papers, as summarized by Zhang, report a 5% increase in total retail prices in low-spillover counties (Yue and Zeng) and a 4% increase using a shift-share instrument (Alvarez et al.), while another study finds a small, statistically insignificant effect on residential prices.
• A Federal Reserve working paper uses a dispatch model rather than observational data. It finds existing data centers have already raised wholesale prices by 3% to 5% nationwide on average, with much larger effects in regions with major data center corridors.

These studies aren’t necessarily in conflict. They measure different things:

• Outcome: Wholesale nodal prices respond immediately to scarcity. Retail rates move through regulation and spread fixed costs across all customers.
• Period: 2015 to 2024 was a period of grid headroom in many places. Post-2023 PJM is a period of scarcity.
• Geography: PJM, with its capacity market, is not WEIM, where projects often bring their own generation.
• Estimand: The effect of a county’s first AI data center is not the same thing as the effect of doubling capacity statewide.

Two forces make this especially tricky: selection (data centers choose where to go) and heterogeneity (the effect depends on local grid conditions). To see how they interact, I simulated 400 regions. Each region gets a baseline retail price and a level of grid headroom. Where headroom is ample, new load spreads fixed costs and lowers prices. Where the grid is tight, it raises them. Developers prefer cheap regions with spare capacity. Across all regions, the true average effect is essentially zero.

Overall Effects:

• Naive post-period gap (treated vs untreated): -2.04 cents/kWh
• Difference-in-differences ATT: -0.24 (true ATT -0.23)
• True ATE across all regions: -0.02

Difference-in-Differences by Headroom Group:

• Tight grid: +0.63 estimate (+0.70 truth, 46 treated)
• Medium grid: -0.01 estimate (-0.04 truth, 70 treated)
• Ample headroom: -0.82 estimate (-0.83 truth, 92 treated)

The naive comparison says data centers cut prices by 2 cents, but that is almost entirely selection: they went where prices were already low. Difference-in-differences removes that bias and correctly recovers the effect on the regions that were actually treated. Even so, that treated-region average is not the effect a planner would get by pushing new load into a random region, because the treated regions were chosen for their headroom. Split by grid condition, the effect changes sign.

This is a plausible way to reconcile the real studies. An estimate from years of spare capacity and an estimate from constrained PJM counties after 2023 can both be correct and still point in opposite directions. The question worth asking is not “do data centers raise prices?” but “for which regions, under what grid conditions, and in which price?”

A toy world with known answers

Real data never reveals the true causal graph, so to test methods I built a structural causal model where I know it. Seven variables are observed at each time step:

• W: Heat index (exogenous)
• F: Fuel cost (exogenous)
• A: AI adoption
• D: Data-center load
• P: Electricity price
• I: Grid investment
• R: Regional development

Each variable depends on its own past and on the one-step-lagged values of its parents:

Xt = B * Xt-1 + et

The five endogenous equations are below. Every right-hand variable is lagged one step, and each error term (e) is independent noise.

At = 0.60 At-1 + 0.25 Rt-1–0.25 Pt-1 + et Dt = 0.40 Dt-1 + 0.40 At-1 + 0.35 Wt-1–0.25 Pt-1 + et Pt = 0.30 Pt-1 + 0.35 Dt-1 + 0.40 Ft-1 + 0.30 Wt-1–0.35 It-1 + et It = 0.50 It-1 + 0.30 Pt-1 + 0.20 Dt-1 + et Rt = 0.50 Rt-1 + 0.30 It-1 + 0.20 At-1 + et

These mirror the real links above. Adoption raises data-center load. Load raises price. High prices and load trigger grid investment. Investment lowers price and supports regional development, which feeds adoption. Heat drives both cooling load and price, and fuel cost drives price. The spectral radius of B is 0.84, so shocks eventually die out.

The graph looks cyclic, but it is acyclic once you unroll it over time because every effect takes one step. That time ordering makes discovery tractable, and it is also an assumption. Real systems have effects that happen within a single sampling interval, and those are much harder to orient.

Job one: discovering the structure

Causal discovery tries to learn the graph from data. The standard time-series tool is PCMCI (Runge et al., Science Advances, 2019), which the Jiang review found in only one of its 65 papers. PCMCI runs in two stages. First, a condition selection stage prunes each variable’s lagged candidate parents. Then a momentary conditional independence (MCI) test checks each candidate link while conditioning on the parents of both the target and the source:

X(target) independent of X(source) | Parents(target), Parents(source)

Conditioning on the source’s own parents is what keeps autocorrelation from producing false positives, and energy data is full of autocorrelation. I built a minimal version using partial-correlation tests. It adds one refinement: the first pruning pass can drop a true parent when its direct effect and an indirect path cancel out, so the procedure reruns MCI using the links it found. That cancellation problem is a violation of faithfulness, and feedback loops create it naturally.

With 3,000 time steps and all variables observed, the search recovers the true graph exactly. The structural Hamming distance (SHD), which counts the edges that differ from the truth, is zero.

Then I hid the heat index. That is realistic: weather is often missing, aggregated to the wrong region, or measured at the wrong resolution. The search then reported two lag-2 edges into data-center load, one from AI adoption and one from grid investment. Neither exists. The algorithm worked correctly: those dependencies really are in the observed data. They just aren’t causal, because the variable that explains them is gone.

The best real-world application I found takes this problem seriously. Madadkhani, Sturma, Drton and Ikonnikova studied ERCOT day-ahead prices from 2019 to 2024. They acknowledge that omitting relevant common causes is unavoidable, which breaks a core assumption of standard discovery methods, so they use LPCMCI, an extension designed for latent confounders and autocorrelation. They find that West Texas wind has become the dominant causal driver of day-ahead prices, with effects more than three times those of natural gas prices, and that rising load in South and West Texas is changing both system prices and regional price differences.

They also estimate separate graphs for peak and off-peak hours and warm and cool seasons, on rolling two-year windows, and they report that the causal structure of price formation is not static, so analyses assuming fixed relationships risk missing these shifts. That is exactly the fixed-graph problem the Jiang review warns about, handled directly.

Job two: how much does a data center move the price?

Discovery says an arrow exists. Planning needs its size. For this I built a separate hourly dataset of 20,000 observations. The treatment D is data-center load in units of 100 MW, the outcome P is local wholesale price, and the true effect is 4.0 $/MWh per additional 100 MW.

The confounders are deliberately nonlinear: temperature (cooling raises load, and price is U-shaped in temperature), hour of day, wind output (convex, lowers price), gas cost, and weekday. If the covariates X block every backdoor path, the adjustment formula identifies the causal effect.

Double Machine Learning (Chernozhukov et al., Econometrics Journal, 2018) handles the nonlinear confounding without letting regularization bias into the coefficient. It fits one flexible model predicting price from the covariates and another predicting load, uses cross-fitting so no observation is predicted by a model trained on it, and then regresses residual on residual:

Estimate = Sum(D_residual * P_residual) / Sum(D_residual squared)

P_residual = P — predicted_P(X) D_residual = D — predicted_D(X)

Results against a true effect of 4.0:

• Naive OLS, no controls: 10.04 (95% CI 9.90 to 10.18)
• OLS with linear controls: 7.59 (95% CI 7.44 to 7.75)
• Double ML, boosted trees: 4.13 (95% CI 4.04 to 4.22)

The naive regression overstates the effect by 2.5 times, because hot hours bring both high cooling load and high prices. Linear controls remove less than half of that bias. Double ML comes close, although its interval narrowly misses 4.0. That is a fair reminder that the method is consistent in large samples, not exact in any particular one.

The strongest published use of this approach in power markets is Cacciarelli, Pinson and colleagues, working with the UK system operator. Using a local partially linear Double ML approach on 2018 to 2024 data, they find wind’s effect on UK wholesale prices is U-shaped: at low penetration, an extra GWh of wind cuts prices by up to 7 GBP/MWh, but the effect nearly disappears at 20% to 30% penetration before strengthening again. Its value here is the method: the effect depends on the state of the system, which is exactly the heterogeneity that shows up in the data-center studies.

When the confounder can’t be measured

Double ML only removes confounding you observe. The reverse question, how price affects where data-center load goes, is confounded by things like fiber access, land, tax deals and water, most of which aren’t in any dataset. The usual tool is an instrumental variable: something that moves price but affects siting only through price. With a single instrument Z:

Effect = Covariance(Z, D) / Covariance(Z, P)

In my simulation, the instrument is a fuel-cost shock interacted with the share of gas on the regional margin. Across 600 regions with a true elasticity of -0.60:

• OLS: -0.03 (standard error 0.03)
• 2SLS, fuel shock: -0.61 (standard error 0.06), first-stage F = 557

OLS says price barely matters, because the unobserved regional pull raises both price and load. The instrument recovers the true effect. You can check instrument strength (the F statistic of 557). You can’t test the exclusion restriction, the assumption that the instrument affects the outcome only through price. That has to be argued from domain knowledge.

Real studies use three kinds of variation of this sort:

• Randomized encouragement: The weatherization study randomized who was encouraged to participate, which works as an instrument for actual participation.
• Shift-share instruments: Alvarez and coauthors use this approach for data centers.
• Natural experiments: Heiss and coauthors treat the ChatGPT release as an outside shock to expected demand.

Each one depends on an argument about exclusion, not on anything the data alone can prove.

Stress-testing the estimate

Following the Jiang review’s recommendation, I ran three refutation tests on the Double ML estimate:

• Double ML estimate: 4.13
• Placebo treatment (permuted D): 0.03 (should vanish)
• Random common cause added: 4.14 (should hold)
• Random 50% subset: 4.07 (should hold)

All three pass. Then I added an unobserved confounder U that raises both load and price, left it out of the model, and turned up its strength:

• U strength 0.00: Estimate 4.04
• U strength 0.25: Estimate 4.17
• U strength 0.50: Estimate 4.56
• U strength 0.75: Estimate 5.12
• U strength 1.00: Estimate 5.50
• U strength 1.25: Estimate 5.88
• U strength 1.50: Estimate 6.22

Every one of those biased estimates passes the same three refutation tests. A permuted treatment always gives zero, and a random covariate never changes anything. Refutation tests catch coding errors and fragile specifications. They can’t detect a confounder you never measured. That is the job of formal sensitivity analysis, such as Cinelli and Hazlett’s omitted-variable bounds (2020), combined with domain knowledge about whether a confounder that strong is plausible.

The same caution applies to difference-in-differences. Pre-trend tests are the standard check, and they aren’t always clean. In Zhang’s preferred sample, treated counties had slightly cheaper wholesale prices than controls in the earliest pre-periods, though not significantly different closer to treatment. That is honest reporting and exactly what readers should look for.

The assumption energy breaks by design: interference

Standard causal methods assume one unit’s treatment doesn’t affect another unit’s outcome. This is the Stable Unit Treatment Value Assumption (SUTVA). Grids violate it by construction. Power flows follow Kirchhoff’s laws, and markets transmit price shocks across regions. The Jiang review notes that SUTVA is rarely discussed explicitly in energy work, and none of the studies it reviewed formally test or relax it.

The newer data-center studies are starting to deal with it directly:

• Yue and Zeng: In counties with more spillover, their estimates weaken or lose significance, which they interpret as a measurement problem rather than evidence of no effect.
• Zhang: He restricts control counties to second-order neighbors of treated counties, because immediate neighbors may share affected grid nodes and would be contaminated controls, and he uses node-level prices to check how far spillovers reach.
• Wikle and coauthors: Covered in the Jiang review, they go further for power plant emissions. They split treatment into a direct component and an upwind component and estimate the interference structure from atmospheric data rather than assuming it.

For grid questions, the equivalent would be to estimate spillover structure from power flow sensitivities or network topology, instead of treating counties as if they were islands.

Closing the loop: counterfactuals with feedback

Back to the full dynamic model. Suppose regulators approve a sustained increase in grid investment starting at time step 20. What happens to prices, load and adoption?

I answer this two ways, using the same noise draws so the comparison is a unit-level counterfactual (in a linear model this matches the impulse response). The first version uses the full feedback model. The second treats electricity supply as exogenous, meaning load no longer moves price or triggers investment. That second version is the implicit assumption in most AI diffusion models, which treat power as a given input.

For a sustained shift, the long-run effect has a closed form:

Results With Feedback:

• Electricity price: -0.49
• AI adoption: +1.50
• Data-center load: +1.21
• Regional development: +1.91

Results With Supply Exogenous:

• Electricity price: -0.77
• AI adoption: +1.41
• Data-center load: +1.26
• Regional development: +1.49

The exogenous model overstates long-run price relief by about 57 percent. In the feedback model, prices first fall almost as far, then climb back as cheaper power attracts more load. The exogenous model also understates regional development by about a quarter, because it ignores the extra investment that new load triggers. Neither version gets the direction wrong. Both get the magnitude wrong in ways that would matter in a rate case.

The real policy responses show the same loop. Curtailment requirements in Texas, the renewable-matching condition in Ireland, and PJM’s proposals for separate data-center procurement are all interventions on specific edges of the graph. Each changes the effect sizes on the other edges. Judging any one of them without modeling the loop answers a question nobody is asking.

A checklist for AI and energy modeling

Drawing on the review’s recommendations, the empirical literature, and the simulations:

• Write down the graph before fitting anything. Even a rough diagram forces you to name the confounders. If you use discovery to build it, report which variables were left out and use a method that allows for latent confounding.
• Say which estimand you’re after. Wholesale or retail. First entry or marginal capacity. The effect on treated regions or on a randomly chosen one. Most apparent contradictions in the data-center literature come from mixing these up.
• Take selection seriously. Data centers choose their sites. Model that choice, restrict comparisons to regions that plausibly could have been treated, or find variation the developer didn’t choose.
• Expect heterogeneity. Grid headroom, market design, and whether a project brings its own generation can all flip the sign of an effect. Report effects by subgroup, not just an average.
• Report effect sizes with uncertainty, not just accuracy metrics. A good forecast doesn’t validate a causal claim.
• Run a placebo or refutation test, then a sensitivity analysis. The first catches mistakes. The second is the only defense against confounders you didn’t measure.
• Handle interference explicitly. Use network-aware control groups, aggregate to a level where spillovers stay inside the unit, or at least state the likely direction of bias.
• Model the feedback, or say you didn’t. A single-equation estimate describes one edge of the loop, not the effect of a policy.

Limits of this piece

The simulations are linear and Gaussian where discovery is concerned, and their confounding is known by construction. Real data centers cluster for reasons that change over time, prices are set by market rules with discontinuities, and grid investment arrives in lumps years after the signal that justified it.

Most of the data-center studies cited here are 2026 working papers. Zhang’s preferred sample, for example, has eight treated counties. Their estimates will change as data and peer review accumulate. The Jiang review is a methods survey and doesn’t provide evidence for any specific AI and energy effect.

The honest summary of the evidence right now: data centers raise wholesale prices in constrained markets like post-2023 PJM, retail effects are smaller and sometimes negative where grids had headroom, and the answer for any given region depends on conditions that change as the load arrives. That’s why the tools matter. When AI load is large enough to move the system it depends on, “what happens if we build this?” becomes a causal question, and it deserves causal methods used carefully.

Code

The companion script covers every simulation and figure in this article: the structural causal model, PCMCI-style discovery, Double ML, instrumental variables, refutation and sensitivity tests, the selection and heterogeneity simulation, and the feedback counterfactuals. It runs on numpy, scipy, scikit-learn and matplotlib, and it’s available to paid subscribers along with the figure files at https://brianjamescurry.substack.com/. The next piece applies these methods to real PJM and ERCOT data.

About the Author

Brian Curry is a Kansas City-based Machine Learning & AI Strategist, Marketing Economist, and Applied Researcher. He drives Vector1, an independent research studio working at the frontiers of cognitive AI, causal inference, marketing economics, and knowledge engineering to build simulation-driven intelligence systems.

Across more than twenty years, Brian has led advanced data science and analytics initiatives at enterprises ranging from $4B to $125B in revenue, including Koch Industries, Tractor Supply, Vail Resorts, Hallmark, Garmin, and AT&T. His work systematically applies economic thinking and causal inference to connect marketing investment to structural outcomes, operational decisions to economic value, and AI adoption to permanent competitive advantage.

He is also the founder of KC AI Lab, a 1,500-member AI/ML education community in the Midwest.

Key Research & Open-Source Projects

GitHub: @Bodhi8

Substack: https://brianjamescurry.substack.com/

Causal Inference & Marketing Economics

• PyCausalSim: A Python framework for simulation-driven causal discovery and evaluation, mapping counterfactual scenarios to isolate true structural drivers.
• Papilon: A complex-system simulation and optimization library optimized for Media Mix Modeling (MMM) and adaptive architectures.
• Daedalus: A causal analytics framework designed for money-flow mapping and macro-economic tracking.

Algorithmic Engineering & Applied Forecasting

• Regimecast: Applied research and tooling for time-series regime detection, structural breaks, and economic forecasting.
• CarryNet: Open-source research and financial modeling focused on commodity carry trade analysis.
• Foodmesh-ml: A machine learning, optimization, and simulation framework engineered for modeling distributed food systems.

Cognitive AI & Semantic Knowledge Architecture

• Memory-Node Encapsulation (MNE): A cognitive AI architecture and novel data structure for artificial episodic memory (NeoCortex-M).
• The Autonomy Economy: A macroeconomic framework projecting the dual impacts of massive productivity growth and extreme structural price deflation in an AI-automated world.
• MeaningFlow: A semantic intelligence framework focused on content modeling, embeddings, clustering, and automated coverage gap analysis.

Contact: brian@vector1.ai | GitHub | Substack

This article may be shared with attribution under Creative Commons Attribution 4.0 International (CC BY 4.0).

References

• Alvarez, F., Argente, D., Chow, J., and Van Patten, D. (2026). Data Centers and Local Economies in the Age of AI: A Shift-Share Approach. NBER Working Paper 35194.
• Bloomberg News (2025). AI Data Centers Are Sending Power Bills Soaring. September 2025.
• Burlig, F., Knittel, C., Rapson, D., Reguant, M., and Wolfram, C. (2020). Machine Learning from Schools about Energy Efficiency. Journal of the Association of Environmental and Resource Economists, 7(6), 1181 to 1217.
• Cacciarelli, D., Pinson, P., Panagiotopoulos, F., Dixon, D., and Blaxland, L. (2025). Do we actually understand the impact of renewables on electricity prices? A causal inference approach. iEnergy, 4, 247 to 258.
• Central Statistics Office Ireland (2026). Data Centres Metered Electricity Consumption 2025.
• Chernozhukov, V., et al. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1).
• Cinelli, C., and Hazlett, C. (2020). Making sense of sensitivity: extending omitted variable bias. Journal of the Royal Statistical Society Series B, 82(1).
• EPRI (2026). Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States.
• Fowlie, M., Greenstone, M., and Wolfram, C. (2018). Do Energy Efficiency Investments Deliver? Evidence from the Weatherization Assistance Program. Quarterly Journal of Economics, 133(3), 1597 to 1644.
• Heiss, A., Sautner, Z., and Schmid, T. (2026). Powering AI: How Do Data Centers Affect Renewable Energy Investment? Swiss Finance Institute Research Paper 26–29.
• IEA (2025). Energy and AI. International Energy Agency.
• Jiang, F., De Vos, M., Wang, Z., Deconinck, G., and Kazmi, H. (2026). Causal inference and machine learning in power and energy systems: A review of assumptions, applications and opportunities. Energy and AI, 25, 100774.
• Madadkhani, S., Sturma, N., Drton, M., and Ikonnikova, S. (2026). Mapping the causal structure of price formation in Texas’s transitioning electricity market. arXiv:2604.14257.
• Monitoring Analytics (2026). State of the Market Report for PJM, January through June.
• NERC (2025). Incident Review: Considering Simultaneous Voltage-Sensitive Load Reductions.
• Norris, T., Profeta, T., Patino-Echeverri, D., and Cowie-Haskell, A. (2025). Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems. Nicholas Institute, Duke University.
• PJM Interconnection (2026). 2028/2029 Base Residual Auction Results. July 14, 2026.
• Runge, J., Nowack, P., Kretschmer, M., Flaxman, S., and Sejdinovic, D. (2019). Detecting and quantifying causal associations in large nonlinear time series datasets. Science Advances, 5, eaau4996.
• Shehabi, A., et al. (2024). 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory.
• Wiser, R., et al. (2025). Factors Influencing Recent Trends in Retail Electricity Prices in the United States. Lawrence Berkeley National Laboratory and The Brattle Group.
• Yue, D., and Zeng, Y. (2026). The Local Economic Effects of Data Center Entry. Working paper.
• Zhang, C. (2026). AI data centers and electricity prices: Much ado about something? UCLA working paper.