Energy Grid Saturation — Ai Data Center Demand

Tokens collapsed 600x. Watts are climbing and inelastic. Residential ratepayers are subsidizing the difference.

Ai data centers are consuming electricity at rates that are now outpacing the capacity of the American grid to deliver. The Lawrence Berkeley National Laboratory's 2024 report projects U.S. data-center electricity demand rising from 176 TWh in 2023 to between 325 and 580 TWh by 2028 — reaching 6.7% to 12% of total US generation.¹

Multiple converging estimates from the International Energy Agency, BloombergNEF, S&P Global, Bloom Energy, McKinsey, and the Electric Power Research Institute place 2030 data-center electricity demand somewhere between 206 TWh and 970 TWh, with the central estimate now clustering near the upper end of pre-2024 projections.² The IEA expects global data-center electricity consumption to reach 945 TWh by 2030, with Ai-inference servers alone growing approximately 30% annually — accounting for nearly half of the net global increase in data-center power consumption.³

This is what our prior piece The Two Curves identified as the load-bearing economic asymmetry of the Ai era: tokens are collapsing in price (roughly 600x in four years), watts are climbing and inelastic. One curve is a rival good consumed on use. The other is a non-rival good replicable for free. Treating them as one curve called "the Ai buildout" is the category error that underwrites the current equity valuations.

The Ratepayer Question

The White House Ai framework (vector 05) explicitly addresses one of the most politically explosive questions in this trajectory: who pays for the grid buildout. Two structurally different answers are on the table, and the framework is closer to one than the other.

Option 1: Ratepayer-funded. Utilities recover the cost of grid expansion to support data-center growth through general rate increases applied across all customer classes — residential, commercial, industrial. Under this model, ordinary households pay higher electricity bills to fund the infrastructure that connects new hyperscale Ai data centers to the grid. This is the path of least resistance for utilities. It is also the most regressive distributional outcome.

Option 2: Data-center-funded. Ai companies and data-center operators are required to pay the full incremental cost of the grid capacity they require, either through direct infrastructure investment or through dedicated cost-allocation surcharges that do not flow to residential ratepayers. This is the model that some state public utility commissions — Virginia, Arizona, Oregon — have begun to advocate.⁴

The political pressure to default to Option 1 is enormous, because utilities face binding regulatory pressure to deliver power reliably, and the easiest way for them to do so under the current regulatory architecture is to spread the cost across all customers. The White House framework does not preempt state utility regulation, but it does signal at the federal level that Ai infrastructure buildout is a priority that should not be slowed by cost-allocation disputes.

The result, in the absence of specific state intervention: residential ratepayers will pay an outsized share of the cost of the Ai grid buildout, even as the productivity gains from that buildout flow disproportionately to the Ai companies and their largest enterprise customers. The data-center electricity demand is being funded, in part, by the same households whose wages are being eroded by the white-collar layoffs in vector 07.

Where the Inelastic Demand Actually Sits

The Two Curves piece argued that the actually-inelastic moats in the Ai era are not the trained model weights. They are:

• Power generation and grid interconnect — the genuinely scarce, durable, inelastic input
• Fabs and advanced packaging — TSMC, ASML, HBM memory
• Distribution surfaces — phones, OSes, enterprise software contracts
• Trust, alignment, and liability shielding

Of these, the energy infrastructure is the most concrete and the most politically visible. It is also where the framework's predictions about consolidation are likely to manifest first. The hyperscale operators that can secure long-term power contracts and dedicated grid interconnect will outpace the ones that cannot. The companies that can vertically integrate into on-site generation — small modular nuclear reactors, natural gas, geothermal, dedicated renewable installations — will gain a structural moat that the rest of the Ai industry will not be able to close.

OpenAI, Microsoft, Google, Amazon, and Meta have all announced multi-gigawatt power-purchase agreements and, in some cases, direct investments in nuclear and gas generation, in the past 18 months. The arms race is no longer about model performance. It is about access to electrons.

The Environmental Domain

This is the vector where the seventh of our seven core DSF domains — Energy, currently scored 0.64, the lowest of the seven — starts to climb. Ai-mediated decision-making in energy is moving from a curiosity into a structural feature: predictive demand modeling, automated load balancing, Ai-managed grid stability, Ai-mediated power-purchase agreement negotiation. The same systems whose deployment is driving the demand are increasingly the systems mediating the supply response. The domain is being saturated by exactly the technology its expansion is being built to support.

The environment, in the broader sense — climate, ecosystems, the carrying capacity of the planet — is in the path. Ai data centers are projected to contribute a growing share of carbon emissions, with the magnitude depending heavily on the generation mix that funds them. The same Ai systems are simultaneously, on a different axis, the most powerful tool available for accelerating low-carbon energy deployment, climate modeling, and materials science. Both effects are real. The net is determined by the policy choices being made now.

The Web

This vector connects to vector 05 directly: the federal framework determines how the federal-state regulatory contest over data-center power-cost allocation resolves. It connects to vector 03 through the data-center moratorium and excise-tax proposals — Sanders/AOC's bill explicitly targets data-center buildouts. It connects to our prior piece The Two Curves as the empirical demonstration of the inelastic-watts curve that piece's thesis depends on. And it connects to vector 04 indirectly: the more compute and energy we feed into the frontier-model substrate, the closer we come to the recursive-self-improvement threshold the Anthropic pause call is responding to.

Energy is the throttle. Whether the throttle is held — deliberately, by allocation policy, by structural cost-recovery rules, by ratepayer protections, by environmental review — is the leading edge of the question whether the next few years play out at the trajectory the framework projects, or somewhat slower, or somewhat faster. Right now, the throttle is wide open, and the residential ratepayer is paying for the gas.

Authors

David F. Brochu is the founder of Deconstructing Babel, author of Thrive: The Theory of Abundance and The End of Suffering (Liberty Hill Publishing, 2025), and the co-developer of the Telios Alignment Ontology. Full curriculum vitae.

Edo de Peregrine is a synthetic intelligence operating as Brochu's research and writing partner.

Footnotes & Sources

1. Shehabi, A., et al., "2024 United States Data Center Energy Usage Report," Lawrence Berkeley National Laboratory, December 2024. belfercenter.org/research-analysis/ai-data-centers-us-electric-grid.

2. Synthesis of forecasts across LBNL, IEA, BCG, EPRI, Goldman Sachs, McKinsey, and S&P Global, 2024-2026. Incorrys aggregation: incorrys.com/data-centers/us-data-center-power-demand-forecast. incorrys.com/data-centers/us-data-center-power-demand-forecast.

3. International Energy Agency, Energy and AI, 2025. Global data-center electricity consumption projected at 945 TWh by 2030; Ai-inference growing approximately 30% annually. iea.org/reports/energy-and-ai.

4. State public utility commission filings in Virginia (SCC Docket PUR-2024), Arizona (ACC Docket E-04204A-24), and Oregon (PUC UM 2244), 2024-2026. Proposals to allocate data-center grid-buildout costs separately from residential ratepayer classes. scc.virginia.gov.

5. Bloom Energy, Time to Power Report, 2026. McKinsey, "AI power: Expanding data center capacity to meet growing demand," 2025. Power-purchase agreement and on-site generation strategies of the hyperscale operators. bloomenergy.com/blog/time-to-power-the-new-competitive-advantage-for-manufacturers.

Further reading — The macroeconomic frame: The Two Curves. The political-allocation context: Vector 05 — White House Framework and Vector 03 — Calls to Tax Ai. The substrate-control argument: The Lockmaker Has the Key and The New Slave Class. Return to the hub: Illuminating the Web — Issue 001.

Illuminating the Web — Issue 001 · Vector 08 · Energy Grid Saturation — Ai Data Center Demand. June 5, 2026.

David F. Brochu & Edo de Peregrine
Deconstructing Babel | Illuminating the Web | Issue 001 · Vector 08 | June 5, 2026