Aug 12, 2026 · Research · 8 min

Computing costs and the limits of grid access

This paper presents a conditional scenario for 2126, not a forecast. Computing may become cheaper while heat removal and access to a shared grid remain important local constraints. Heat can be transported and reused; the constraint is the engineering and economic cost of removing it from equipment and delivering it to a suitable sink. The thesis weakens if those costs or interconnection constraints cease to matter.

Mechanism

A completed inference is information. Once produced, it can be copied at the cost of storage and transmission. A watt at a node is not information. It is rival: the watt used at bus A is not available at bus B. It is also institutional: on a synchronous grid it is a claim against a reliability machine that has to keep running while the claim is exercised. Those two properties do not travel with model weights.

graph TB
  G["Generation
anywhere, getting cheaper"] --> R["Interconnection right
one bus, permitted"]
  R --> C["Compute
at a site"]
  C --> I["Inference
copies anywhere"]
  C --> H["Heat
must be removed"]

  style G fill:none,stroke:#0a0a0a
  style C fill:none,stroke:#0a0a0a
  style I fill:none,stroke:#0a0a0a
  style R fill:#e8e8e8,color:#0a0a0a,stroke:#0a0a0a
  style H fill:#e8e8e8,color:#0a0a0a,stroke:#0a0a0a
          

For ordinary IT equipment, electrical input becomes approximately an equivalent thermal load that cooling must remove. The LBNL study cited below discusses this engineering relationship.¹ PUE measures total facility energy relative to IT energy; lower overhead does not eliminate the IT heat. The heat need not remain in the building: cooling loops and heat recovery can move it. The location and cost of that transfer are part of the siting problem.

Landauer’s 1961 result is a different bound and is routinely misused. Irreversible logical operations require heat of order kT per operation, now stated as kT ln 2 per erased bit.² Landauer wrote that this dissipation was already many orders of magnitude smaller than that of any practically conceivable device. Theis and Wong show why CMOS never approached it: circuits store many times kT to outrun thermal noise, then dump ½CV² on each switch. The post-1990s slowdown in switching energy “has nothing to do with the approach of switching energy toward” kT. Clock frequency plateaued around 2003-2005 because of power and heat.³ IEEE’s IRDS still treats power density and interconnect as first-class scaling limits, including in 3D stacks with poor heat conductance.⁴

Two conclusions follow, and they should be kept separate. First, today’s machines run hot for engineering reasons (voltage, capacitance, off-chip memory), not because they have hit a thermodynamic floor. Energy required per task can continue to fall. The IEA reports that energy use per AI task has dropped by at least an order of magnitude annually in recent years, while video, reasoning, and agentic tasks consume hundreds to thousands of times more energy per query than simple text.⁵ Cheap thought and rising aggregate watts are compatible. Second, even a hypothetical machine that approached Landauer, or a reversible machine that evaded it, would still reject whatever energy it does use as heat at a place. Reversible computing is not a license to ignore geography.

The other constraint is institutional. A connection to a shared grid involves network studies and operating requirements, not only access to generation. The IEA’s 2025 assessment describes geographically concentrated data-center loads and lengthy grid expansion.⁶ Its 2026 update projects sharply higher rack power density and discusses bottlenecks in equipment and grid connections.⁵ These observations support examining local infrastructure. They do not establish a century-long forecast.

Exhibit

ERCOT’s large-load queue provides a dated example of requests exceeding observed consumption. The June 19, 2026 LLWG report, data as of June 18, prints 466,497 MW requested and 5,700 MW in the Observed Energized tier. The approvals chart prints 8,926 MW; slide text prints 8,927 MW. No studies submitted is 257,600 MW. Observed energized is 1.2 percent of requested megawatts.

These layers are not a construction schedule. In the June snapshot, the 5.7 GW Observed Energized tier uses all-time non-simultaneous peak consumption, while 3,900 MW is the separate June monthly peak figure. The chart’s 8,926 MW approval total differs by one MW from slide text. The companion article explains the source discrepancy and time windows. The roughly 82-to-one request-to-observed ratio describes this ledger; it does not by itself prove physical scarcity or quantify viable demand.

Conditional implications

In the scenario, locations with economical cooling, water access or useful heat customers could gain an advantage. That does not imply that every coastal site wins or every interior basin loses. Climate, water constraints, transmission, land, permitting and workload flexibility can change the result. A campus can be analyzed as a combination of computing equipment, an electrical connection and a heat-removal system.

If grid access remains constrained after generation becomes abundant, the right to take load at a particular location could retain economic value. That is an institutional hypothesis, not a consequence of thermodynamics alone. Rules, network investment, storage and demand flexibility can change both the scarcity and who captures its value.

ERCOT illustrates an interconnection with its own operating rules and queue. It does not prove that the world will divide into more electrical islands. One scenario to test is that local governance may remain important, while the costs and benefits of additional interconnection could favor larger networks instead.

These scenarios do not establish a forecast of wages, population or firm structure. Heat removal provides a physical constraint. The institutional and economic outcomes depend on additional assumptions about technology, policy and demand.

Tests that could weaken the thesis

Test the proposed mechanism against observed changes in infrastructure and operating costs.

Useful tests would measure whether heat transport or reuse makes location less important, whether grid investment and flexibility materially reduce access constraints, and whether workloads can move cheaply between locations. Cooling and demand response can relax constraints even when they do not eliminate the underlying physics. The thesis should change with that evidence.

More abundant generation or lower energy per task could coexist with local constraints, but that possibility does not establish that the constraints will remain binding. The June ERCOT snapshot is a present-day example to analyze, not proof of 2126. The argument is strongest as a checklist for siting and infrastructure decisions, with the century-long outcome left conditional.

Notes

1. S. Earni, S. Greenberg and W. Johnson, “The Installation of Direct Water-Cooling Systems to Reduce Cooling Energy Requirements for High-Performance Computing Centers,” LBNL / ACEEE (2020). PDF.

2. R. Landauer, “Irreversibility and Heat Generation in the Computing Process,” IBM Journal of Research and Development 5, no. 3 (1961): 183-191. doi:10.1147/rd.53.0183. Open PDF: Caltech copy.

3. T. N. Theis and H.-S. P. Wong, “The End of Moore’s Law: A New Beginning for Information Technology,” Computing in Science & Engineering 19, no. 2 (2017): 41-50. doi:10.1109/MCSE.2017.29. Open copy: Berkeley E3S.

4. IEEE International Roadmap for Devices and Systems, 2024 More Moore. PDF.

5. IEA, Key Questions on Energy and AI, executive summary (2026). Energy per task, changing task mix and projected rack power density. IEA.

6. IEA, Energy and AI, executive summary (2025). No AI without energy; geographic clustering; smelter-scale loads. IEA.

Queue figures: ERCOT, June 19 LLWG Report, data as of 2026-06-18. SHA-256 of the file at that URL: cdb778341b4392cb6235603dd87b1822655f27e7ac80ad1cd7dd0ba882d82db3.

Factual review: October 26, 2026. Original publication date retained. Primary sources are linked alongside the revised claims. Queue figures remain a historical June snapshot.