Aug 12, 2026 · Research · 6 min
The Scarce Thing in 2126 Is the Right to Use a Watt
This paper argues that the cost of cognition can fall for a century without relaxing two local constraints. Heat must be rejected at the site where computation occurs. A watt drawn from a shared grid is a claim on a machine that other parties cannot halt. On that view, 2126 is not organized around minds. It is organized around the right to take and dump energy at a node. The claim does not require generation to stay scarce. It requires heat rejection and interconnection rights to stay site-specific. If they do not, the claim is false.
Can get cheaper
- Models and tokens
- How the electricity is generated
Stays stuck to a place
- Heat leaving the building that used the watt
- The legal right to take that watt from the grid
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
stays at the site"] 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
| A finished model | A watt at a grid node | |
|---|---|---|
| Can you copy it? | Yes. Another copy costs storage. | No. The watt at bus A is not at bus B. |
| Does someone else have to agree? | No. | Yes. The grid is shared and cannot be turned off. |
The heat constraint is first-law accounting, not a slogan about Landauer. Lawrence Berkeley National Laboratory states the point without remainder: all electrical input power to IT equipment is converted into heat that must be removed by cooling.1 PUE measures facility overhead on top of that heat. It does not cancel the IT watts. A rack that draws P still presents about P as heat at the site if those watts were made by gas, photovoltaics, or a reactor a hundred kilometres away.
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
| Limit | What it means | Does it decide where compute sits in 2126? |
|---|---|---|
| Landauer (kT ln 2) | Theoretical minimum heat to erase a bit. Real chips are far above it. | No. |
| Switching energy (½CV²) | Why CPU clocks stopped rising around 2004: too much heat on the chip. | No. That is a chip problem, not a siting problem. |
| Site heat = the watts used | Every watt into the building becomes heat that must leave that building. | Yes. Cheaper power plants do not move the heat. |
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. Efficiency per task can keep falling. 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 right constraint is institutional. A shared interconnection is not a private generator. Taking load changes voltages, flows, and contingency margins for everyone else on the island. That is why interconnection is a study-and-permit process. Cheaper distant generation does not mint a position on a bus or a thermal budget in a basin. The IEA's 2025 assessment is already this argument in contemporary form: data centres can match smelter-scale loads but are far more geographically bunched; nearly half of U.S. capacity sits in five regional clusters; new transmission in advanced economies takes years; onsite gas is being tried under grid delay and still does not remove the need to fix interconnection.6 By 2027, the Agency writes, an AI rack the size of a refrigerator may need to evacuate heat equivalent to 30 natural-gas boilers.5 That is a siting problem. It is not a training-run problem.
Exhibit
The cleanest public measurement of node-demand outrunning plant and permit is ERCOT's large-load queue. 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. Approved to energize is 8,926 MW. No studies submitted is 257,600 MW. Observed energized is 1.2 percent of requested megawatts.
These layers are not a construction schedule. They are a request ledger against a reliability island. 257,600 MW has not submitted studies. 5,700 MW is the energized tier in the projection table; it is not the 3,900 MW June non-simultaneous peak of already-approved load printed on a different slide. The anatomy of that ledger is a companion article. The interpretation this paper needs is narrow: by June 2026, asked-for interconnection in ERCOT exceeded observed energized load by two orders of magnitude. That is what node scarcity looks like while cognition is already cheap enough to generate the requests.
Implications the mechanism licenses
If heat remains a site quantity, industrial compute sits where rejection is cheap: cold water, coast, high latitude, or purpose-built radiator mass. Interior basins without a heat host keep people and lose load. A campus is then a node plus a thermal contract. The processors are replaceable. The right and the sink are not. This is the mill-and-river pattern under a different working fluid. It does not require a new skyline. Atoms stay slow.
If a watt on a bus remains a claim on a non-halting machine, interconnection stays a legal object after generation is abundant. The scarce capital good is not the model. It is a current right to take load at a location, under rules that can be changed without collapsing the island. Firms that only produce copyable cognition shrink toward the cost of compute they can actually energize. Status, in the narrow sense of who can bind the shared machine, attaches to whoever holds the node, the sink, and the authority to sign an irreversible act.
If rule changes must occur without a blackout, large always-on systems stay bounded. A single planetary bus is a poor place to version reliability rules. ERCOT is an existence proof of an interconnection island with its own queue, protocols, and commission. The prediction is more islands, not fewer, and industrial capacity tracking those islands rather than model-training rank. Flags remain. They sit on the node map; they do not replace it.
| If this stays true | Then this follows |
|---|---|
| Heat has to leave the building that used the watt | Big compute sits next to cold water, coast, or purpose-built cooling, not wherever the model was trained. |
| Taking a watt from the grid needs a permit | Cheap generation does not skip interconnection. The scarce thing is the right to plug in. |
| The grid cannot be shut down to change the rules | Rules get changed inside bounded islands (ERCOT is one), not on one planetary grid. |
None of these implications licenses a forecast of wages, fertility, or the death of the firm. They are what local heat and a non-halting grid already imply once cognition is no longer the scarce input.
Kill tests
The argument dies if any one of the following becomes true in the record, not in a slide.
| The paper is wrong if | This does not make it wrong |
|---|---|
| Heat from a building can be dumped somewhere else at almost no cost, so location stops mattering. | Better cooling, lower PUE, or reusing waste heat nearby. |
| A large grid can be routinely shut down and restarted to change the rules. | Batteries, demand response, or onsite generators that keep the grid up. |
| The right to take a watt at a specific grid node can be copied like a file. | Cheaper power plants, cheaper chips, or cheaper AI queries. |
Heat rejection stops being local if site thermal load can be moved off-site at negligible cost. Shared machines halt cheaply if large synchronous islands can be taken down and restarted as routine operations. A watt at a node copies if the right to take load at a specified bus becomes as non-rival as a weight file. Demand response, liquid cooling, and cheaper chips satisfy none of those.
Abundant fusion, orbital solar, or another 10,000-fold drop in energy per task does not, by itself, satisfy any test. Each of those can hold while heat is still dumped at the rack and while a bus position is still a permit. That is the content of the claim. The June 2026 ERCOT ledger is consistent with it. It does not prove 2126. It shows the constraint operating while thought is already cheap. If later primary sources show heat or rights becoming non-local, this paper is wrong. Until then the scarce object remains the right to use a watt.
Notes
1. S. Greenberg et al., "The Installation of Direct Water-Cooling Systems in an Operating Data Center," LBNL / ACEEE (2020): "All of the electrical input power to the IT equipment is converted into heat that must be removed by appropriate cooling." 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; 17% and 50% demand growth in 2025; rack heat "equivalent to 30 natural gas boilers." 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.