Two numbers landed in the same week, one at grid scale and one on a bench, and they are the same kind of number. Utilities in 53 US service territories report 537 GW of data center load. They have signed contracts for about 120 GW of it, which is 22 percent. On my own machine, a power limit set to 100 W produced a measured draw of 161 W and the same tokens per second as a limit set to 150 W. Both are the same failure: a stated limit read as a measured one.
What was actually published
Latitude Media published its figures on November 13, 2026, through Latitude Intelligence, a research platform it launched this month that tracks data center load by stage: operating, signed, and reported. The definitions do most of the work here. Load counts as signed only when a customer has an executed agreement to take service. Some utilities report every inquiry, others only executed agreements, and the published totals mix both. Applying one definition across 53 territories gives 120 GW signed against 537 GW reported.
The signed figure is still about 5.5 times the 22 GW of data center load operating in those same territories today. So this is not a story about nothing happening. It is a story about what counts as happening.
Jefferies reached the same shape through a different method, sorting 187 GW of disclosed commitments across regulated utilities into tiers by contract strength. Under the strictest tier, formal electric service agreements only, the total fell to 69 GW. Two independent methods land in the same band: 22 percent of reported load under Latitude’s definition, and 37 percent of disclosed commitments under Jefferies’ strictest tier, which is 69 of 187 GW, my own arithmetic on their figures. A single utility figure quoted without its tier is not comparable to any other.
The instrument decides what the number means
An electric service agreement is a service contract. A letter of agreement is a commitment to move toward one. Those are not the same object, and the largest published figure in the set contains both.
AEP publishes 69 GW of contracted load growth through 2030, supported by a mix of fully executed ESAs and letters of agreement. The mix is not uniform: the 45 GW in AEP Texas is backed by letters of agreement, while the company’s claim of $16B in residential cost offsets attaches specifically to fully executed take-or-pay ESAs at its vertically integrated utilities, which are not in Texas. The headline number contains its own two tiers, and the instrument that makes contracted load meaningful for an equipment buyer attaches to the smaller tier.
Where signed load exceeds a utility’s current peak demand, the contract is for power the utility does not have yet. AEP Ohio’s signed load equals 235 percent of its peak. NIPSCO in northern Indiana is at 169 percent, PPL in Pennsylvania at 145 percent. Those are executed agreements against capacity that has to be built, and the cost allocation for that build is still open. FERC issued show-cause orders to six regional grid operators in June over the tariff rules governing large loads. Responses from five operators, including PJM and MISO, were due November 16, 2026, and Southwest Power Pool’s are due November 20.
At bench scale, a power limit is a request to a governor
The same category error runs at the machine level. nvidia-smi -pl 150 does not set a ceiling on measured power. It tells the clock governor what to aim for. Whether the card can honor that request depends on what the governor has left to cut. On the card I measured, it ran out of things to cut at roughly 160 W.
I measured the same trade in October on this hardware and got the direction right: what power capping actually costs a local LLM build. This run extends it downward, and the curve does not stay linear. Below a floor, the setting stops doing anything.
The measured curve
Setup: a Ryzen 9 5900X serving a 27B model at Q5_K_M through llama.cpp on an RTX 3090, confirmed over SSH before writing this. The requests were sent from a different machine, a Ryzen 7 2700X host, so every wall-clock figure here includes the network round trip. Power and clock samples were read on the serving box. One request, 400 tokens, single stream, no batching, draw sampled at 1 Hz during each request. Energy per token is arithmetic on a board-level estimate, not a meter reading.
| Power limit set | Engine decode (t/s) | Measured draw (mean) | SM clock | J per token |
|---|---|---|---|---|
| 370 W (stock) | 34.96 | 360.1 W (7 samples) | 1650 MHz | 10.3 |
| 200 W | 14.76 | 188.0 W (15 samples) | 429 MHz | 12.7 |
| 150 W | 7.91 | 161.8 W (29 samples) | 226 MHz | 20.4 |
| 120 W | 7.91 | 161.7 W (2 samples) | 225 MHz | 20.5 |
| 100 W | 7.91 | 161.1 W (2 samples) | 225 MHz | 20.4 |
The table rows are single passes per setting. Three passes at the stock limit gave a median of 9.97 J per token (8.78, 10.3, 9.97). Three passes at 150 W gave a median of 20.72 J per token (20.72, 20.79, 20.68), a pass spread of 0.5 percent. The 370 W row is looser because the busy window is only 11.4 seconds, so fewer samples land inside it. Sample counts are printed for that reason.
Where the cap stops working
At 200 W the card honored the request: measured draw came in under the limit at 188 W, and the SM clock dropped from 1650 MHz to 429 MHz. At 150 W, 120 W, and 100 W the measured draw stayed between 161.1 and 161.8 W regardless of the setting, and throughput was identical at all three: 7.91 tokens per second, 126.4 milliseconds per token. The limit was accepted and reported as set, and it did nothing.
The memory clock explains why. It stayed at 9346 to 9501 MHz across every setting, including the 100 W one. A decode step is a memory operation before it is a compute operation, so the power a governor can remove is the compute power. Once the SM clock is at 225 MHz, what remains is the memory subsystem, the board, the VRM, and the fans. That is the floor, and on this card it sits near 160 W. A power cap below the floor is a number in a settings panel.
This is the practical version of the grid argument. A curtailment commitment is only worth something if the facility can actually bend. A power cap is only worth something if the hardware can honor it. Both are measured, not declared.
Energy per token got worse, then stopped changing
Power capping is a demand measure, not an efficiency measure. That was already the conclusion in the earlier run, and the extended curve confirms it. Energy per token rose from 10.3 J at the stock limit to 12.7 J at 200 W and 20.4 J at 150 W, then stopped moving at 100 W. Per million tokens that is 2.86 kWh at the stock limit, 3.53 kWh at 200 W, and 5.67 kWh at the floor. The shape is the point: capping moves the peak at the meter and costs energy per token, and below the floor it does neither.
The same distinction shows up in the load path. On a warm slot, wall-clock throughput matched the engine’s reported rate within 1.3 percent: 34.5 against 34.96. The first request after a model swap took 16.56 seconds wall-clock for 32 tokens, of which 0.9 seconds was generation. The other 15.7 seconds was loading the model into the slot. An agent loop that switches models per call pays that 15.7 seconds every time, and no benchmark table figure includes it. The energy-per-token framing at facility scale is covered in what a 24-month power timeline says about energy per token; at bench scale the equivalent number is the one above, and it is only defensible if you can show the meter.
What this means for a local build
First, a cap below the floor is decoration. For a real reduction, set the cap above the floor and verify it holds during generation, not at idle. On this card the useful range is roughly 200 W and up. Below that, the reported limit and the actual draw disagree.
Second, deferrable work is the only work that can be capped honestly. Batch inference, embeddings, indexing, and fine-tuning can run at a capped power because they have no deadline. An interactive agent call does. That is the same asymmetry that makes facility flexibility a range rather than a commitment, and it is why the market pays for speed to power rather than for gigawatts on paper. ECL’s Santa Clara facility reached full commitment across three tenants eight months after its April 2026 announcement and before construction finished, because its purchase-order-to-keys timeline is nine to 12 months against the two to three years typical of grid-dependent development.
Third, the number you can defend is the one at the outlet. Board power, fans, PSU losses, and the CPU are part of the bill, and a GPU’s reported draw is an estimate the driver makes about its own board. A reduction claim that has to survive a utility or your own electricity bill needs a meter on the circuit.
What to buy if you want to prove a reduction
A plug-in meter is enough for a single machine. The MECHEER watt meter plug is about $15.97 as of November 16, 2026, and reads watts, kWh, volts, amps, power factor, and cumulative cost across seven display modes. It holds data through a power failure, which is what makes a before-and-after comparison survive one.
For continuous logging rather than a display, a Kasa Matter smart plug with energy monitoring is about $18.99 for a two-pack and reports current draw into Home Assistant or Apple Home, which turns a capped run into a time series you can show someone.
For the whole panel, a Tuya WiFi power meter with two 80 A current transformer clamps is about $26.99 as of November 16, 2026. It measures 0.2 to 80 A per channel and can meter generation and consumption in opposite directions on the same device, which is the only way to separate what a machine used from what the panel delivered.
None of these change the throughput curve. They make the claim verifiable, which is the part an auditor checks. If you are choosing the machine itself, the tiers are covered in the best hardware for running local LLMs in 2026, and the choice that matters here is not which card but whether you can measure it.
Check before you claim anything
Set the cap, run a request, and read the draw during generation. If the measured draw sits above the setting, the cap is not binding and the number in your settings panel is not a result. Report engine throughput and wall-clock throughput separately, name which loop produced each, and report a median over several passes. A gigawatt figure without its contract tier is unusable, and a watts figure without a meter is the same kind of claim.

