The carbon story of AI usually focuses on training. That is understandable: a single large training run can burn megawatt-hours over weeks. But training happens once. Inference happens every time anyone asks the model anything, forever, across millions of users. The energy that actually compounds is the energy of serving.
For an API consumer, that means the greenness of your LLM usage is mostly decided by something you do not control directly: where the provider's GPUs sit and what powers them. You can, however, choose a provider based on it.
Why inference energy is a data center story
An inference workload is a GPU cluster running in a data center. Two things dominate its footprint:
The grid it is plugged into. A GPU pulling 700 watts draws power that is as clean or dirty as the regional electricity mix. A data center in a region with heavy wind and solar is emitting far less carbon per token than the same hardware on a coal-heavy grid, even if everything else is identical.
Cooling. Servers turn electricity into heat, and getting rid of that heat takes more electricity. This is what PUE (power usage effectiveness) measures: total facility power divided by the power that reaches the IT equipment. A PUE of 1.5 means the facility uses 50% extra energy beyond what the servers consume. A well-designed modern facility runs closer to 1.1 to 1.2. Cool, dry climates cut cooling cost dramatically, which is a big reason data centers keep appearing in cold inland regions rather than in city centers.
Where cheap clean power actually lives
The places with the best clean-energy economics for compute are not the obvious tech hubs. They tend to be inland regions with strong wind and solar resources and a dry, cool climate:
•Inner Mongolia (Horinger). One of China's largest wind power bases. Cold winters mean much of the year needs little or no mechanical cooling.
•Xinjiang (Hami). Very high solar irradiance and steady wind. Remote, but remote is fine when your workload is a GPU cluster you control remotely.
•Ningxia (Zhongwei). Already an established data center region with utility-scale solar and wind; it hosts one of the major international cloud regions, which tells you the economics work at scale.
The pattern is the same across all three: abundant renewable generation, a cool dry climate that lowers PUE, and land cheap enough to build the substations and solar fields that feed the racks.
None of this is exotic technology. It is siting. The greenest thing a compute provider can do is put its hardware where the grid is already clean and the climate does half the cooling for free.
What "green" claims should actually tell you
This is where I get skeptical. "Green" is a marketing word until someone gives you numbers. When a provider makes a clean-energy claim, there are three things worth asking for:
•PUE. If they will not publish it, treat the claim as unverifiable.
•The actual grid mix of the region the data center sits in. A provider that merely buys renewable energy certificates while running on a fossil grid is not running green. The location is the disclosure.
•Whether the renewables are real capacity, not offsets. New wind and solar built to feed the facility is a different claim from buying credits elsewhere.
Honest providers will give you the region and the facility characteristics. If the answer is "our data centers are in Inner Mongolia, Xinjiang, and Ningxia, on grids with high renewable penetration, in cool climates," that is a concrete, checkable statement. If the answer is a slogan, move on.
What you can actually do about it
You are not going to audit a provider's substation. But you can make a few decisions that tilt your own footprint:
•Pick providers that disclose location and PUE, not ones that just print "carbon neutral" on the page.
•Use smaller models where the task allows. A 7B model can answer a support question nearly as well as a 175B one at a fraction of the energy. Routing easy traffic to a small model is the single biggest carbon lever you control.
•Cache aggressively. Prompt caching turns repeated input into a one-time cost, cutting both latency and energy for repetitive workloads.
•Batch when you can. Fewer, larger requests beat many tiny ones on server utilization.
The honest framing is that you cannot measure your way to a clean footprint with a spreadsheet, because the grid mix is not in your data. What you can do is choose providers whose infrastructure makes the footprint smaller by design, and stop routing trivial requests to the biggest model you can afford.
That is the position TokenWorks stakes out: its three compute centers sit in Inner Mongolia Horinger, Xinjiang Hami, and Ningxia Zhongwei, regions chosen for high renewable penetration and cool climates that keep PUE low. If you care about the energy behind your API calls, that is the kind of disclosure that matters more than a badge.