When we launched Cobble, we made a claim that was easy to say and hard to prove: that inference could be built from what the world discarded, and that it could run on power the world was not already fighting over. Reclaimed GPUs took care of the first half. Today we can report on the second.
Phase I of the Sustainable AI Initiative is complete. Ten percent of the power behind every completion served on the Cobble Network now comes from solar generation. Not ten percent of a pilot cluster, and not ten percent on a good afternoon: ten percent of the fleet's total completion power, measured across the hardware that answers your requests.
What Phase I Changed
The first phase was about proving the model rather than maximizing the number. We needed to show that a distributed fleet of reclaimed hardware could be fed by on-site solar without compromising the thing you actually pay for: completions that arrive quickly and reliably.
Over the past months we have:
- brought solar generation online for the first slice of the fleet, feeding the GPU pools that serve the public catalog;
- metered that generation against actual inference load, so the ten percent figure reflects power consumed by completions, not nameplate capacity on a datasheet;
- kept the fleet on air cooling with no evaporative water use, so adding generation did not add water consumption; and
- carried the measurement into our operations tooling, where it now sits beside utilization, latency, and cost so that every capacity decision sees it.
Ten percent is a modest number on its own. It is a large number for a fleet that was built from hardware other operators had already written off.
A Third of the Power per Completion
The solar share is only half of the story. The other half is how much power a completion needs in the first place.
By our measurements, a completion on Cobble already uses roughly one third of the power of a comparable completion at a typical industry provider. That comes from choices we made before the first panel went up: serving open-weight models at sizes matched to the work, caching repeated prompt prefixes so warm requests do not recompute what they already know, running the fleet at high utilization rather than keeping idle capacity hot, and avoiding the energy cost of manufacturing new silicon altogether.
Put the two numbers together and the picture is this: every completion starts from a third of the energy, and a tenth of that third now comes from the sun. Phase II is about moving that tenth.
Phase II: Sixty Percent by Q2 2027
Phase II has one headline target. By the second quarter of 2027, we plan for sixty percent of completion power across the network to come from solar.
Getting there means more generation, more storage so that the share holds after sunset, and scheduling that leans into daylight hours where the workload allows. It also means growing the fleet without growing its footprint: every server we add is reclaimed, and every new rack is sited with generation in mind rather than bolted on afterward.
We will publish progress the same way we published Phase I: as a measured share of completion power, reported against real load, with the methodology visible to anyone who asks.
Why This Matters to You
If you build on Cobble, nothing about your integration changes. The API is the same OpenAI-compatible endpoint, the models are the same open-weight catalog, and the price list is the same published per-token rates. What changes is what sits behind the request.
Sustainability on Cobble is not a surcharge, a separate tier, or an offset purchased elsewhere. It is a property of the infrastructure itself: reclaimed hardware, a fraction of the energy per completion, and a growing share of that energy generated on site. You get it on every plan, for every model, by default.
Thank you to everyone building with us. Phase I was the proof. Phase II is the scale.
The Cobble Team