GPU compute for AI work, on hardware we own.
We run inference, agentic workloads, and research computing in the Pacific Northwest — buying the machines, powering them, and keeping them running ourselves.
01 — 04
What the machines are for
Four kinds of work, all of which want dedicated GPUs rather than a shared queue.
Model serving
Hosted inference for open-weight language and vision models. Dedicated GPUs, predictable throughput, and pricing that doesn't move when someone else's traffic spikes.
Agentic workloads
Compute for long-running agent systems — tool use, retrieval, multi-step reasoning — where jobs run for hours and per-token API pricing stops making sense.
Scientific computing
Training and inference for research groups: simulation, imaging, materials modeling, large-scale data analysis. For teams who need real GPU time without waiting on a university cluster allocation.
Research discovery
Our own work — automated annotation and training pipelines for computer vision, and tooling for searching large technical corpora. It runs on the same hardware we rent out.
Energy
The name started as a joke about the weather. It became the question the business runs on.
One eight-GPU node draws about 15 kilowatts, continuously — roughly a dozen households, without pause. That number, not the price of silicon, is what decides where this industry can physically be built.
It's why compute keeps landing in Washington, where about two-thirds of in-state generation is hydroelectric and has been since long before anyone needed a GPU.
We also run a small solar and storage pilot. A few kilowatts — nowhere near enough to power a node, but enough to learn what on-site generation and battery buffering actually take before we build somewhere that depends on them. We'd rather understand the constraint early than inherit someone else's assumptions about it.
Washington in-state electricity generation
Approximate. It's the reason the map of American AI infrastructure has a cluster along the Columbia.
Infrastructure
Small, owned, and run by the people who bought it
Our fleet is NVIDIA H200 NVL and RTX PRO 6000 class hardware, expanding to HGX B300. Everything is owned outright and operated by us, in Washington State, on commercial three-phase power — with colocation in the Seattle metro for production workloads.
Being small is the point. We can tell you which machine your job is on, what it costs us to run, and when it's due for replacement. If you need a specific configuration, ask — we size purchases around work we've actually been asked to do.
Tell us what you're trying to run
Send the workload and roughly how much of it. We'll tell you honestly whether we're the right fit.
contact@wheresun.com