Where's Sun
Bellevue · — Book a call

Your models. Your data. Your infrastructure.

Private AI infrastructure in the Pacific Northwest, with the engineering to evaluate, deploy and operate open-weight models in production. Your data stays on dedicated hardware we can name, in a facility we can take you to.

Dedicated tenancyOpen-weight modelsForecastable cost
Book a call See how it works
Sun over BellevueWhere is it? →
—Elevation now
—Sunrise
—Sunset
Single tenancy by defaultThe machine is yours for the term. No shared pools, no neighbours, no contention you didn't plan for.
No egress chargesMove weights, checkpoints and datasets in and out as often as you need. We don't meter the door.
No multi-year commitmentMonthly terms. Add capacity when adoption grows, release it when a project ends.

Why bring models in-house

Three reasons organizations stop sending data to a public API

Your data carries obligations

Customer records, contracts, clinical notes, source code. Once that traffic transits a third party, it becomes something you explain in every security questionnaire, every renewal, and every diligence process that follows.

Cost that grows with adoption

Consumption pricing is comfortable during a pilot and difficult at rollout. Dedicated capacity is a fixed line item your finance team can forecast twelve months out.

Continuity of the model itself

Versions are deprecated, behaviour shifts without notice, limits arrive at inconvenient moments. A model you host stays exactly as you validated it, for as long as you need it to.

What we do

From model evaluation through production operations

Most teams would rather not build an infrastructure function to find out whether this works. That part is our job.

01

Evaluate

We benchmark open-weight candidates against your task and your data, and give you a written comparison with the API you use today — quality, latency, and cost side by side.

02

Deploy

Serving architecture, batching, quantisation, context handling. The result is an endpoint your application already knows how to call, so the migration is a configuration change rather than a rewrite.

03

Adapt

Fine-tuning and adapters where they earn their place. Training data stays within your environment and never becomes part of another organization's model.

04

Integrate

Retrieval, agent orchestration, evaluation harnesses, batch processing. The work between a model and a product, built to your requirements rather than from a template.

05

Operate

Monitoring, patching, failover and capacity planning, with a named engineer who has physical access to the hardware your workload runs on.

06

Scale

Capacity added as adoption grows and released when it doesn't, without a multi-year commitment to hardware you have not yet needed.

Some organizations engage us for all six. Others want the infrastructure and their own team on top of it. Both are straightforward.

Models

Run the open models your team already evaluates

When a new release lands you can be testing it the same week, rather than waiting for a vendor to add it to a catalogue.

LlamaQwenDeepSeekMistralGemmaGLMgpt-ossKimi
FluxStable DiffusionWanWhisperembedding & rerankers

Language, vision, image, video and speech families, plus anything else with published weights. If you've already fine-tuned a model elsewhere, bring the adapters with you.

The difference

What changes when the model runs on your side of the boundary

Public API

Capability you rent

  • Every request transits a third party
  • Spend rises with each new user and use case
  • Model versions change on the vendor's schedule
  • Shared capacity, with limits you don't control
  • A recurring item in every security review

Where's Sun

Infrastructure you govern

  • Data remains on named, dedicated hardware
  • A fixed monthly cost, whatever the volume
  • The model stays exactly as you validated it
  • Capacity reserved for you and nobody else
  • Documented answers your reviewers can verify

The hardware

Dedicated machines, not a share of a pool

Single tenancy as standard. You can ask which machine your workload runs on, and we will tell you.

Available now

NVIDIA H200 NVL

141 GB HBM3e

Memory enough to serve a 70B-class model at full precision with long context. The usual starting point for replacing a frontier API.

Available now

RTX PRO 6000

96 GB · Blackwell

Image, video and audio generation, mid-size language models, and batch workloads where throughput per dollar is the deciding factor.

Arriving 2026

NVIDIA HGX B300

8 GPUs · 2.3 TB HBM3e

Eight Blackwell Ultra accelerators on a single fabric, for workloads that have outgrown one card. Capacity is worth reserving in advance.

If the configuration you need isn't listed, tell us the model and the timeline. We purchase around work our customers have actually committed to.

Story

We named the company after a weather joke. It turned out to be the whole business.

The domain was meant for a small tool that found the sunny corner of a grey Puget Sound afternoon. Then we started buying accelerators, and discovered that the question we'd named ourselves after was the one that actually governs this industry: where does the power come from.

A rack of GPUs is a machine for turning electricity into answers. Ours draw roughly fifteen kilowatts each, continuously — about a dozen households apiece. Washington generates close to two-thirds of its electricity from moving water, which is why this corner of the map filled with compute, and why the same workload here costs less and carries a lighter footprint than almost anywhere else in the country.

We run a small solar and storage installation of our own as well. It is nowhere near enough to power a node. It is enough to keep us honest about what generation and storage actually require.

Washington in-state electricity generation

Hydroelectric~66%
Nuclear~8%
Natural gas~14%
Wind & solar~12%

Approximate, and the reason the map of American AI infrastructure has a cluster along the Columbia.

Questions

What security and procurement ask first

Where does our data reside, and can we verify it?

On storage attached to dedicated hardware in a named facility in Washington State. Under NDA we will tell you exactly where, provide an architecture diagram, and put the data handling in writing so it can go straight into your own customer documentation.

Do you access our data, or train on it?

No. We do not copy it, we do not inspect it, and nothing from it enters any model. At the end of an engagement we wipe the storage and confirm in writing that we have.

How does the cost compare to what we pay now?

Below a certain volume a public API is genuinely the better answer, and we will say so. Above it, dedicated capacity at a fixed monthly rate usually wins by a considerable margin. Send us a recent invoice and we will work through the comparison with you.

Which models can we run?

Any open-weight family — Llama, Qwen, Mistral, DeepSeek, Gemma, and the image, video and speech models alongside them. When a new release lands, you can evaluate it the same week rather than waiting for a vendor to offer it.

We don't have an ML infrastructure team.

That is the usual position. Most of our customers have strong product engineers and nobody who wants to own GPU drivers and serving stacks. We take that layer; your team keeps building the product.

Can you support our security review?

Yes. NDA, architecture documentation, data flow description, and a direct conversation with whoever is completing the questionnaire. We would rather spend an hour on that than have it stall a deployment for a month.

What happens if we move on?

You leave with open weights, your own fine-tuned adapters, and a stack that runs anywhere. That is a materially better exit position than leaving a hyperscaler, and we would rather earn the following year than make leaving expensive.

Talk to us

Start with a conversation

Tell us what you're running and what it costs today. If moving it doesn't make sense yet, we'll tell you that instead.

contact@wheresun.com
Office 3003 Northup Way #100
Bellevue, WA 98004