Run frontier open-weight models, Cut the API bill by up to 91%.
GLM 5.3, Qwen 3.8, DeepSeek and Mistral, served from dedicated GPUs inside your own network behind an OpenAI-compatible API. Today 1,000 users cost you around $3,000 each per month in API tokens. On your own GPUs the same headcount costs $272.
What 1,000 users cost, two ways.
One side is a price list you rent by the token, at $3,000 per user per month. The other is capacity you own.
| Line | Per year |
|---|---|
| API tokens today1,000 users at $3,000 per user per month | $36,000,000 |
| TUYO Private AI DeploymentUnlimited bandwidth and usage, with scalable infrastructure for peak hours | $3,267,584 |
| You keep, measured against what you pay today | 11.0× $32,732,416 |
$3,000 a month is what 1,300 million tokens cost at ChatGPT Sol 6.1 list rates with cache. The same tokens on Claude Opus 5.5 high would be $6,000 a month: $72,000,000 a year for these 1,000 users.
The price of a token is the whole business model. Open weights don't have one.
Your statement, in four entries.
Answer four questions. The one-pager comes back sized to your headcount: what your AI costs today, what it costs with us, and what you keep.
Statement of inference costs
Entry 1 of 4Statement of inference costs
Prepared for your company
- Users
- —
- Dedicated GPUs
- —
- VRAM
- —
- API tokens today
- —
- With TUYO
- —
Sized as you answer · 188 GB per GPU · and no model names, just capacity
—
What the contract includes.
- Dedicated capacity. GPUs reserved for you rather than shared tenancy, so nobody else's traffic sets your queue.
- Frontier open-weight models, and new ones as they ship. You choose which model serves which workload.
- OpenAI-compatible API. Switching is a base URL and a key: the tools your engineers already use keep working.
- Runs in your country, inside your private network. Your code never crosses a border to reach a provider's datacentre.
- Zero data retention. Prompts and responses are not stored, logged for training or read by anyone.
- 99.5% uptime SLA, 1-hour response for critical incidents, chat support in business hours, security monitoring.
- Setup and onboarding included. Infrastructure, first model, and your engineers' first session.
- No per-token metering. Heavy use does not move the bill, which is the entire point of owning capacity.
Questions we get before the first call.
Is it really that much cheaper?
It is arithmetic, not a claim. A rented token carries a provider's margin; a GPU-hour does not. At 100 users the deployment costs 4.5× less than the same headcount in API tokens, 10.6× at 500 and 11× from 1,000 up. Even the smallest reservation, six GPUs, pays for itself more than four times over. The statement, with your own headcount, comes out of the form above.
Are these models as good as GPT or Claude?
They're frontier open-weight models, not the same weights. For coding agents, chat and retrieval, GLM 5.3, Qwen 3.8, DeepSeek and Mistral are close enough that most teams make them the default, and keep a proprietary model on the hardest 5% of traffic, which still costs a fraction of running everything there.
How long does setup take?
Days, not quarters. The capacity is reserved, the API is a drop-in replacement for the one you already call, and your engineers are working against the first model inside the first week.
What happens as we grow?
One more GPU per 36 users. Adding people doesn't mean renegotiating a contract or re-pricing your tokens: the reservation grows in a straight line.
Where does our data live?
In your country, inside your private network. Nothing is retained after the response is returned and nothing leaves the deployment.
Where does $3,000 per user a month come from?
It is what daily agent use costs on a rented API: about 1,300 million tokens a month at ChatGPT Sol 6.1 list rates with cache, and $6,000 a month for the same tokens on Claude Opus 5.5 high. The statement uses whatever your own headcount implies, and if your spend is lower, the smaller reservation still saves you money, just less of it.
What's the catch?
Two honest ones. Below 100 users you still reserve six GPUs, so the saving is 4.5× rather than 11×: the floor is the floor, and we would rather tell you than let you find out. And you are buying reserved capacity for a year, not a subscription you can cancel in a bad month.