Startups
Your startup’s AI.
Own the model your product runs on. Your token bill becomes a cost you can plan.
Where it runs
- GPUDedicated, yours alone
- RegionYou choose
- WeightsYours to keep
- A dedicated GPU a month, proposed
- $7,700
- Sonnet 5.5 tokens a month, at list prices, cost about as much as one GPU
- 1.9B
- Weeks from scope to first release, indicative
- 6
- To roll back behind your API, by design
- 60 s
What we hear
“Our token bill grows faster than our revenue, and the model is not ours.”
Rented plans moved to usage meters in 2026. Every customer you win adds to the bill.
- Billed by useEach customer adds tokens.
- Cheaper tokens, the same billPrices halved in 2025; use grew 4.5 times.
- Someone else’s roadmapIt can reprice, change or retire your model.
Bain, How token economics will change opex, 10 June 2026.
Frontier and owned
The learning loop exists. Make it train your model.
Keep renting
- Broad and varied requests
- Features you are still testing
Own the model
- The heavy, repeated work your product runs on
- Your traffic, and what your team approves
- A model nobody can reprice or retire
Cursor’s Tab model: a 28% higher accept rate with 21% fewer suggestions (vendor figure, 2025). Fin’s Apex 1.0, post-trained on its own service work: 73.1% resolved against 71.1% for GPT-5.4 (vendor figure, via VentureBeat, 26 March 2026).
The Formula, for you
Your work makes an open model yours.
- Base model
- An open model, for example Qwen, Llama, Mistral or Kimi.
- Your traffic
- Real requests, and the answers your team approves.
- Your startup’s model
- Trained by us, owned by you, with a Model Deed.
The model is yours, with a signed Deed.
The FormulaOpen bases are examples; every licence is checked before use.
How your people use it
From your real traffic to your own model.
- 01CaptureReal requests from your product.
- 02ApproveYour team picks what it learns.
- 03DistilA large teacher trains a small model, where the licence allows it.
- 04GateIt must beat your current model.
- 05ReleaseLive behind your API. Back in 60 seconds.
Release 5
Learning from approved traffic
LearningRelease 4
Passed the gate
ReleasedRelease 3
A tie: current kept
HeldRelease 2
Rolled back in 60 s
Rolled back
What you will ask
Five objections, answered plainly.
- 01
Data
Where does our traffic go?
On Eigna Cloud, onto your own volumes in our GPU region, with keys you hold. It never moves between clients.
- 02
Team
We have no one to run models.
You do not need anyone. We rent and run the GPU, serve your model behind your API, train it on approved traffic, test it and release it.
- 03
Quality and risk
Will our own model match the frontier on our task?
Only a test on your task can say. The Test Lab measures it against your current model first. When a better open model ships, we re-base on it, and it goes live only if it wins.
- 04
Cost
Is owning really cheaper?
At heavy, steady use. Proposed: $7,700 a dedicated GPU a month, about what 1.9 billion Sonnet 5.5 tokens cost at list prices. For light use or the cheapest tiers, renting wins.
- 05
Investors and customers
What do we own at the end?
The weights, your sealed test set, the release record and a Model Deed signed by both. You can leave with all of it, at any time.
Where it runs, for you
Eigna Cloud: a dedicated GPU, no infrastructure team.
A GPU we rent and run for you, serving your model behind your API. No token limit: the GPU is the limit.
Eigna Cloud
Data stays in the region you choose.
Your data sits on your own volumes in our GPU region, with keys you hold.
Where owning pays, and where it does not.
| Compared | Blended price per million tokens | Tokens a month for the cost of one Eigna Cloud GPU |
|---|---|---|
| Claude Haiku 5.5 | $0.20 | About 38 billion |
| Claude Sonnet 5.5 | $4.00 | About 1.9 billion |
| Claude Opus 5.5 | $8.00 | About 960 million |
| GPT-5.5 | $11.25 | About 680 million |
Proof
Published, losses included.
17.8%
Our first training run cut forecast error 17.8% against its base model. It is not yet ahead of a simple last value baseline, and we publish that too.
Our first client: a forecasting model for trading. It shares founders with Eigna, said up front. Then a media agency, for image and video, and a web agency, for code.
How it grows
How we work, weeks 1 to 6.
- Week 1ScopeOne task, a traffic sample, today’s bill. We agree the sealed test.
- Weeks 2 to 4Test LabA first candidate against your baseline. You keep the report.
- Weeks 5 to 6First releaseBehind your API: shadow first, then live.
- ThenKeeps learningReleased only when better.
Timings are indicative until the first projects run.
Two ways in
Pay a fee, or part in equity.
As a client
- Eigna Cloud, $7,700 a GPU a month, proposed
- No equity
- The weights, the test set and the Deed are yours
With Eigna Ventures
- Engineers who build with your team
- Compute to train and serve
- Part paid in equity, on terms set with counsel
No promise of returns. Client data never moves between ventures or clients.
Start
Send us one task and your token bill.
One task and the last three months of your bill. We answer with a fixed-cost plan and a Test Lab proposal. You decide.