Seed to Series B · New York
Gothic Holdings backs companies that were built around models from the first commit — not retrofitted with them. We underwrite decades rather than demos, and we run a research lab of our own, so the people judging your architecture have shipped one.
$180M
Under management
Across two vehicles — Fund I, vintage 2022, and Fund II, opened 2025.
23
Companies backed
Seed through Series B, concentrated across four sectors.
78%
Follow-on rate
Share of the portfolio we have re-invested in at least once.
What we look for
Most firms describe what they invest in. We would rather tell you what we check, because that is what you will actually experience if we meet.
The model is not a feature bolted onto a product; it is the reason the product can exist. We look for teams whose architecture, data pipeline and pricing all assume inference is a primary cost line. Retrofits rarely survive the second model generation.
Architecture review at first meetingSoftware that is ten percent better rarely dislodges an incumbent. Software that removes most of a cost line does. We underwrite the unit economics of the market you are entering — not the demo you are showing — and we model what happens when inference gets an order of magnitude cheaper.
Model-economics underwritingTaste is the ability to tell which results will hold. We look for founders who read the literature, run their own evaluations, and can say precisely why the obvious approach fails. Credentials are optional. Judgment is not.
Technical review by lab researchersPortfolio snapshot
We hold positions in a small number of places we understand well. Companies are described by what they do rather than what they are called — most of the portfolio is not public about its investors, and we prefer it that way.
8 companies
6 companies
5 companies
4 companies
The Lab
Gothic runs a small research group in-house — four researchers working on efficient inference, domain-specific small models, and evaluation tooling. It exists for one reason: an investor who has trained a model can tell the difference between a hard problem and an expensive one.
Findings move in both directions. Lab work sharpens our diligence; portfolio companies get the benchmarks, the ablations and the engineers who ran them. Where a result has value beyond the portfolio, we license it selectively rather than let it sit in a drawer.
Serving cost — USD per 1M tokens
Gothic serving stack Commodity baseline
Current sweep
Recent runs
| Run | Params | Eval | Δ | State |
|---|---|---|---|---|
| gh-sm-legal · r42 | 1.4B | 0.913 | +2.1 | promoted |
| gh-sm-legal · r41 | 1.4B | 0.892 | +0.4 | archived |
| gh-router · v9 | 340M | 0.874 | +3.8 | promoted |
| gh-sm-claims · r17 | 3.1B | 0.861 | — | running |
They asked to see our eval harness before they asked to see our deck.
Introductions
No warm introduction required. We reviewed 2,340 companies last year and met 190 of them; roughly half arrived cold. Tell us what you are building, what the model does that ordinary software could not, and what the capital is for.
If it is a fit, you will hear from a partner and a lab researcher within one business day. If it is not, you will get a real answer and the reason behind it. Our standard is a decision inside fourteen days of the first meeting.
Five fields. Read by a partner, not a filter.
Thanks — we'll be in touch within one business day.