Gothic is a fund and a laboratory in the same building. The fund writes the check; the lab does the work that tells us whether writing it is a good idea.
What follows is how it actually runs — the check sizes, the diligence, the fourteen-day standard, and what you can call on once we are shareholders.
Engine one
$180M across two vehicles, deployed into companies from seed to Series B. Concentrated by design: five to seven new positions a year, and more than half of each fund reserved to keep backing them.
Engine two
Four researchers training small models and building evaluation tooling. Its output is not a product line; it is why our diligence is technical rather than narrative, and why portfolio teams call us at eleven at night.
The fund
Fundraising is expensive for founders mostly because nobody says what they actually do. Here is ours, in the order it usually comes up.
We take a board seat at Series A and above, and an observer seat at seed. Nine board seats and eleven observer seats across the portfolio today. Between meetings we answer quickly and interfere rarely; the operating decisions are yours, and the reason we did the technical diligence was so we would not have to re-litigate them.
We prefer priced rounds with standard documents and no structure that would embarrass anyone in a downside case. No participating preferred, no ratchets, no pay-to-play provisions written into a first round. Investment is concentrated in the United States and Canada, with a small European allocation.
How we decide
Our diligence is technical because our lab makes it possible. The schedule below is the one we hold ourselves to; if a step will run long, you hear it on day one with a date attached.
Day 0
Forty-five minutes with an investor and, wherever we can arrange it, a lab researcher. No deck required. We would rather see the system than the story, and we will ask about the failure modes before the roadmap.
Partner & researcher
Days 1–4
A lab researcher reads the architecture, the evaluation setup and the data pipeline, and runs your benchmark on our own hardware where you allow it. Most founders tell us this is the first time an investor has done this rather than described it.
The Lab
Days 5–8
We build a cost model of your inference and support load against your pricing, then stress it two ways: against inference an order of magnitude cheaper, and against a competitor with a two-year head start and worse taste.
Investment team
Days 9–12
Six to ten conversations, weighted toward buyers rather than boosters. You see the reference list and the questions before we make the calls, and you get the substance of what we heard whichever way the decision goes.
Investment team
Days 13–14
A yes with terms, or a no with the memo behind it. We do not run silent processes, we do not sit on companies to keep an option open, and we do not use a slow no to negotiate.
Partnership
The Lab
Three research areas, chosen because each one changes what a company in our portfolio can afford to build. The lab publishes internally, ships tooling, and is measured on whether founders use it.
Quantization, speculative decoding and scheduling work aimed at a single number: cost per useful completion. Our internal serving stack now runs at $0.62 per million tokens against a commodity baseline of $2.90 — and against our own $4.10 eighteen months ago.
We train sub-2B models on narrow, well-specified tasks — clause extraction, claims triage, defect classification — to find where small genuinely beats large. Across fifteen task benchmarks we have built so far, it does in eleven.
Harnesses that catch what a leaderboard will not: distribution drift, benchmark contamination, and silent regressions after a base-model upgrade. This is the first thing we hand a company after the wire clears.
Findings flow outward. Every portfolio company receives the benchmark suites, the ablation results and an engineer to walk through them — not a link to a repository and good luck.
Findings also flow inward. No investment reaches the partnership without a written technical memo from the lab, and we have declined companies on the strength of a two-page ablation more than once. It is an uncomfortable document to write about a founder you like. It is the reason the portfolio looks the way it does.
Where a result has value beyond the portfolio — the serving stack, most recently — we license it selectively rather than spin it out. The lab is a research capability, not a second business, and we intend to keep it that way.
What you get
Every fund promises support. These are the six things a Gothic company can actually book, with last year's numbers attached to each one.
1,100 subsidized GPU-hours per company each year through our compute partners, plus help negotiating your own capacity before the month you desperately need it.
1,100GPU-hours per company, per year
A standing hour with the lab every week. Bring an evaluation behaving strangely, a training run that will not converge, or an architecture you are about to commit to for three years.
WeeklyOpen to every portfolio company
An independent read of your model, your evaluation setup and your inference bill — annually, or ahead of a raise when a technical diligence process is coming for you anyway.
41Audits conducted in 2025
Research and infrastructure candidates we have interviewed ourselves, drawn from the lab's network rather than a database. Fourteen portfolio hires placed last year.
14Placements in 2025
Introductions to the growth funds that actually price rounds in this sector, with a data room already in a shape they recognize and a technical memo they can verify.
11Portfolio rounds supported in 2025
On valuation, on whether the round is real, on whether the hire is a mistake. We would rather be the investor a founder calls before a decision than the one who reviews it afterward.
Same dayResponse standard
Portfolio principles
Five to seven new positions a year; twenty-three in total since 2022. If we cannot say in three sentences why a company wins, we do not have a position — we have a lottery ticket. Concentration is what makes the diligence worth doing.
55% of each fund is reserved. We follow on when the thesis is intact and the price is defensible, and we tell founders where we stand early rather than leaving them to read silence. Seventy-eight percent of the portfolio has seen a second check.
We underwrite decades, not demos. Our model assumes a ten-year hold and a secondary market we do not depend on. That assumption is why we can be relaxed about a flat year and unrelaxed about an architecture that will not survive one.
For founders
Write two paragraphs to [email protected]: what the model does, and what could not be built without it. A deck is optional. A working system, or a precise account of why one does not exist yet, is not.
Do not spend a week on the note. We are reading for the shape of the problem, not the prose.
Fourteen days from the first meeting to a yes with terms or a no with reasoning. If we are going to be slower — usually because a technical review needs a second researcher — you hear that on day one, with a date.
Only what you choose to share, under a mutual NDA signed before any technical review begins. Researchers who read your code are walled off from portfolio work in an adjacent space for the duration, and we name the adjacent companies before you decide whether to proceed.
If you would rather run the benchmark yourself and send us the artifacts, that is a normal request and we will not read anything into it.
We do not invest in a direct competitor of a portfolio company. There is no version of that sentence with an exception in it.
Adjacency is a judgment call, and we make it with the existing founder in the room rather than after the fact. If we are already looking at something close to what you are building, you will hear it in the first meeting.
Often. We led nine of the last fourteen rounds we participated in and are comfortable setting terms, taking the board seat and doing the unglamorous work that comes with a first institutional check. We are also happy to follow a lead we respect and to say so quickly.