AI is making code easy to copy. The invention underneath isn't. We discover novel primitives in language models, file on them at the USPTO before public disclosure, and license the IP. That's the whole company.
For a decade, AI progress was measured in GPUs. That era is ending. The frontier is no longer who can compute the most — it's who can invent the primitives that make compute matter, and who files on them first. An invention-first, IP-native lab is the shape of a company built for the window that just opened. The labs that move now own the next decade.
Gains from bigger runs and more parameters are slowing. The next breakthroughs are primitive-level, not scale-level — and invention, not compute, is what finds them.
Frontier-quality base models are becoming commodity. The advantage moves to what you can invent to run inside them — and who owns the rights to those mechanisms.
Every downstream AI company needs defensible, differentiated mechanisms. Few have any. IP-first invention labs are the natural supply side of that market.
Once a primitive is publicly disclosed, it is no longer patentable. Labs that file before the field rediscovers a mechanism own it for the next decade.
Named after the black hole in Interstellar — a place at the edge of what's known, past which the rules change. That's the work: invent where the published literature hasn't looked yet, then patent and license what we find.
We don't host inference. We don't sell per-seat subscriptions. We don't staff your team. We invent novel capabilities in language models, file patents on what we find, and license the IP to the companies that need it.
Founded and based in the United States. Currently fundraising. Invention first, IP second, commerce third. The base model stays neutral — our licensees own their deployments end-to-end.
Agentic coding is collapsing the cost of building software. What stays scarce is the underlying invention — and the patents that surround it. We file broadly on purpose. In a world where anyone can build anything, defensibility moves from code to IP.
1// conjecture2hypothesis: "models can be written3 to, not just trained"45// validate across families6for model in [qwen, llama, mistral,7 phi, gemma]:8 prove(model, production_scale=True)910status: replicated
Each invention starts with a conjecture the published literature treats as expensive, impossible, or irrelevant. We test it, patent it, and license it.
A method for writing information directly into the weights of a frozen pretrained language model — in milliseconds, using only forward passes. The output is a portable Knowledge Module: loadable, stackable, swappable, and revocable at will. It replaces fine-tuning, RAG, and prompt engineering with a single persistent write operation on the model itself.
A transformer architecture trained from scratch to be populated by Knowledge Modules at runtime. Reasoning layers handle language, logic, math, and code. Knowledge-bank layers start empty and take their content from whichever KMs are loaded. A small reasoning engine plus a strong capability module approaches frontier performance at a fraction of the serving cost.
Next primitives are in private research. Ask us about the pipeline under NDA.
Most AI research optimizes a single axis: bigger models, longer training, more expensive inference. Valuable work — but crowded. We work on new operations on language models that change what a deployed model can do in milliseconds, without retraining, retrieval, or context-window bloat.
One patent is not a company. A portfolio that compounds across primitives, model families, and verticals is. Four reasons this goes venture-scale.
The cadence is designed to compound. A licensed primitive generates revenue. That revenue funds the team building the next one. Every new invention is a separable asset the prior work did not have.
A fine-tune is obsolete when the next base model ships. A granted patent runs for roughly twenty years. We accumulate long-dated, cross-model assets in a field that churns quarterly.
Enterprise and regulated teams cannot ship a black-box SaaS. They need capability they can run in their own stack, with their own data, under their own compliance. That is what licensing gives them — and what a hosted product cannot.
Every company running a language model is a potential licensee. We do not sell infrastructure or hosting. The foundation-model ecosystem — open-source and frontier — does the scaling for us.
Gargantua Labs invents novel AI primitives, files patents on them, and licenses the IP. That structure is deliberate — and it rules out the four positions most AI companies occupy.
"A world where a language model is an operating system and knowledge is a file."
The end state
Our vision, Gargantua Labs
Where we're headed
Invention first · IP second · Commerce third
Investors. Licensees. Research partners. We're sequencing a short list of each this quarter — every engagement starts the same way, a private conversation under NDA.
We're raising now.
Your model is the bottleneck. Your knowledge is the moat.
Academic collaborations and strategic co-development.
Capital raise is live. Licensing slate is opening. We're prioritizing a small number of flagship investors and licensees this quarter — early conversations are how we set the order. Whichever side you're on, the first step is a private one.
Investor intros and licensing inquiries welcome. NDA standard.