Why VCs Must Build Their Own LLMs
Every company knows things the outside world doesn't.
In a way, this knowledge is their whole business. It is usually siloed in cluttered inboxes, human brains, and outdated procedures.
AI gives companies a new way to put this knowledge to work.
The competitive advantage of a VC comes from knowledge accumulated over years of investing. Put simply, the job is to use your insight to get into the best deals first.
At Arkane, we built an agentic AI system to write the first checks for Deep Tech founders.
Now, powered by NVIDIA, we are creating our own models to achieve both scale and velocity impossible to replicate via traditional sourcing channels.
From General AI to Sourcing AI
We played with our first agents in March 2023. In 2025, we cobbled together systems to gather and analyze scientific evidence, the IP landscape, and government interest. We digested millions of documents, parsed hundreds of databases, and used dozens of models to find the most lucrative investment areas. In 2026, we closed the first deals sourced by our system.
Our goal was to build a specialized system that could answer the key questions that matter to us:
- What important problems have suddenly become solvable?
- Who are the people working on these solutions?
- Who exactly should we meet right now?
Better models led to better answers, and it worked for a time. However, we ran into the limits of general-purpose AI1.
From understanding arcane neuroscience to using proprietary heuristics, our own model would have solved a lot of problems. The next step was to find the best way to build an integrated solution. NVIDIA stood out for the combination we needed: open-weight models, proprietary hardware, and tools to train and deploy custom models as an integrated stack.
Our First Results with NVIDIA
Data markup turns raw source material into information our system can use. It was a demanding first test: our existing process was already cheap and operated at scale.
Initial tests with Nemotron 3 Super produced:
- 11× lower costs than GPT-4.1
- 18× more data availability for founders’ search
Reducing data markup costs changed how much raw information we could afford to bring into our process. More available information gives us a broader base to investigate and compare investment areas. It also helps us identify more people working in these areas.
Why This Matters for a VC
Lasting flywheels are rare in business. Competition catches up, products get replaced, and market forces shift. VCs usually build their flywheels in capital access, marketing, founder support, and other non-technological means.
Internal systems that can learn from every data search, partner decision, and investment outcome are such lasting flywheels. Every day these powerful systems grow the gap between VCs that have them and those that don’t.
The real advantage comes from what belongs to you and you alone.
Building the next VC stack on NVIDIA
Our next step is to build a full AI stack on NVIDIA hardware, drawing on the full breadth of its model portfolio. We will fine-tune models like Nemotron 3 Super 120B A12B and Nemotron 3 Ultra 550B A55B on our accumulated research, proprietary heuristics, and subject-matter expert feedback.
That is the flywheel we are building at Arkane: a family of proprietary models sourcing the best Deep Tech founders before everyone else.
This is the future of how investment firms will operate. Some will build models for investment decisions; others will craft them for relationship management.
We believe that proprietary models will become table stakes for every serious VC firm.
With NVIDIA’s technology, we are building for that future now.