The Deep-Tech VC Research Stack in 2026

Ask a GP at a leading deep-tech fund how they source deals and they'll describe a patchwork. ArXiv alerts forwarded to Slack. A patent monitoring service that emails a weekly digest nobody reads. A CRM that technically tracks lab contacts but hasn't been updated in three months. A data room full of one-page memos someone wrote by hand.

This is the current state of the VC research stack for most funds: five or six disconnected tools, none of which talk to each other, all of which require manual effort to maintain. The data is scattered. The signal gets lost. And the deals that required moving fast — the ones where getting to a founder six months before their raise defined whether you were on the cap table — got missed because no one saw the signal in time.

The funds pulling ahead in 2026 are not doing more manual work. They've rebuilt the research stack around a unified, automated workflow. Here is what that looks like layer by layer — and where the point-tool approach breaks down.

Layer 1: ArXiv and Research Database Monitoring

The research layer is where deep-tech deals form. A computational biology startup begins as a Nature Methods paper. A fusion energy company traces back to a plasma physics preprint on arXiv. The formation signal exists long before any investor deck does.

The manual approach is an arXiv RSS feed filtered by keyword. It catches papers that match your exact search terms and misses everything at the intersection of domains — which is often where the most interesting work lives. Photonics meeting machine learning. Synthetic biology meeting materials science. No keyword filter catches what you don't know to search for.

The automated approach ingests the full feed across arXiv, bioRxiv, medRxiv, and IEEE Xplore — roughly 10,000 new documents per day — and applies a scoring model. Papers are evaluated for technical novelty, application domain clarity, author commercialization history, and institutional track record. The top 0.5% surface for human review. An analyst reads 50 high-signal papers instead of skimming 500 mediocre ones.

Layer 2: Patent Intelligence

A research paper signals capability. A patent signals intent.

When a university lab files a narrow, application-specific patent — rather than a broad academic claim — it typically means the technology transfer office is involved and a spin-out is being organized. The filing is often the first hard evidence that a company is coming, months before any fundraising begins.

The manual approach: a weekly USPTO digest, filtered by class code. The problem is that the most interesting filings are cross-class — a biotech patent that also touches semiconductor manufacturing, for example — and class-code filtering misses them entirely.

The automated approach correlates patent filings against the research database. When a PI whose papers scored highly in the last 18 months files a commercially-structured patent, that correlation is flagged. The system doesn't just track patents in isolation — it understands whether the patent represents a progression of a known research trajectory.

Layer 3: Lab Network Mapping

The most durable sourcing edge in deep tech is not data — it's network. The funds that see deals first are often the ones with a PI relationship built over 18 months, not the ones who found the company on AngelList the week the round launched.

Network mapping in the manual model is a spreadsheet of professor contacts, organized by whatever system the most organized analyst on the team invented. It doesn't capture co-authorship graphs, grant award histories, or which PIs have spun out companies before — all of which are predictive signals for who is likely to form a company next.

The automated approach builds a live graph of research relationships: which PIs collaborate, who has DARPA or NIH funding (a strong commercialization predictor), who has prior spin-outs in their history. When a new paper from a known node in your network appears, you get a warm-contact alert, not a cold discovery.

Layer 4: AI Scoring and Prioritization

The scoring question every deep-tech investor needs answered is not "is this research interesting?" It's "is this research investable — and when?"

Those are different questions. A paper on quantum error correction may be scientifically brilliant and commercially irrelevant for 15 years. A paper on a new battery electrolyte formulation might be six months from a seed raise. Telling the difference requires evaluating market size, IP defensibility, timeline to commercialization, and team strength simultaneously.

Manual scoring means an associate writes a one-paragraph take on each interesting paper, which a partner skims once in a weekly meeting. There is no consistency, no calibration across analysts, and no way to track whether the scoring criteria shift over time.

AI scoring applies a consistent rubric across every paper in the pipeline: market size estimate, IP novelty relative to prior art, institutional track record, estimated time-to-company based on research maturity. The output is a numerical score with a breakdown the analyst can interrogate — not a black box, but an accelerant for human judgment.

Layer 5: Deal Memo Generation

The last bottleneck in the research workflow is the memo. For every paper that crosses the investment-interest threshold, an analyst needs to produce a structured writeup: market context, comparable investments, commercialization path, suggested next steps. In a manual workflow, that takes 3–4 hours per paper. Most papers never get a memo because the analyst queue is always full.

Automated memo generation produces a first-draft one-pager for every scored paper — market sizing, comparable bets, open questions, suggested outreach timing — in seconds. The analyst edits and extends, not starts from scratch. The memo backlog goes from weeks to hours.

Where Point Tools Break Down

Each of these five layers has existing point-tool solutions. The problem isn't that the tools don't exist. It's that they don't connect.

PitchBook is the institutional standard for market data and deal tracking. It's the tool of record once a company exists and has raised money. It is not a research-layer tool — it doesn't monitor arXiv, doesn't track PIs, doesn't score papers for commercial viability. For deep-tech sourcing, PitchBook tells you what happened. It doesn't tell you what is about to happen.

Affinity solves a different problem — relationship intelligence and CRM for deal flow. It tracks who you know and how warm the relationship is. It is not connected to the research layer either. A warm contact alert in Affinity only fires when you've already met the person. It can't tell you that a PI in your extended network just published a breakthrough paper and is six months from a spin-out.

Capability	PitchBook	Affinity	Conduit
ArXiv / research paper monitoring	No	No	Yes
Patent filing intelligence	Limited	No	Yes
PI / lab network mapping	No	Contacts only	Yes
AI commercial viability scoring	No	No	Yes
Automated deal memo generation	No	No	Yes
Post-formation company tracking	Yes	Partial	Yes
Relationship / CRM management	Basic	Yes	Yes

The pattern is consistent: PitchBook and Affinity are strong post-formation. Conduit operates upstream — at the research, patent, and lab layer where the best deep-tech deals are actually won.

This isn't a critique of either product. PitchBook is doing exactly what it was designed to do: track the market of formed companies. Affinity does relationship intelligence well. The problem is that a deep-tech fund running both still has a gap — the pre-formation sourcing layer — that neither tool fills. That gap is where the best deals live.

What the Modern Stack Actually Looks Like

The funds consolidating onto a unified research platform are not abandoning PitchBook or Affinity. They're adding a pre-formation intelligence layer that feeds into those tools when a company actually forms. The workflow becomes:

The Compounding Returns of an Automated Stack

The immediate benefit of an automated research stack is obvious: more coverage, less analyst time wasted on triage, faster memo production. But the more important benefit is the one that shows up 18 months later.

A fund that started automated research monitoring in early 2025 now has a database of every high-scoring paper from the last 18 months, tagged by field, author, institution, and patent activity. When a company comes to market today, they can run a lookup: did we score this PI's research? Did we flag the patent? Did we track the co-authors? In most cases, the answer is yes — and the partner walking into that first meeting has context that no newcomer to the deal can replicate.

That information asymmetry is the moat. It doesn't come from any single paper or any single deal. It compounds from 18 months of systematic research coverage that most funds haven't built yet.

Read how the upstream sourcing motion works in our earlier piece: How VCs Are Using AI to Find Deep-Tech Deals 18 Months Before Formation.