Data moats and flywheels

LAST REVIEWED 2026-10SOURCED FROM 3 SESSIONS, APR–OCT 2026

TL;DR

• Proprietary data can compound into an advantage. Algorithms and compute are available to others too.
• That data is what nobody publishes, or years of how customers use the product.
• Reward the user on the spot. That is what produces labeled data.
• Small data still signals. Learning is log-linear.
• Show why the product improves with use and why customers keep choosing it.
• For a marketplace, explain what value remains after buyer and supplier meet.

Why data is the moat

A deep tech GP’s framing: startups can’t out-algorithm the labs or out-compute the hyperscalers — data is the one ingredient a small company can genuinely own. And it doesn’t have to be exotic. The example told in the room: a high-schooler who scraped a retail-trading community’s ticker chatter for three years and sold it to hedge funds — unassailable, because history can’t be re-collected.

Two repeatable sources of differentiated data, from a seed investor whose thesis is exactly this:

1. Collect what nobody publishes. Do the unglamorous work of gathering data that exists in the world but not in any dataset.
2. Decades of domain fluency. Knowing how customers actually use data — what’s signal and what’s noise in a niche — is itself proprietary.

The corollary from the roundtable: “alpha is everything; ordinary data gives no signal.” Commodity enrichment sources are table stakes, not a moat.

Building the flywheel

The mechanics that came out of the practitioner roundtable:

• Make feedback pay instantly. Users label your data when doing so immediately improves their own experience — that instant reward, not goodwill, is what powers explicit feedback loops.
• Interview your ICP about what’s missing. The people doing delivery and support know which context the dataset lacks; mine call recordings the way teams “study film.”
• Back-test creative sources. Permit and remodel records against subsequent outcomes, demographic age-out data, web-scraped alternative data, satellite imagery — the roundtable’s real-estate examples all shared one property: a testable predictive claim.
• Small data is enough to start. Learning is roughly log-linear: a small dataset surfaces the strongest signals and directional insight; granular personalization is what needs scale. Don’t wait for big data to ship the first loop.

Contracts feed the flywheel

The first-customer playbook connects here: deep early discounts traded for utilization minimums and feedback obligations are how the flywheel gets its first turns — see First customer contracts.

For compliance-grade products, the roundtable’s reliability pattern: multiple models cross-checking with a tiebreaker, tasks decomposed into micro-steps with intermediate outputs as an audit trail, and human-in-the-loop encoded as a company value rather than a patch.

Pitching the moat

Investors don’t want the nuance. The roundtable’s framing advice: “we’re building the largest dataset in X; it trains the best predictive models; customers stay because the product keeps getting better” — a network effect in one breath. If you keep having to explain why the data compounds, you’re pitching the wrong investors.

“

It’s much better than… trap them, then please them. Right? Trap them in the correct way.

Managing partner, deep tech seed fundsession, Apr 2026, on products users can’t leave

When the moat depends on a marketplace or physical operations

The October pitch review brought a marketplace question into the discussion: after you introduce a buyer and supplier, why do they keep transacting through you? A directory or introduction can lose its value once both sides know each other. The participating founder described services such as financing and insurance as potential reasons to stay. Those services were part of the proposed approach; the session did not establish that they had prevented buyers and suppliers from going direct.

Make the continuing value explicit, then test whether customers actually use it. The reviewer also asked whether an infrastructure business would remain needed after the market bottleneck that made it timely began to ease. A large problem today is not, by itself, proof of a durable business.

The discussion broadened beyond data to physical and operational advantages. For a business involving hardware, explain who owns the assets, who pays for them, and what advantage the company retains. The investor described financing structures in which other parties carry hardware costs; this session does not establish that such financing is available to every startup or that software companies need hardware to be defensible.

Editorial synthesis: a useful test is whether the advantage survives when competitors can access similar software, buyers know their suppliers, and the immediate bottleneck improves. That question combines the reviewer’s challenges; it was not a quoted checklist from the room.

Sources

Three sessions: an AMA with the managing partner of a deep tech seed fund on defensibility (Apr 2026), an in-person member roundtable on proprietary data and flywheels (May 2026), and an October 2026 pitch-deck review with a venture investor and former founder on marketplace durability and physical or operational advantages. Roundtable material is presented without individual attribution. October material is paraphrased, with founder proposals distinguished from demonstrated outcomes.