Sancus Ventures backs the people behind the plumbing

A database can be correct and still tell you yesterday’s story. At Substack, according to data-replication company Artie, information once reached the warehouse in batches, sometimes hours apart or overnight. Analysts waited. Experiments waited. The software was doing its job, but the business needed a different job done. Artie says its streaming approach reduced that lag to 10-15 seconds.

That small interval offers a useful way into Sancus Ventures. The San Francisco investor backs companies working on software’s underlying machinery. Its portfolio includes Artie, Anyscale and Predibase. These businesses address different problems, but share a familiar predicament: an ambitious application can only move as fast as the systems underneath it.

A customer hiding inside the network

Sancus describes a community of more than 500 infrastructure experts and decision-makers, including engineering leaders at large public companies and growing technology businesses. For a founder, the attraction is specific. Somewhere in that group may be a person who understands the product, controls a relevant budget, or knows why a seemingly sensible deployment will get stuck.

In her April 2025 AngelList interview, founder Lake Dai explained that potential customers participate in diligence. The same conversation can test an investment and introduce a founder to a buyer. She also described seeking patient, strategically useful limited partners, and a Fund II approach emphasizing ownership and valuation discipline.

The appealing mechanism is the shorter distance between technical judgment and commercial feedback. Consider the questions an engineering buyer might ask: Who maintains this? What happens during an outage? How difficult is migration? A founder who hears those questions before finishing the roadmap has something more useful than encouragement. The product now has an informed critic.

The operator who became the investor

Dai’s experience helps explain the preference for practical questions. Carnegie Mellon identifies her as Alibaba employee #84, leading its early product team. Her subsequent career included Yahoo! and Overture. She later moved into investing and teaching, bringing the experience of making products work inside operating businesses.

Sancus launched in 2022. Its team page names Dai as founder and managing partner, Angela Hsu as vice president, and advisors whose experience spans infrastructure, developer platforms, crypto and law. Ted Nyman, for example, previously served as GitHub’s CTO. The mix reflects a business in which a promising technical idea still needs organizational, recruiting and commercial judgment.

Dai has taught Applied AI at Carnegie Mellon since 2016, according to Sancus. Teaching adds another vantage point: explaining how a model becomes useful requires attention to the application, the available data and the consequences of getting things wrong. Those are also reasonable subjects for an investor meeting.

When the plumbing stops the product

Return to Artie. Its founders describe a problem that extended beyond making an initial data pipeline work. Maintaining consistency, handling changing schemas and supporting different data types made the production version harder. The lesson is familiar to infrastructure builders: a demonstration can conceal the maintenance burden that determines whether anyone will rely on it.

Predibase approached another production problem: customizing open-source models and serving them efficiently. Rubrik announced an agreement to acquire the company in June 2025 and completed the acquisition in July. Its subsequent financial disclosure reported acquisition-date purchase consideration of $109.1 million. That figure describes Rubrik’s transaction; it does not reveal Sancus’s proceeds.

The interesting customer is the one who can explain why the demo will struggle on Monday morning.

An editorial reading of the Sancus approach

Monterey AI supplies a different example. Reforge announced its acquisition in January 2025 as it introduced Insight Analytics. The problem there was fragmented customer feedback: information scattered across tools and teams, difficult to turn into product decisions. Infrastructure, in this broader sense, includes the systems that help an organization understand what to build.

Follow the bottleneck

By July 2026, Dai’s public writing emphasized the software surrounding AI agents: permissions, evaluation, recovery, observability and cost control. Her argument was that stronger models do not automatically produce dependable systems. This is an evolution of the infrastructure thesis toward the work required to operate AI after the prototype succeeds.

The portfolio provides two timely illustrations. In July, Anyscale signed a definitive agreement to join Nscale, connecting distributed-computing software with physical compute infrastructure. The announcement remained subject to closing conditions. In August, Silicon Data announced a $30.5 million initial Series A closing with Sancus participating. Its business measures compute pricing and performance, helping buyers understand a market where identical chips can produce different results.

These developments show strategic demand for infrastructure capabilities. Reading them as proof of a particular fund return would go further than the evidence allows. For founders, the more useful observation is where the demand appears: around expensive, difficult work that customers need someone to handle reliably.

The useful question to copy

Sancus is an investment firm. Founders seek capital and support; limited partners supply the investment capital. Its software products belong to portfolio companies. That distinction matters when deciding how to engage: a founder is assessing an investor relationship, while a developer choosing a tool needs to evaluate the relevant product directly.

Founders can borrow one practical test from the approach: ask an investor which technical buyer would challenge the product’s assumptions, and what a useful introduction would help establish. The answer should be concrete. A large network helps when its members fit the problem and have time to engage. For a product without a compelling use case, another introduction simply creates another polite meeting.