Edition 4

In SaaS, the classic durable moat has been the system of record: the place that stored the critical data and where workflows kicked off. Salesforce for customer data, Workday for employees, NetSuite for financials. Once you were embedded there, ripping you out was painful enough that customers tolerated a lot, including price increases, mediocre product velocity, and difficult upgrades.

The AI-era version moat may be one that borrows from clearinghouses in financial markets. A clearinghouse sits between parties that cannot fully trust each other, verifies and authorizes what they do, and keeps the receipt. As agents from many vendors act autonomously against critical data, something has to sit in the middle and decide which agent is cleared to act, on what data, with what limits, and prove what happened after the fact.

That middle layer controls four things: memory (what agents know), context (what they see), execution (what they are allowed to do), and governance (who is allowed to do what, plus the audit trail).

"If migrating off a system of record was painful, migrating off the thing that holds your policies, your permissions, and your entire audit history is probably harder."

"It's no longer 'is the model good?' Every model is good (or good enough). The question becomes 'can I see what every agent did, set policy on what it can touch, and prove it to my auditors?'"

Microsoft, Salesforce, Snowflake, and Databricks are all incumbents racing to be this toll booth every agent action passes through.

For startups on the other hand, two ways in may be 1) earn clearinghouse status vertically, in an industry the horizontal players will not go deep on, or 2) become the neutral single pane of glass across a multi-vendor mess incumbents cannot credibly govern.

Either way, "the source-of-truth era is transitioning into the source-of-permission era."

Public software staged a +38% rebound off its April 10, 2026 low in just 34 trading days, reversing a ~40% drawdown from the November 3, 2025 peak.

Aside from software being still down ~20% overall, the rebound has also been remarkably narrow. About 10 names, largely the AI and security winners (e.g., CrowdStrike, Datadog, Palantir, Snowflake), account for roughly 80% of the market cap regained, re-rating the top 10 from 9.9x to 14.8x forward revenue.

The median software company, by contrast, still trades at just 3.3x forward revenue, up only 19% off the lows and still off ~30% from the highs.

A podcast with Bill Gurley is always worth a listen. In this one, Bill shares the mental models he returns to most, including systems thinking, which matters because complex systems are multivariable and nonlinear, so optimizing one metric tends to produce second-order and third-order effects that you may only find out way later. Note he recommends a book Thinking in Systems here.

A common successful founder trait Bill sees is obsessive learning, but with one twist. They study both the disruptive edge that moves fast and rewards top-percentile expertise, but also the history and masters of a given field. This allows them to have a valuable firm understanding of the bedrock, but also recognize the need to innovate on top of it.

On AI market structure, Bill believes whether one model dominates comes down to optimization, pricing, and regulation. Gurley warns heavy regulation could entrench an oligopoly, with some large players wanting regulation as a moat against open source. Open-source approaches, however, like in China, create faster learning systems, comparing approach to two farming societies, where one shares best practices and thus evolves faster.

Benedict is always a bit of a skeptic, as far as VCs go. In this episode, he notes that coding is AI's first real product-market fit, but cautions against too much extrapolation of this fact.

He also says to treat chatbots and foundation models as low-level infrastructure, not finished products; they lack clear network effects and need verticalized skills, templates, and interfaces built on top to become useful.

Benedict believes capex has hard limits, i.e., big tech can spend hundreds of billions, not trillions a year, so a taper is coming as it did with telecoms and oil.

Despite his skepticism, Benedict believes AI multiplies software rather than replacing it, noting that every new capability spawns new problems and new tools, so he expects a messier, but larger SaaS landscape.

Note: Benedict puts out lengthy, rich presentations twice a year exploring macro and strategic trends in the tech industry, including his latest Spring 2026 edition: AI Eats the World.