Deedy Das on the Anthology Fund, competing with AI labs, and why missionaries beat mercenaries in venture

Summary

Deedy Das joined Menlo Ventures roughly 18 months ago after a career in engineering and product, most recently at Glean. His pitch for being in venture is simple: technical founders want to talk to someone who actually cares about the engineering, not just the market size.

His main focus at Menlo has been leading the Anthology Fund, a $100 million vehicle backed by Anthropic. The fund has deployed roughly 30 checks and operates differently from a standard series A process. The goal is to meet founders months before a formal round, build conviction over six to eight months, and then convert the best into full ownership positions. Two companies have made that leap so far: Goodfire, an AI interpretability startup, and Open Router, which recently closed a large series B.

The Anthropic structure

The Anthology model is a direct response to the incentive problem that plagues most corporate venture funds. Teams running balance-sheet vehicles typically aren't compensated on carry, so the investments tend to serve ecosystem optics rather than return discipline — the Slack Fund's portfolio of Slack apps being the obvious case study. By routing capital through Menlo, Anthropic outsources the sourcing, diligence, and portfolio management to a team whose economics depend on actual returns. The portfolio companies don't even need to use Anthropic's models; the mandate is companies that benefit the AI ecosystem broadly.

Das is direct that cloud credits — Anthology offers $30,000 to $100,000 — are a nice-to-have rather than the draw. What founders actually want is visibility into what the labs are building and a relationship before a competitive dynamic forces the question.

Competing with labs

Das's guidance to founders worried about lab competition comes from his Glean experience. Google had a cloud search product when Glean launched, but the Glean team knew the team running it and understood it wasn't a priority at Google's scale. The playbook is to target a market that looks like ~$50 million in revenue — too small for a lab running at $1 to $4 billion in annualized revenue (Anthropic's disclosed range for the first half of 2025) to bother with — and build fast enough to reach escape velocity before the opportunity becomes legible at that scale. Labs now face enough revenue pressure that they're unlikely to chase small opportunities speculatively.

On the Anthology conflict question specifically, Das says Dario Amodei has been unusually missionary about it. The fund has backed at least one company with an alternative approach to model development, and Amodei signed off without friction. The line Das draws is companies going directly head-on with Anthropic's core bets with significant capital behind them — that's the actual conflict threshold.

Data wars and Glean

On Salesforce and other SaaS incumbents tightening connector access to push competing enterprise search products, Das argues the framing is mostly wrong. Glean increases daily active users on Salesforce — people who weren't using the platform discover content through Glean and go back to the source. Glean also enforces source permissions strictly, so a customer can't buy 10 Salesforce licenses and expose that data to a thousand employees through Glean. The relationship is more synergistic than adversarial, and the competitive pressure from Salesforce's own AI search ambitions is the real driver of any friction, not data protection.

Where agents are actually working

The consumer agent Das is most enthusiastic about right now is 19pine.ai, a service that makes phone calls on your behalf — insurance negotiations, IRS hold queues, airline refunds, carrier cancellations. The business model is outcome-based: users pay a percentage of whatever they save, not a subscription. Das frames this as the billing model the market has been waiting for in agentic AI.

At a more structural level, he thinks the bigger unlock is domain-specific RL stacks that train agents to operate enterprise SaaS tools the way Claude Code was trained for agentic coding work. Computer use hasn't crossed the reliability threshold yet, but he sees that changing. The gap he's focused on is the last stretch from roughly 95% automation to full task completion — the residual human-in-the-loop clicks and review steps that keep agents from replacing headcount outright.

Missionaries versus mercenaries

On the AI talent war and its effect on early-stage venture, Das agrees the dynamic is real but not new. Startups have always paid $170,000 to $200,000 against Google and Meta offers of $500,000 or more. What's changed is the ceiling — $10 million annual packages for top AI researchers shift the mercenary-missionary spectrum enough that founders who would have stayed get pulled out. Das says the single most important thing he does in diligence is pressure-test whether a founder will still want to work on this problem in five years regardless of what lands in front of them. He can't fully solve for it, but it's the question he keeps coming back to.

On whether the rest of the Mag7 replicates Meta's approach to talent, Das is skeptical. Zuckerberg has a specific tolerance for long-duration bets with uncertain returns — Oculus being the precedent — that most large-cap tech CEOs don't share. The internal cost is also real: Meta employees on Blind are openly frustrated seeing colleagues recruited at $25 million a year, and the cultural math of being a fraction of a star hire's comp is corrosive enough that the productivity gains may not net out.