All podcasts / Invest Like the Best / Summary

Gokul Rajaram - Lessons from Investing in 700 Companies

2026-01-29 - 77 min - source - Read full transcript

Patrick O'Shaughnessy (host)Gokul Rajaram

Key insights

Long-horizon coding agents have collapsed the traditional PM/designer/engineer split into a bottoms-up, engineer-led workflow.

Rajaram pinpoints December 2025-January 2026 as the inflection point: resilient, long-running agents now let engineers, researchers, and PMs build directly in code, so PMs mainly define the 'why' at a high level and evaluate output rather than write detailed specs. Companies are adding explicit 'prototyping interviews' to test whether PMs can work hands-on with these tools.

ai-native-product-development

Judgment, not code output, is the one skill that stays futureproof as AI-generated code and design proliferate.

With large numbers of AI agents producing code, the bottleneck shifts to deciding which output matters and reviewing it for correctness, security, and design-system fit. Rajaram argues this editorial judgment - not raw production capacity - is what humans uniquely contribute in an era of near-infinite output.

ai-native-product-development

Non-deterministic AI software forces PMs to own evaluation systems, since a slight input variation can produce a wildly different output.

Unlike deterministic software where the same input always produced the same result, AI products can behave inconsistently across near-identical inputs. Rajaram says PMs and researchers now must build both human and AI evaluation ('evals') to judge whether output quality is acceptable across use cases, sometimes writing AI to evaluate AI because humans can't keep up.

ai-native-product-development

Software companies priced on seat-based utility are far more exposed to AI disruption than those holding non-timeless, hard-to-migrate data.

Rajaram singles out Zendesk as an example: an AI agent can sit alongside it and gradually siphon off paid seats since customers don't need an all-or-nothing switch. Companies like NetSuite or Salesforce are more insulated because their data has a long half-life and ripping them out is career-limiting for whoever champions the switch, buying them time to build their own AI agents on top of that data.

software-durability-and-moats

System-of-record incumbents are actively cutting off API access to stop AI agent startups from hollowing them out from within.

Slack, owned by Salesforce, cut Glean's data access in 2025 to prevent it from becoming a thin AI layer that slowly captured Slack's value. Rajaram sees this pattern spreading: incumbents are blocking APIs outright, charging per API call, or bundling their own free agents, which forces agent-layer startups to build their own systems of record or years-long migration tooling to survive.

software-durability-and-moats

Durable products need at least one of five moats: a scarce asset, a control point over money or data, hardware, essential-workflow lock-in, or network effects.

Citing Hamilton Helmer's 'Seven Powers' framework, Rajaram points to DoorDash's three-sided network, Toast's free hardware paired with embedded payments, Mercury's regulatory switching friction, and Sierra's exclusive access to Bret Taylor as concrete examples of each moat type.

software-durability-and-moats

There are only three durable ways to make money in advertising: own coveted first-party inventory, drive a measurable outcome at scale, or become the exclusive channel for a large demand source.

Google Search monetizes intent data and Facebook monetizes identity data; ChatGPT is unusual in combining both. AppLovin built a business purely on delivering app installs at a set cost without owning inventory, while The Trade Desk succeeds as the exclusive spend allocator for advertisers like Procter & Gamble. Middlemen that build on top of Google, Facebook, or OpenAI's inventory get squeezed because the platform eventually absorbs their capability.

advertising-business-models

Every ad platform needs an explicit 'engagement budget' capping how much user engagement it will sacrifice for ad revenue.

At Facebook, the News Feed and ads teams operated under a shared cap on the maximum engagement dip attributable to ads, measured against a permanent no-ads holdout group. Rajaram argues any newly-monetizing AI interface (ChatGPT, Gemini) needs the same holdout-group discipline before ads erode usage, since engagement reliably drops once ads enter a previously unmonetized surface.

advertising-business-models

A weekly CEO email built around three sections - top of mind, performance update, and miscellaneous - is the most effective way to scale communication once a company outgrows a single room.

Rajaram recommends spending 60-70% of the email on 'top of mind' (what's genuinely keeping the CEO up at night), favoring candor because it invites the team to propose solutions, plus a shorter performance-update section and a lighter miscellaneous section for recognition and announcements. He has used the format himself and seen roughly 15 CEOs adopt it.

leadership-communication

The strongest hiring signal is a live work project done without AI, because talk-based interviews let candidates 'BS their way' through non-engineering roles.

Square gave CorpDev candidates a live acquisition analysis to defend; DoorDash gave candidates $10-20 and challenged them to acquire 1,000 customers in a few hours. The goal wasn't success but observing resourcefulness, creativity, and whether candidates pushed back on a flawed premise rather than blindly executing it.

founder-and-talent-evaluation

Founder authenticity - an origin story rooted in real lived experience with the problem - is the top predictor Rajaram screens for as an investor.

Google, Facebook, and DoorDash all began as founders' curiosity-driven toy problems rather than deliberate business plans. Rajaram distrusts founders whose main motivation is starting a company with a friend, and probes the 'idea maze' by asking why they rejected five or six alternative approaches, testing whether they're genuine students of their industry's history.

founder-and-talent-evaluation

Career job-hopping (12-18 months per role) is one of the biggest hiring red flags, because it takes three to four years to have real impact at a company.

Rajaram treats a pattern of consecutive short stints as evidence someone hasn't built lasting value or a durable network, distinguishing it from a single early-career misstep that didn't work out. He calls this a top disqualifier as a hiring manager.

founder-and-talent-evaluation

Books referenced

7 Powers - Hamilton Helmer - Rajaram cites Helmer's framework directly when listing the handful of durability sources (scarce assets, control points, hardware, essential workflows, network effects) a company needs embedded in its business model from day one.

Companies

Marathon Management - Rajaram's venture firm, referenced as the lens for his investing framework and founder-evaluation questions.
Google - Where Rajaram built AdSense and observed Larry Page and Sergey Brin's technology-first, scale-obsessed culture; source of the self-serve and North Star metric philosophy he later applied elsewhere.
Facebook (Meta) - Rajaram led the ads product team; Zuckerberg originated Custom Audiences after a Zynga complaint, and Facebook enforced an explicit engagement budget between News Feed and ads.
Square - Jack Dorsey and Jim McKelvey's company; example of design removing friction, risk pushed to the transaction level, and GPV as North Star metric.
DoorDash - Cited as a network-effects moat (restaurants, dashers, consumers) and for its unconventional work-project interview (give candidates $20 to acquire 1,000 customers).
Zendesk - Named as one of the most exposed legacy software companies because its per-seat pricing is pure utility that AI agents can quietly siphon off.
Slack - Cut off Glean's API access in 2025 to stop it from becoming a thin AI layer on top of Slack's data; called out as precarious because its data has a short half-life.
Salesforce - Owns Slack; discussed as a system of record whose CRM customer data gives it more durability than utility-priced tools.
NetSuite - Example of a durable system of record - ripping it out is career-limiting for whoever championed it, regardless of per-seat cost.
Glean - AI search/agent company that lost API access to Slack data, illustrating incumbents blocking agent-layer startups.
Toast - Stickiness example combining free hardware (with a return penalty) and embedded payments.
Mercury - Banking example of stickiness via money flow plus regulatory friction to switching.
Sierra - Cited for a unique-asset moat: exclusive access to co-founder Bret Taylor's network and credibility.
AppLovin - Built a 100+ billion dollar ad business purely by driving one outcome - mobile app installs - at a defined cost, without owning first-party inventory.
The Trade Desk - Example of the third ad business model: exclusive channel for a large demand source (e.g., Procter & Gamble's ad spend).
Figma - Self-serve insurgency example - blocked by a top-down sales push at Square, then displaced Sketch two years later via bottoms-up adoption by a design manager.
Sketch - Incumbent design tool that Figma displaced through grassroots, self-serve adoption rather than direct sales.
Palantir - Outcome-based selling example: commits to solving a client's top problem in six months and only gets paid if it delivers.
OpenAI / ChatGPT - Discussed as combining intent and identity data for the first time, making it a uniquely powerful future ad surface; also cited for articulating that ads should not influence AI recommendations.
Fair - Max Rhodes' B2B marketplace, used as a founder-authenticity example rooted in his college umbrella-brand distribution problem.

Techniques and frameworks

North Star Metric with check metrics - A single growth metric (e.g., DoorDash's GMV, Facebook's DAU, Square's GPV) must be paired with guardrail metrics like margin or retention so teams can't inflate the North Star by degrading the business.
Customer behavior change hypothesis - Rajaram's product philosophy: no feature should ship without a stated hypothesis that customers will move from behavior X to behavior Y.
Self-serve definition test - A product only counts as self-serve if a customer can onboard and use it without ever talking to a company employee; Rajaram treats this as the gate for scaling distribution beyond direct sales.
Founder authenticity / origin story screen - Rajaram's first investing question is always the founding story, screening for lived experience with the problem rather than 'starting a company with a friend.'
Idea maze questioning - Probing why a founder chose their specific approach over five or six alternatives, testing whether they're students of their industry's history.
Board buddy system - Pairing each board member with a specific management-team member for regular contact between board meetings, which Rajaram considers more valuable than the meetings themselves.
Weekly CEO email (top of mind / performance / misc) - A three-section format - 60-70% of it on 'top of mind' - that Rajaram has used himself and seen roughly 15 CEOs adopt to scale communication past the single-room stage.
Prototyping interview - A new interview stage companies are adding to test whether product managers can hands-on prototype with AI coding tools rather than only write specs.
Outcome-based pricing and selling - Shifting from seat/utility pricing to charging per resolved outcome (e.g., per ticket resolved instead of per seat), which Rajaram says legacy software companies need to survive AI-agent competition.
Span of control minimum - Rajaram's rule that managing fewer than 10 people should not be allowed at a company - anyone below that threshold should be an individual contributor instead.
Ads engagement budget - A shared cap between growth/engagement teams and ads teams limiting how much user engagement can be sacrificed for ad revenue, tracked via a permanent no-ads holdout group.

Summary

Gokul Rajaram - who built ads and product at Google, Facebook, Square, and DoorDash before founding Marathon Management and investing in more than 700 companies - joins Patrick O'Shaughnessy to trace how product building, software moats, and advertising economics are all being rewritten by AI at once. His central claim is that long-horizon, resilient coding agents crossed a threshold in December 2025 and January 2026: product managers no longer prescribe detailed specs, engineers and PMs build together bottoms-up directly in code, and the scarce skill left to humans is judgment - deciding what's worth building and evaluating whether non-deterministic AI output is actually good, since the same input can now produce wildly different results.

That same disruption reshapes which software survives. Rajaram splits legacy companies into two camps: those priced on seat-based utility (Zendesk is his example), which AI agents can quietly siphon share from seat by seat, and those sitting on durable, non-timeless data (NetSuite, Salesforce), which are protected because switching is organizationally risky and their data can be used to train their own bundled agents. He points to Slack's 2025 decision to cut Glean's API access as the leading edge of a broader pattern - incumbents blocking, metering, or bundling around their APIs to stop AI agent startups from becoming thin layers that hollow them out. Durable products, in his framing (borrowing Hamilton Helmer's Seven Powers), need at least one of five things: a scarce asset, a control point over money or data, hardware, essential-workflow lock-in, or network effects - illustrated through DoorDash, Toast, Mercury, and Sierra.

On advertising, Rajaram distills two decades of building Google AdSense and Facebook Ads into three business models that actually work: own coveted first-party inventory (Google, Facebook, and now ChatGPT, which uniquely combines intent and identity data), drive a measurable outcome at scale without owning inventory (AppLovin and app installs), or become the exclusive channel for a large advertiser's spend (The Trade Desk and Procter & Gamble). Everything else - especially middlemen building on top of Google, Facebook, or OpenAI's own inventory - gets squeezed because the platform eventually absorbs whatever capability the middleman proves valuable. He also describes the "engagement budget" discipline Facebook used to cap how much user engagement ads teams were allowed to sacrifice for revenue, which he thinks every newly-monetizing AI interface will need.

The conversation closes on leadership and talent, drawing on stories from Larry Page, Sergey Brin, Mark Zuckerberg, Jack Dorsey, and Eric Schmidt. Rajaram credits Schmidt with teaching him that strategy presentations should use only images because "people don't remember words, they remember how things made them feel," and credits Dorsey with treating product managers as "editors" who cut rather than add. His own leadership toolkit includes a three-section weekly CEO email (top of mind, performance update, miscellaneous), a board-buddy system pairing each board member with a management-team counterpart, and a hard rule against managing fewer than 10 people. For hiring and investing, his consistent filter is authenticity and doing: real work projects instead of talk-based interviews, founder origin stories rooted in lived experience rather than "wanting to start something with a friend," and skepticism toward candidates who job-hop every 12-18 months, since he believes real impact takes three to four years to show up.

Notable Quotes

"The one thing I think that's going to be truly future proof is judgment." - Gokul Rajaram

"People don't remember words. They remember how things made them feel." - Gokul Rajaram, recounting Eric Schmidt's instruction to present strategy using only images

"Apparently, Rick Rubin would say that he wasn't a producer. He was a reducer." - Gokul Rajaram

"You cannot lead with what your product does anymore. You've got to lead with what is the outcome you can deliver." - Gokul Rajaram

"It's not the board meeting that truly matters, it's all the things between the board meetings that are the real thing when things get done." - Gokul Rajaram