Founders Fund alumni clash live: high-margin software vs. capital-intensive hard tech

Summary

Two Founders Fund alumni, Delian (now at Founders Fund) and Everett (now at Kleiner Perkins), staged a structured debate over whether capital-intensive hard tech or high-margin software produces superior long-term returns. The rivalry traces back to their time together at Founders Fund, where Everett half-jokingly proposed routing all negative-gross-margin deal flow to Delian via CRM filters.

The Core Disagreement: Gross Margin vs. Terminal EBITDA

Delian's central argument is that gross margin is the wrong lens for early-stage evaluation. Terminal EBITDA margin, he contends, is what matters, and that is determined primarily by monopoly strength. He points to NVIDIA as the clearest validation: a hardware company with the highest EBITDA margins in the Mag Seven precisely because it holds a near-monopoly. Tesla, by contrast, faces real competition and suffers accordingly. Salesforce, often cited as the canonical SaaS monopoly, holds only 25% CRM market share and produces EBITDA margins around 40% once full sales and marketing costs are loaded in.

Everett's rebuttal frames the question around ROIC, drawing on Hamilton Helmer's Seven Powers framework. His argument is not that software always wins, but that digital product form factors more naturally generate the conditions for durable power, including network effects and scalability, than most physical businesses do. He cites Costco as a counterexample of a great atoms-based business, but argues the structural advantage still skews toward digital.

The Competitive Intensity Problem

Everett concedes a meaningful shift in the software investing environment. During the 2010s, a founder targeting a vertical like HVAC software or a horizontal function like CRM typically faced only two or three competitors. That low competitive intensity meant SaaS companies could achieve good unit economics early without heavy capital deployment. Today, nearly every software category has become what he calls a "mini rideshare market," with roughly 10x more capital chasing each opportunity than existed fifteen years ago, combined with meaningful marginal inference costs that did not exist in the zero-marginal-cost SaaS era.

Delian uses this opening to argue that the same dynamic that made investors skeptical of Uber and DoorDash in the mid-2010s, namely negative gross margins and capital intensity, is now playing out in AI foundation models. Investors who swore off those businesses missed enormous outcomes. He draws a direct parallel to those now pouring capital into AI application-layer companies selling inference at a loss, quoting what he attributes to Chamath Palihapitiya: "selling a dollar for 90 cents."

Where Value Is Actually Accruing in AI

Delian challenges the application-layer thesis directly. If foundation model intelligence is becoming commoditized, value should be migrating to the app layer, where companies can swap models and capture margin. In practice, he argues, the opposite has happened. The fastest revenue growth and user growth over the past eighteen months belongs to the foundation model companies, not the application layer. He notes that Anthropic has moved to positive gross margins with expanding economics, while Kleiner Perkins has invested in neither OpenAI nor Anthropic.

Everett pushes back by reframing ChatGPT not as a foundation model product but as a consumer application with massive brand power. He argues ChatGPT is the first billion-plus user consumer application built by a new company in a long time, and that its brand equity constitutes genuine power independent of which underlying model serves it. Swapping Claude 3 Sonnet into ChatGPT, he suggests, would not dislodge its user base.

Token Cost Trajectory and Inference Economics

On the question of how quickly inference costs will fall, Everett draws a distinction between frontier and non-frontier workloads. The roughly order-of-magnitude-per-year decline in token costs held for a period but has slowed materially at the frontier. For tasks that do not require frontier intelligence, he expects costs to continue falling sharply via open-source and distilled older models. The strategic question for any AI company is therefore twofold: does it have enough pricing power to charge above marginal token cost, and how much of its inference actually requires frontier capability versus cheaper legacy models.

He highlights Cursor and a recent essay by Chris Paik at Pace Capital as a live test case. Developers, as taste makers with high sensitivity to model quality and price, represent a particularly demanding cohort. Whether Cursor has genuine user lock-in or is simply a well-designed interface to frontier models remains, in his view, unresolved.

Revenue Quality Across Both Domains

Both agree that revenue quality analysis is underdeveloped outside software. In hard tech and defense, Delian flags a wide spectrum from a program of record contract, which carries high durability, down to an SBIR grant, which is essentially experimental budget and should carry a very different multiple. He expresses concern that investors new to the space are applying SaaS-era heuristics, including Rule of 40, without understanding the years-long path to positive gross margin in manufacturing or aerospace. He notes seeing a 100x revenue multiple applied to a hardware company recently.

Everett applies Bill Gurley's "10x Revenue Club" framework to AI, identifying lower gross margins and uncertain customer stickiness as current weaknesses relative to SaaS. He identifies one potential compensating factor: agentic products that displace labor can be monetized on metered consumption at contract sizes that dwarf traditional SaaS. He points to Claude Code as the first product where this is demonstrably happening, with annualized developer spend representing roughly one-tenth of a developer's fully loaded cost to an employer.

Portfolio Reference Points

Both cite Rippling and Hadrian as 2021 bets that have performed well while illustrating opposite profiles. Rippling carries high gross margins but requires sustained sales and marketing spend to acquire customers. Hadrian started with deeply negative gross margins but, at scale, serves only ten to fifteen customers and generates inbound revenue demand with minimal sales effort, a dynamic Delian frames as the hard-tech equivalent of a monopoly.

Delian also takes a pointed shot at Everett's investment in Captions, an AI video captioning company, arguing two Stanford graduates and $100 million could replicate it. He contrasts that with Varda Space Industries, which manufactures in orbit and faces no credible replication threat regardless of capital.

Macro and Government Capital

On the question of government capital entering early-stage hard tech, including a potential U.S. government stake in Intel, Delian is sanguine. Cheap sovereign capital marking up reindustrialization companies is, in his framing, a feature rather than a threat. Everett notes that Founders Fund's enterprise value likely tripled the night Trump was elected, a point Delian does not dispute.