Ev Randle of Benchmark argues the old SaaS playbook is broken, since AI inference is expensive and high gross margins can signal customers are not using AI features.
Randle says the most popular AI companies invert nearly every software metric, and bigger AI companies no longer get safer, with devaluation from a prior round a live risk at any scale.
Randle says the market is not zero sum, because demand for frontier, open source and on-device inference is growing together, and he expects Anthropic's roughly $380bn round to yield huge pre-IPO returns.
AI Breaks the Software Investing Rulebook, Says Benchmark's Ev Randle
Benchmark's Ev Randle argues that the spreadsheet-era rules for judging software companies no longer hold in AI. Inference is costly, margins are lower, and size does not guarantee safety. Revenue now flows from monetizing inference, and the growing pile of private capital and coming IPOs is built on that shift.
Takeaways
Old SaaS rules largely fail for AI companies. Software was prized for 70 to 90 percent gross margins, little services load, high retention and rule-of-40 scores. Randle says popular AI companies invert most of these. Inference is expensive, so a high margin can signal customers are not using the AI features.
Scale no longer makes AI companies safer. Some AI businesses are already well over $1bn in revenue without proven unit economics or durable differentiation. Randle calls the risk-reward relationship broken, because being devalued from a prior round is a live risk at any size.
AI revenue comes from monetizing inference. Randle frames AI economics as price times quantity times margin. App companies have lower margins than SaaS, while inference platforms sign nine-figure contracts. Developers spending $3,000 a month each on Claude Code show a small software line item becoming very large.
Demand for inference is not zero sum. Demand is growing for frontier models and for open source and on-device inference alike. Cognition showed an open source model saving 95 percent on simple tasks. Open source reaching 95 percent of frontier quality would scare the labs.
Enterprises need independent agent platforms. Gumloop extends what happened in coding to white-collar work. Models are uneven, with Claude strong at coding and open source cheaper, so enterprises want a vendor that is not tied to one lab.
Private markets now favor late-stage investors. Companies stay private longer, neolabs may need $2bn of compute, and late-stage holders see more upside than Series C investors. Randle says General Catalyst and Andreessen Horowitz now act as alternative asset managers rather than venture firms.
AI IPO liquidity could dwarf earlier cycles. Pre-IPO investors in Slack, DoorDash, Snowflake and NewBank made two to five times. Anthropic's $380bn round could return 35 times the Snowflake round at a $1.5tn listing. Randle expects the cash to reach San Francisco housing.
How Benchmark approaches its bets
Benchmark invests in founders, not categories.
It closed its Starcloud investment weeks before Elon Musk began promoting orbital data centers.
The portfolio includes Cerebras, Sierra, LangChain, Legora, Fireworks, Mercor and HeyGen.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Randle says he may look wrong about much of this in a year.