AI Turned Early Revenue into Noise, So Seed Investors Must Underwrite Earnings Power Instead, Marathon's Gilroy Argues

1 hour ago

4 min read

What's New

AI erased the barriers to starting a company and, at the same time, broke revenue as a signal of product-market fit, which forces early-stage investors to underwrite seed and Series A companies the way growth investors underwrite mature ones. Michael Gilroy, Founding Partner of Marathon Management Partners, argues this in a podcast on How I Invest. Every trend now has 20 companies instead of 2 or 3, and enterprise buyers purchase AI tools out of fear of being left behind, so a company can run from zero to $10 million of ARR and back to zero. Marathon replaces revenue multiples with gross profit and projected net income conversion, replaces call transcripts with 15-plus back-channel customer conversations, and does 2 to 3 deals a year. Anyone still pricing seed rounds on ARR is pricing an experiment.

Why It Matters

The prevailing model treats fast ARR growth as the proof point and AI-generated transcript analysis as the diligence edge. Gilroy says both are shared inputs with zero differentiation: every investor reads the same transcript, and the revenue behind it may be a three-month pilot annualized. On the other side are the multi-stage platforms with 5,600 portfolio companies and a media arm, and the seed funds chasing 15% ownership in the hottest AI category. Marathon manages $400 million, so the concentrated, human-heavy approach is also its pitch to founders and LPs. Gilroy concedes it is not scalable by design, and that his 100% seed-to-Series-A graduation rate on nine deals could be luck.

Big Picture Drivers

Barriers to entry are gone: Infrastructure from the Twilio era through today's tooling means a product can stand up overnight, and Google, Meta, OpenAI, and Anthropic release IPO-grade founders at record pace. Capital is more abundant than in 1999, so a weekend idea can have a term sheet by Wednesday.
Fear-driven buying degrades revenue quality: Security buyers Gilroy interviewed over 12 to 18 months bought out of FOMO, often adding all 20 solutions in a subsector. Cohort damage shows up in 6 months, after the round has closed.
Gross profit variance in AI now matches fintech: Software once clustered at 65 to 80% gross margin, which made revenue multiples workable. AI companies range from negative 10% to 80%, so Marathon models net income conversion off gross profit and a public-market earnings multiple at maturity, even at Series A.
Capital incinerators versus capital savers: One portfolio company spent two years building to feature parity, showed near-80% gross profit at Series A, and burned about 5% of that round before raising again. Founders who promise inference costs will fall and margins will inflect in five years need giant economies of scale and network effects to be believed.
The number two problem: Being right on trend and market but backing the second or third company loses the money and the opportunity cost of concentrating in the winner. When the pick is unclear, Marathon waits for the A or the B.
Pre-portfolio work replaces the 48-hour term sheet: Marathon picks one or two companies per trend, works for them as if on the board, and keeps 3 to 4 such relationships at a time. The Striker security round came after the firm brought tens of customers before investing.

By The Numbers

20 versus 2 to 3: Companies per trend now versus before AI, by Gilroy's count.
15+: Back-channel customer calls Marathon runs per deal, up from 2 or 3 historically, on top of 5 front-door references.
2 to 3: Investments per year across four investors, for a fund one of 10 to 15 names.
~5%: Share of its Series A round one portfolio company had burned a year and a half later.
Negative $5 million to $300 million: CloudWalk's gross profit run rate at Gilroy's entry versus its net income today, the case for backing negative-margin businesses when the cost curve is knowable.
0 and 7.5, rising to 12.5: Marathon's co-invest carry, which steps up only if the deal beats the model it showed LPs.

Key Trends to Watch

Gross-profit underwriting spreading to seed: Watch whether other early-stage firms abandon ARR multiples for margin-adjusted models as AI companies with 10% and 70% gross margins raise at the same revenue multiple.
Cohort data as the AI diligence battleground: With pilots inflating run rates, six-month retention on 2025 to 2026 cohorts will separate need-driven from fear-driven revenue.
Vertical versus horizontal disruption from the labs: Gilroy expects labs to absorb horizontal tools and leave specialized verticals, his Toast-over-Block example, alone. Watch which categories the labs enter next.
Performance-linked co-invest fees: A carry that rises only above the GP's own model gives LPs a reason to trust the numbers. Watch whether it spreads beyond first-time firms.

Memorable Quotes

"Revenue quality is much much lower if somebody's buying it out of fear instead of out of need." Why ARR no longer settles the question at seed.
"AI has created an environment where having agents running around booking meetings for you and doing diligence is no longer an edge." Automation as table stakes, with the human call as the remaining differentiator.
"The biggest mistake you can make in our business is backing the number two, three, or four business in any given market that ends up working." The cost of picking wrong in a 20-company trend.
"This job is not spreadsheets." The model is wrong on delivery by design; its value is the questions it forces.

The Wrap

The thesis holds if AI companies underwritten on gross profit and earnings power at seed keep graduating and attracting growth capital while ARR-priced peers show cohort collapse, and if concentrated four-person firms doing 2 to 3 deals a year win allocations against platforms through pre-portfolio work. It fails if inference costs fall fast enough that today's capital incinerators reach margin parity on schedule, if labs push deeper into verticals than Gilroy expects, or if fear-driven revenue turns out to stick because switching costs arrive before the cohorts churn. The 2025 and 2026 AI seed cohorts will show their six-month and twelve-month retention over the next two years, which is when revenue either becomes a signal again or confirms it was noise.