Key points
- Leonis Capital is closing a $75m Fund III with $1.3m first cheques, deploying 2027 to 2030.
- Zhao says AI agents will need new infrastructure for identity, compliance and payments, and software moves from systems of record to systems of action.
- Leonis's Alpha data finds three- and four-star founders outperform five-star pattern-recognised founders, while professors struggle as founders.
Leonis Capital bets on AI agents before the market does
Leonis Capital is a small research-driven fund that tries to buy AI themes one or two years before the market does, and its structure matches that goal. Zhao keeps funds small, takes most of his ownership at the first cheque, and backs agent infrastructure and systems of action instead of pattern-recognised founders. His LP base of AI lab engineers and researchers is the audience for that research.
Takeaways
- Leonis invests ahead of the AI capability curve. Zhao looks for companies one or two years ahead of AI capability rather than chasing current hype. Leonis's LPs are early engineers and researchers at AI labs, and the firm publishes its research on its website and Substack.
- Software must become systems of action. Agents will multiply across companies and consumers, so systems of record like Salesforce and Anaplan give way to software where agents act reliably. That needs new infrastructure for agent identity, compliance and payments. Leonis backed Agent Phone, a Twilio for agents.
- Leonis refuses to raise a bigger fund. An LP suggested keeping a fund open longer to raise $200m while the market is hot. Zhao keeps his cadence instead, arguing a $200-300m fund would fight established firms at Series A, where 10 or 15 term sheets show up.
- Three and four star founders win. Leonis scores founders from one to five stars in its internal AI, Alpha. Five-star founders are pattern-recognised names from big labs, yet three- and four-star founders have produced the best outcomes. Professors struggle because building a company needs full-time commitment.
- Agent infrastructure is the underhyped bet. Zhao compares agent infrastructure to early Stripe and to Marqeta riding the on-demand wave with DoorDash. Agents struggle with human-formatted data in CRMs like Salesforce, leaving room for new tools. Physical AI is overhyped, and interactive world models look weak.
- Leonis skips regulation-dependent deals. Healthcare, finance and legal carry both headwinds and tailwinds, so Leonis stays at the application level and skips deals where regulation decides the outcome. His underwriting leans on the technology and the team.
Leonis fund by fund
- Fund I was $10m, with cheques of $100,000 to $200,000 and no reserves. Five years in, Zhao calls it a really good fund with one write-off and probably two unicorns.
- Fund II raised $40m when interest rates were high and allocators preferred savings accounts. Twenty percent of its companies met the ownership requirement.
- Fund III is $75m, with a $1.3m core first cheque and 20% reserved for follow-ons. Leonis expects 50% of companies to meet the ownership requirement and will deploy it from 2027 to 2030.