🔥More Shots on Goal: How AI Rewrites Portfolio Construction

Most seed fund models still assume that roughly half of your companies raise a Series A. The current number is closer to a fifth.

That one input decides how many companies you enter, how much capital you hold back, and how many positions ever become a follow-on decision. Move it from 55% to 20% and the whole model needs rebuilding.

This is the fund-model sequel to AI Creates More Startups But Fewer Winners, which ended on a barbell. Here is what the hidden-gems end costs you in practice.

Two independent datasets say the same thing about seed graduation, using different definitions.

Definitions differ. Crunchbase counts any later round or exit with no time cap, Carta counts a Series A inside 24 months. The direction is identical.

The 2024 cohort is young and will drift up, so treat 16% as a floor rather than a finding.

Three things moved alongside it. The median seed round tripled since 2018 to roughly $3M, the time between Seed and Series A doubled to more than two years, and the revenue bar more than doubled from about $1M ARR to $2-4M ARR (Crunchbase, May 2026).

Part of this is the rate cycle. The larger part has a precedent.

When cloud computing collapsed the cost of starting a software company, Ewens, Nanda and Rhodes-Kropf documented what venture capital did about it. Their finding, in their own words, was an increased prevalence of a "spray and pray" approach, where investors provide a little funding and limited governance to an increased number of startups that they are more likely to abandon, but where initial experiments significantly inform beliefs about the future potential of the venture (Journal of Financial Economics, 2018).

Smaller cheques, more companies, more abandonment. That is the conclusion of this essay, already documented for a smaller shock a decade ago.

Three mechanics hold the rate down.

Cheaper to start and more expensive to prove is exactly what widens the top of the funnel while narrowing the exit. Outliers raise more money faster than ever before while an increasing majority of the market raises less and slower.

Start with the simplest version. Hold the number of companies you want to see reach a Series A constant at 12, and solve for how many you have to enter.

No fund triples its entry count and keeps everything else the same. The cheque size falls, the ownership falls, or the number of graduates falls. Usually all three.

The power law pushes the same way. Of more than 21,000 financings between 2004 and 2013, Correlation Ventures found 65% returning less than 1x and only 4% returning 10x or more (via Seth Levine). At a 4% rate, 22 entries buys an expected 0.9 investments at 10x or better. Sixty entries buys 2.4.

Take a €100M seed fund with €85M investable. Most early-stage managers run close to a 1:1 split between initial cheques and reserves.

Model A is that fund under 2019 conversion. Model B is the identical structure under 2026 conversion. Model C rebuilds it.

Model B is the interesting column, because it is what most funds are running right now without having decided to.

The reserve pool did not shrink, but the number of companies it can legitimately be deployed into fell by roughly two thirds. That leaves €8.5M sitting behind each graduate.

A pro rata right on a €15M Series A for a 10% seed position absorbs about €1.5M. So Model B holds more than five times the capital that its own pro rata rights can take up.

That capital has three places to go. Super pro rata at Series A prices, where your seed-stage edge is already spent. Bridge rounds into companies that missed the bar. Or nowhere at all. All three are worse than having deployed it at entry.

Model C fixes the ratio rather than the reserve. Forty entries at a 20% rate produces 8 graduates, and €25.5M of reserves gives each of them €3.2M, which is roughly two times pro rata and leaves real room to lean into the best two or three.

The price sits in the fourth row. Entry cheques fall from €1.70M to €1.49M while seed rounds have tripled since 2018, so ownership per entry drops.

Pricing is also bifurcated. Carta puts an AI foundational model Series A at a $300M median valuation against $55M for a non-AI company at the same stage (State of Private Markets Q1 2026), so what €1.49M buys depends entirely on which half of the market you are entering.

The ownership floor moves with the outcome distribution. The static version of this objection says smaller cheques buy less ownership, so each winner returns less. In Model C you enter at €1.49M for roughly 8% of an €18M post and dilute to about 3.5%, so a €2B outcome returns €70M on a €100M fund against €90M for Model A's larger entry.

That objection is weaker than it looks, because the outcome is not fixed at €2B. The fifty most valuable private companies were worth a combined $5.1T in September 2026 (Multiples.vc). At the top of that distribution, 3.5% clears any floor you care to set.

Outlier outcomes have increased more than dilution-adjusted shareholding at exit has decreased. Said differently, you can afford lower initial shareholding if the likelihood adjusted outcome can be big enough.

Partner attention is scarce, and spreading it too thin has real costs. Abuzov finds that startups backed by VCs during periods of unusually high workload are 9% less likely to subsequently IPO or be acquired (JFQA, 2025).

The implication is not that every company needs equal attention, but the opposite: as portfolios grow, attention has to become more selective. Back more companies where the economics justify it, but concentrate partner time and board seats on the investments where active involvement can make the biggest difference.

Width comes from more entries, not a bigger fund. Kaplan and Schoar found the relation between fund size and performance concave, and that a GP raising a larger fund posts weaker returns than their prior one (Journal of Finance, 2005). Raising €200M so you can write 40 cheques at the old size is the move that finding warns against.

Five changes, in the order I would make them.

The conversion rate between seed and Series A fell by more than half, and most fund models still carry the old number.

The response the evidence supports is more entries, smaller cheques, a reserve ratio nearer 30% than 50%, and follow-on capital concentrated into two or three names. This is what venture capital did the last time the cost of experimentation collapsed. AI is the same shock with a larger amplitude.

One caution outranks all of it. Width pays only if your 40th entry is as good as your 25th. If widening means lowering the bar rather than casting the net further, every number in this piece works against you.

The firms that get this right will look undisciplined for about two years. Then their entry cohorts will start clearing, and the arithmetic will look obvious in hindsight.

Stay driven,
Andre