Every angel carries a mental model of the upside, whether or not they have written it down. Mine is that I have a 1 in 10 chance of making 10x, a 1 in 20 chance of 20x, a 1 in 100 chance of 100x. Scale and probability move together, inversely, all the way out.

That is Zipf’s law: a power law with a tail index of exactly 1. Plot it on a log-log scale and my mental model is a straight line sloping down. Knowing the name is not the interesting part. As a mathematician, what I wanted was the consequences. If that is really the shape you are investing into, what follows? The answers turn out to be uncomfortable for almost everyone in the asset class, including me.

The shape is right, my optimism is not

The data says the shape is real, and it isn’t new. Correlation Ventures looked at more than 21,000 financings and found that about 65% failed to return the capital invested and only 4% returned 10x or more. Moonfire Ventures, pulling those studies together, fit real venture returns to a power law with a tail index of about 2.

My model has a tail index of 1, which makes it optimistic, and the gap is worth sitting with because it is not a rounding difference. A tail index measures how fast the probability falls as the payoff grows. At an index of 1, halve the probability and you double the payoff, forever. At an index of 2 the payoff grows far more slowly against the same drop in odds. My 1-in-100 shot at 100x is, on the real numbers, a good deal rarer than one in a hundred.

There is a second thing hiding in that number. A power law with an index at or below 2 has no well-defined average. The integral that should give you the mean doesn’t converge, because the rare enormous outcomes dominate any finite sample you could ever collect. Real venture sits right on that edge. Whatever the true index is, it is close enough to 2 that the average return is a fragile idea, a number that lurches every time another outlier lands in the dataset. Canva returned something like 2000x to its earliest Australian backers and Uber returned roughly 5000x to Jason Calacanis’s angel cheque. One of those in your sample and the average is a different number. Which is a mathematician’s way of saying the mean is the wrong statistic. There is no version of this asset class where returns cluster around it, and building a portfolio as though they might is the original error.

Everything I learned about portfolios assumed a bell curve

At BT in the early 1990s I did quant work on portfolio construction. It was early days for the discipline, and most of the effort went into understanding how asset classes interacted. I assessed different means, different variances, correlations between them, and out of that an efficient frontier you could optimise along. The distributions were assumed normal. That assumption sat so deep in the machinery it never came up for discussion.

That framework is still what people mean when they say portfolio construction, yet it’s largely unexamined. Yes, it is rigorous, it has a Nobel in it, and it is the reason a sophisticated investor’s first instinct about a ninety-company portfolio is diminishing returns. They are not being lazy. They are running a model that was properly derived for a distribution that is not this one.

Under Zipf, the tools don’t work. Mean-variance optimisation prices risk as variance, and here the variance is the return, the thing you are trying to buy. There is no frontier to sit on. Step back and describe the objective plainly: you are trying to hit the top 10%, the top 1%, the top 0.1% of the distribution. That is the entire job. And if you cannot reliably pick which company that will be, and the evidence says you cannot, then selection is not a lever. Nor is sizing or timing. The only lever left is how many draws you take.

“More shots on goal” is an unsatisfying answer for anyone trained the way I was trained, which is part of why the industry resists it.

What the tail does to your IRR

The way I think about the consequence is simpler: the more investments I have, the higher my IRR. That is not a hunch. Abe Othman and Nigel Koh at AngelList looked at more than 10,000 investor portfolios and fitted a regression of median IRR against angel portfolio size. The coefficient was 9 basis points per investment. On their reading, the typical investor holding 100 investments outperforms the typical investor holding one by almost 9% a year.

Read that carefully, because the obvious reading is the wrong one. The 9 basis points is not what the marginal deal earns on its own capital. It is the IRR of the entire portfolio moving. The ninety-first investment doesn’t earn 9bp on its own cheque and stop there. It lifts the return on all ninety-one, because what it changes is not the size of the book but the distribution the book is drawn from. Every deal is a draw from the same fat tail, and each additional draw improves the odds that the tail shows up somewhere in the portfolio. The new position isn’t a position, it’s a parameter.

Which means: add eleven investments and the IRR on the entire portfolio rises by a full percentage point. Sit with that for a moment. In public markets, alpha is fought over in basis points. Careers are built on thirty of them, and funds market themselves on less. Here is an asset class handing out a hundred at a time, for nothing more than turning up more often, and the received advice is to stop early.

That is where the objection breaks, and the objection is usually phrased badly. People call it diminishing returns, which conflates two different things. Is the first derivative positive – does IRR still rise as the portfolio grows? Yes, and that is the 9 basis points. Is the second derivative negative – is that rate of increase flattening? Probably. I would expect it to. But a concave increasing function is still increasing. To justify stopping you need the first derivative to reach zero, and nobody has produced an angel portfolio size at which it does. AngelList fitted a straight line through their data; if the curve turned over inside that range, the fit would have found it. The orthodoxy has mistaken flattening for a maximum.

Look at the chart in that paper and the honest thing to say is that the scatter is wide. At fifty investments the median IRRs run from roughly zero to twenty-five per cent. The line slopes up from about 5% at the left to about 14% at a hundred, and the dots are all over it. That noise is not a defect in the data. It is what a fat tail looks like from the inside, and it is why any individual angel’s experience, mine included, proves nothing. The slope is the finding. The slope is enough.

So how many shots?

The received number is 20 to 25, and it is a good number, badly used. It is roughly where I landed in Roll the die, which works the Bernoulli maths through properly: at around twenty-five investments you have a 90%-plus probability of catching at least one winner. That is a floor. It is the answer to “how few can I get away with,” which is a real question, and twenty-five is a real answer to it.

Then the floor became a target. Somewhere between the arithmetic and the received wisdom, “you need at least twenty-five” turned into “twenty-five is the number,” and a minimum acquired the authority of a maximum. Nobody decided this, it is what happens to a threshold when it is the only number anyone quotes.

Read what the floor says. It gives you a high probability of one winner. It says nothing about what happens at fifty, or ninety, or two hundred, and under a power law, one winner is not the objective. The objective is exposure to a tail that pays in orders of magnitude, and a 90% chance of catching one is a long way from the best you can do.

My own simulation model makes the point in a less abstract way: the more the merrier. I built it to show why fifteen a year beats twenty-five in total, and that is the version in the video. Fifteen years at fifteen a year, which is 225 investments. But 225 is not the model’s answer. It is my strategy, drawn. Feed it twenty a year and it draws a bigger number and prefers it. The model has no view on where to stop, because there is no stopping point in it to find. All it will tell you is which of two strategies is better, and the answer is always the one with more independent draws in it.

Fifteen a year has been the working cadence for the last five or six years. The book is at ninety as of June 2026.

The constraint I’ve kept

Here is where I must be honest about my own number, because fifteen was never really a decision.

Brisbane Angels runs ten pitch nights a year, three companies each. That is thirty pitches, and the thirty is a format constraint rather than a demand constraint. There’s just no room for a fourth on the night. Of those thirty, roughly half get funded, which is a function of how much money is in the room. I invest in most of what gets funded. So, my cadence is not a policy I set and defend. It is Brisbane Angels’ throughput, and I am downstream of it.

I don’t invest outside the group, and the reason is the infrastructure itself. MooCoo’s bare trust is what turns ninety positions into one annual report, and it works because the paper sits in one place. Invest direct and I am back to my own cap table, corporate actions, and paperwork. The ninety-first investment stops being a parameter and goes back to being a headache. That is the trade. The admin that makes a large book possible is the same admin that decides where the book can come from.

It is a trade I made knowingly and it is currently worth it. But by the logic of everything above, it is a constraint waiting to be engineered away rather than a law of nature. The maths wants more independent draws. I have bound my draw rate to a single source.

Why I’m optimistic anyway

The near-term answer is that the source is getting better.

Brisbane Angels has a twenty-year history, which means flow arrives without anyone having to chase it. MooCoo runs the group and project-manages that flow, and a good deal of what looks like sourcing is really triage, advising a founder whether they are a fit for the funnel or not. The thirty slots stay at thirty. What rises is the quality of what fills them, because a group that reliably funds deals attracts founders who bring better ones. Flow begets flow, which is the flywheel I set out in Angelmatic for beginners.

Better companies pull more money out of the room on the night. And the room itself grows: a group with a track record attracts members, so the pool rises alongside the quality. So more of the thirty clear, and my count rises without me forcing anything. Fifteen becomes eighteen becomes twenty. Some will always miss out, because thirty is thirty (ten nights, three slots, no room for a fourth). That is a calendar constraint, and it is the one thing the flywheel can’t spin away.

The longer answer is that MooCoo starts sourcing outside Brisbane Angels, which is not for a while. Thirty is a good number, but it’s not a bound one.

The question that’s coming

Which leaves me with a problem I don’t have yet.

Today my cadence and the room’s throughput are the same fifteen, so nothing is being turned away. But the flywheel is doing its work, and at some point the room funds twenty. Do I take twenty?

By my own argument I have to. If the funnel produces twenty deals that clear the bar and I take fifteen, I am rejecting five deals I have just declared investible, on some basis other than the filter, because the filter has already spoken. On what basis, then? I am ranking inside the investible set. Which is picking winners, and picking winners is the one thing this entire argument says cannot be done.

My framework is a falsification test. It asks whether I can articulate why a deal will fail, and if I can’t, the deal is in. It is binary by design. It has no opinion about which of two investible deals is better, and if I force it to have one I have stopped using a filter and started using a hunch.

A cadence cap set below the quality of your own funnel is not discipline. It is stock-picking in discipline’s clothes. Fifteen was the right number when fifteen was what existed. It won’t be for long, so it moves to twenty, then twenty-five, tracking the funnel rather than my nerve.

Which is the point at which this stops being a maths question.

Follow-on or a new deal?

Ask a venture capitalist and they will tell you reserves are non-negotiable. Preserve your ownership in the breakout. Do not let the Series A wash you out.

The evidence is thinner than the conviction. Abe Othman ran the question properly: a universe of seed-stage convertible investments, pools of 40, 10,000 simulated portfolios, testing three strategies: never follow on, always follow on, or double down only when the seed position had at least doubled in value. The result was uncomfortable. The selective doubling-down strategy landed very close to the all-or-nothing approaches, which is the same finding as before in a different costume: reliably identifying the future stars is not a demonstrated skill. Absent perfect foresight, the broadly diversified portfolio with no follow-ons tended to match or beat the alternatives in most scenarios.

So why does the industry believe otherwise?

A constraint dressed as a conviction

Because for a venture capitalist, follow-on genuinely is the right answer. It just isn’t the right answer for the reason they give.

A VC’s cost per investment is close to fixed and it is high. Diligence, legal, the board seat, the quarterly reporting, the line item in the LP letter. AngelList’s own work on cheque sizing concedes the point directly: writing a small cheque takes almost as much time and effort as writing a large one, but a large cheque, if it works, returns much more to the fund. That constraint forces concentration before the return distribution gets a vote. Concentration forces reserves and reserves force follow-on.

Then the business model pays for it twice. A chunk of GP compensation is a management fee charged on funds under management, so capital held in reserve is capital still earning. Deploying it into the marked-up round of a company you already know is the lowest-effort way to put a large cheque to work at a defensible price. A survey of fund managers found a median of 55% of committed capital reserved for follow-on opportunities. Over half the fund, held back, on a strategy whose evidence base is this thin.

And the capital held back is capital not writing five or ten new seed cheques. In a game decided by shots on goal, buying more of a company you already own at a higher price is a poor way to find the one you don’t.

None of that is a racket. The GP is optimising correctly inside their own structure. The point is narrower: their conclusion is downstream of a cost structure and a fee structure, not of the maths, and it does not transfer to anyone who shares neither.

Which is why the Othman result has been so comfortably ignored. The evidence isn’t contested because it’s inconvenient. The people best placed to act on it are the people structurally least able to.

The angel’s exemption, and what it costs

The angel has none of those constraints. No LPs or fee base rewarding capital held rather than deployed, or obligation to fill a board seat. On paper, the angel is the only participant in this asset class free to do what the power law actually says, which is take more shots.

That’s on paper. In practice, the angel has a cost per investment too. It is just paid in their own time instead of a fee line, and it is why most angels stall short of the floor, let alone above it. Every objection to a large angel portfolio is one of these costs, and each one has a specific answer. All but one.

• Admin. Ninety companies is ninety cap tables, ninety sets of corporate actions, ninety annual statements. A bare trust structure collapses that: MooCoo holds legal title, I hold the beneficial interest, and ninety positions report to me as one line, once a year. The AFSL is what makes that compliant at scale rather than improvised.
• Diligence. This is the cost the group destroys. A GP pays for judgement: analysts, consultants, partner hours, all charged to the fund somewhere. Brisbane Angels has fifty people in the room, most of them professionals, just not venture capital professionals, and among them is usually someone who has built the exact thing being pitched. The subject matter experts also screen deals out before the pitch night, so the filter is running before anyone has opened a data room. That is not cheaper institutional diligence. It is an input the institution cannot buy at that price, because the people supplying it aren’t being paid. They are investing alongside you. The corollary is that you need less of the formal kind, which is the argument in Due diligence: enough, but not too much.
• Governance. You don’t manage ninety companies. When a decent VC leads the Series A, they take the board seat, the reporting and the fundraising, and they are contractually obliged to their LPs to drive that company toward an exit. The winners get minded by professionals. You stay out of the way.
• Quality and independence. Volume bought by lowering the bar changes the tail index, and a diluted 1-in-500 at 100x breaks the whole argument. Nothing about scale excuses you from a filter, which is what If I can’t articulate why it will fail is for. The draws also have to be independent, and this is where the group format helps rather than hurts: Brisbane Angels sources across Australia and New Zealand, the room’s interests are broad enough that the deals aren’t clustered in one sector, and the group invests year in and year out, so the vintages spread. Thirty pitches a year from one group is a lot less correlated than it sounds.
• Capital. The one the infrastructure doesn’t solve. Some angels simply run out of money, and a cadence you cannot sustain for fifteen years is not a strategy, it is a burst. Cheque size, cadence, and the size of the pool have to be planned together, which is what the simulation model is for. It is not an argument for more. It is the thing that tells you how much more you can sustainably do.

The group answers two of those at once, which is a strong claim, so let me concede the failure mode. Fifty people in a room is not automatically judgement. It works when the person who has built the thing speaks up. It fails when everyones to whoever sounds most certain, and a room full of accomplished people outside their own domains as readily as any other room. Thirteen years of watching that room and the difference is usually obvious within about ninety seconds.

And time. The GP’s hours go to board seats and LP reporting. Mine don’t, which is a real advantage and one nobody puts in a pitch deck.

That is the machinery: a bare trust, an AFSL, one annual report, a room full of operators, a VC to hand the winners to, a filter that still says no. MooCoo and Brisbane Angels have built it, and Angelmatic is the piece that takes the marginal cost of the next deal to something close to zero.

The floor was never a ceiling

Twenty-five is what the distribution says about the floor. It is a good answer to a real question: how few investments can you make and still have the maths work at all. Below it you are gambling, and Roll the die stands.

It says nothing about the ceiling, because there isn’t one in the model. The angel portfolio size people stop at is a different number entirely, and it comes from somewhere else. It is what the back office says, wearing the floor’s clothes. It is the count an angel can reach before the admin, the diligence and the time defeat them, and once you have arrived there exhausted it is a very short walk to deciding that the place you stopped was the place to stop.

I have watched four of those constraints come off, one at a time, and each time the number moved. The fifth one is still on, and it is mine. It’s a draw rate bound to a single group’s throughput, because that is the price of the annual report I like so much. The maths has no opinion about that. It just keeps saying the same thing, patiently, from underneath the plumbing.

Richard Moore is co-founder of MooCoo Ventures, an angel syndicate that co-invests alongside Brisbane Angels, one of Australia’s most active angel groups. He has made over ninety personal angel investments since 2013.

moocoo.vc