The power law debate: concentration, access, and what changed with AI

The case that concentration has intensified

The sharpest current version of the concentration argument comes from CF Private Equity, whose analysis of United States venture-backed exits finds the top one percent of companies accounting for 80 percent of total exit value since 2023, against 17 percent in 2005 to 2010 and 34 percent in 2017 to 2022. The same sentence carries the figure that matters most for reading it: excluding a single event, SpaceX's listing at 1.77 trillion dollars, the 80 percent becomes 45 percent. A statistic that halves when one company is removed is describing something real about concentration and also telling you how much of the measurement rests on one outcome.

Mayfield reaches the same conclusion from the other side of the table, describing the current period as an AI power law era in which liquidity is flowing almost entirely to category leaders, with a bifurcating market where median and average deal sizes have come apart. These are two interpretations of the market, rather than independent tests of a particular return distribution.

Both accounts suggest that rapid scaling in AI can widen the distance between leading companies and the rest. That is an interpretation of the pattern, not a causal result established by the concentration figures alone.

How you would actually know

The claim that venture returns follow a power law is a statistical hypothesis, and there is an established procedure for testing one. Clauset, Shalizi and Newman set out why the usual practitioner method does not work. Plotting outcomes on logarithmic axes, observing a roughly straight line, and fitting a slope by least squares is unreliable at every step: on synthetic data with a known exponent of 2.5 and ten thousand observations, a least squares fit to a logarithmically binned histogram returns 1.19, and the reported standard error gives no warning that anything is wrong. Straightness on those axes is necessary for a power law but nowhere near sufficient, and they demonstrate it by drawing samples from a power law, a lognormal, and an exponential distribution and showing that all three look straight.

Their procedure estimates the distribution, tests how well it fits, and compares it with named alternatives. The goodness-of-fit p-value asks how unusual the observed discrepancy would be under the fitted model and test procedure; it is not the probability that the model is true. Applied to 24 datasets that the published literature had claimed were power laws, seven were ruled out outright, and after comparison against alternatives only one was fully convincing. The alternative that survives almost everywhere is the lognormal, which they could not rule out for any dataset but one.

That alternative is the one this argument has to answer, because Newman's own worked example of a process that produces a lognormal is investment returns. A quantity that is the product of many random multiplicative factors is lognormally distributed, and a lognormal with high variance looks like a power law over any range shorter than about four orders of magnitude. A venture portfolio is a compounding multiplicative process across a few dozen positions spanning perhaps three or four orders of magnitude, which is precisely the regime in which the two are hard to tell apart.

None of this says the concentration is not real. It says that heavy-tailed and power-law-distributed are different claims, and only the first is easy to establish. Cont makes the technical version of the point: even in public markets with hundreds of millions of observations, estimating a finite tail index does not establish a power law, because the tail behaviour is identified only up to an unknown slowly varying function. A General Partner can say with confidence that venture outcomes are extremely concentrated and heavily right-skewed. Saying they follow a power law is a stronger and separately testable claim, and it is one that most published assertions of this kind do not survive.

The affirmative case has its own serious literature and it is worth stating at its strongest. Gabaix documents power laws as a robust regularity across economics and finance, including in stock returns, where the tail exponent sits near 3 on datasets exceeding two hundred million observations. That number is the one to hold onto. A venture fund makes twenty to forty investments, and venture as a whole has no dataset within several orders of magnitude of the public equity record. The evidence that power laws exist in finance is strong. The evidence that any particular venture portfolio follows one is a different and much thinner thing.

What the venture-specific evidence says

One study of private fund returns places venture at the extreme end of its fitted distribution. Lahr, fitting the distribution of valuation multiples across 3,332 private equity funds including 828 venture funds, finds venture tail exponents small enough to imply that the variance of returns over the life of a fund does not converge, and notes that this is a more unsettling property for an investor than it first appears.

Retterath and Kavadias also test venture return distributions directly and report power-law behavior. Their method and sample matter: the analysis uses log-log correlation and R-squared, with between five and twenty-three observations per investor. Those are limitations to keep beside the finding, particularly when the same discussion cites Clauset and colleagues on the weaknesses of that approach.

The genuinely two-sided version

Altos Ventures offers a useful way into the disagreement about what to do with this. On one side, Peter Thiel's rule: the biggest secret in venture capital is that the best investment in a successful fund equals or outperforms the entire rest of the fund combined. One response is to concentrate capital in the companies where conviction is strongest. But identifying the eventual winner in advance is a separate problem.

On the other side, Union Square Ventures, whose stated practice is to invest across a long tail and whose Fred Wilson has described spending much of his time with the companies that will not move the fund, because those founders gave years to something that did not work and deserve better than being written off. That is an ethical position rather than a portfolio one.

The outcome is where the paradox sits. Quartz's 2015 analysis of PitchBook data on United States early-stage rounds since 2000 found that of the 62 companies USV had backed at seed or Series A, five went on to pass a billion dollars, a rate of roughly 8 percent, the highest of the firms it covered. Most firms with at least four such companies came in between 1 and 3 percent, with Sequoia at 5.5 percent on a portfolio five times larger and Y Combinator at 0.7 percent on the largest portfolio in the study. The figure is worth holding loosely, for the same reasons the rest of this page argues for holding such figures loosely. It rests on five companies, it counts a valuation milestone at one date rather than a realised return, its denominator reaches back to 2000 so recent investments had little chance to qualify, and the underlying data excluded companies that reached a billion dollars by going public.

These outcomes illustrate concentration; they do not, by themselves, establish a power-law distribution. They also leave room for different portfolio responses. An investor can expect a few large outcomes while choosing to support a broader set of companies, because the eventual winners are not known at entry.

One figure worth correcting

One widely repeated figure is worth correcting rather than passing on. The claim that ninety percent of venture returns come from a handful of firms is usually credited to Cambridge Associates, and Cambridge Associates published an analysis arguing against it, describing the belief in their own words as “a catchy but unsupported claim.” What they found instead was that an average of 83 different companies a year account for the value creation in the top hundred investments, and that since 1999 most of that value has come from deals outside the top ten. That is a named research house arguing that venture is less concentrated than the industry believes, which is worth more to a reader than another citation agreeing with the consensus.

The two findings are not in tension, because they are about different populations. Value creation concentrates sharply among companies, which is what the exit-value data measures. It does not concentrate nearly as sharply among firms, which is what Cambridge Associates measured. Treating an established fact about companies as though it were a fact about managers is the single most common error in this debate, and it is the one that quietly justifies allocating only to a handful of brand-name funds.

The counterweight worth taking seriously

The VC Factory extends the power law to funds themselves, finding in an analysis of 11,350 startups backed by 259 funds between 1986 and 2018 that about half lost money, that 121 companies, roughly 1.1 percent, returned an entire fund on their own, and that 90 percent of funds returning at least 3x had one of them. In that sample, a fund-returning company was common among funds reaching 3x. The finding is not a universal requirement for outperformance.

Its warning is about what mega-fund scale is actually selling. Because the number of companies producing outlier outcomes is limited and does not increase with the capital chasing them, a very large fund is not positioned to generate power-law returns. What it offers an institutional LP writing hundred-million-dollar cheques is visibility, controlled liquidity, and stable returns, which is a legitimate product and a different one. Power-law language borrowed to justify scale is describing a strategy that scale makes harder rather than easier.

Where this leaves a General Partner

Venture outcomes are often highly concentrated. The exact distribution, and what it implies for a particular fund, require more care. Concentrated and broader portfolios can both have a coherent rationale; the evidence here does not establish one construction rule for every fund.

What the Colibrí Architecture model contributes here is deliberately modest. It has no view on the right level of concentration and does not push a firm toward either pole. What it reads is whether the answer a firm has chosen is carried consistently through the rest of its configuration: whether a concentrated strategy is matched by the ownership and cheque size concentration requires, and whether a broad strategy is matched by the team capacity breadth demands. Both positions are defensible. Declaring one and building the other is the thing worth catching.

Sources

• AI is Steepening the Power Law in Venture CapitalAaron Miller and Joe DeCicco, CF Private Equity, 2026The exit-value concentration series, with and without the SpaceX outlier.
• Q1 2026 VC Outlook: The AI Power Law EraMayfieldReaches the same conclusion from a general partner's vantage rather than an allocator's.
• Power-Law Distributions in Empirical DataAaron Clauset, Cosma Rohilla Shalizi, and M. E. J. Newman, SIAM Review 51(4), 2009The standard procedure for testing a power-law claim, and what happened when it was applied to 24 published claims.
• Power-Law Distributions in Empirical Data (open preprint)arXiv:0706.1062, 2009Open-access version carrying the estimation comparison and the goodness-of-fit results quoted on this page.
• Power laws, Pareto distributions and Zipf's lawM. E. J. Newman, Contemporary Physics 46(5), 2005Why compounding investment returns produce a lognormal distribution that is difficult to distinguish from a power law.
• Power Laws in Economics and FinanceXavier Gabaix, Annual Review of Economics 1, 2009The affirmative case that power laws are a robust regularity in economics and finance, with the public equity evidence behind it.
• Empirical properties of asset returns: stylized facts and statistical issuesRama Cont, Quantitative Finance 1(2), 2001Establishes that a heavy tail and a power law are different claims, and that the second is far harder to demonstrate.
• Fat tails in private equity fund returns: The smooth double Pareto distributionHenry Lahr, International Review of Financial Analysis 86, 2023Fits the distribution of returns across 3,332 private equity funds and finds venture at the extreme of the tail.
• How to Hit Home Runs: Portfolio Strategies and Returns in Formal and Informal Venture CapitalAndre Retterath and Stylianos Kavadias, working paper, 2020A study of venture return distributions and portfolio strategy; the article discusses the limits of its fitting method and sample.
• Paradox of the Power Law in Venture CapitalAltos VenturesFrames the disagreement as Thiel's concentration rule against Union Square Ventures' long-tail practice.
• The venture capital firms that are best at spotting unicornsAlice Truong, Quartz, 2015The primary source for the early-stage unicorn rates on this page, including the 1 to 3 percent band most firms fall in.
• Venture Capital Disrupts Itself: Breaking the Concentration CurseTheresa Sorrentino Hajer, Cambridge Associates, 2015Argues against concentration orthodoxy at the manager level, which is a different question from concentration among companies.
• LPs Beware: The Super Power Law in Venture CapitalThe VC FactoryExtends the power law from companies to funds, and questions what mega-fund scale is actually selling an LP.