The Brilliant People With Broken Compasses

Brilliant People, Persistent Underperformance

The people running active funds are, by any reasonable measure, extraordinarily smart. Their CVs are impeccable. Their analytical frameworks are rigorous. Their models are precise to four decimal places.

And yet, as a group, they persistently underperform.

This isn’t a new observation. The data on active management has been accumulating for decades, and the verdict is not particularly kind. But the standard explanations — fees are too high, too many funds chasing too few alpha opportunities, the information edge has evaporated — while largely correct, feel incomplete. They describe what’s happening without fully explaining why it keeps happening even among the genuinely gifted.

There’s a third explanation that doesn’t get enough airtime. And it has nothing to do with intelligence, work ethic, or even fees. It has to do with the fundamental mismatch between the tools most portfolio managers were trained to use — and the nature of the businesses that are actually generating the most value in the modern economy.

The Map That Doesn’t Fit The Territory

The bedrock analytical tool of the investment profession is the discounted cash flow model. In its essence, a DCF is a forward projection of earnings, discounted back to a present value using some assumed cost of capital. It is a machine built for linearity. You plug in a revenue growth rate, you extrapolate, you discount, you arrive at a number.

For a business that grows at 8% a year and compounds steadily, this works beautifully. The model fits the business like a glove. A trained analyst with a good model and good judgment can do valuable work here.

But here’s the problem: the businesses generating extraordinary returns over the last two decades — and arguably the next two — don’t grow linearly. They grow exponentially, or not at all, and then exponentially. They follow what Alex Sacerdote of Whale Rock Capital calls an “S-curve” — decades of slow adoption followed by a near-vertical acceleration, then a plateau. The skill isn’t just building the model. The skill is knowing exactly where on that curve you are.

A quick aside on Sacerdote, and on Kai Wu of Sparkline Capital, who shows up later in this piece: neither name will be familiar to most readers, and that’s rather the point. Mainstream financial media runs on a fairly shallow rotation of the same handful of sell-side strategists, recycling the same soundbites in seven-minute segments. The long-form podcast world is a different animal entirely — specialist managers given two uninterrupted hours to walk through the actual research, the actual data, the actual reasoning behind a position. Most of what’s genuinely useful in this piece came from that world, not from anything you’d hear on the news.

A DCF model applied at the wrong point in an S-curve will produce consistently bad answers. Applied too early in the growth phase, it will say a business is wildly overvalued. Applied too late, it will anchor to a trajectory that’s already decelerating. The model isn’t wrong, exactly — it’s just solving the wrong problem.

The CFA Curriculum Doesn’t Teach This

Let me be precise here, because it’s easy to strawman this argument.

A CFA qualification is genuinely impressive. The curriculum is deep, rigorous, and demanding. It produces analysts who understand financial statements, valuation methodologies, portfolio theory, and risk management at a sophisticated level. Nobody serious disputes this.

What it doesn’t produce — and probably can’t, by design — is an investor with an intuitive feel for how technology adoption curves work. Or one who can look at a platform business and immediately understand that its unit economics improve as it scales, rather than deteriorating. Or one who grasps that a company like Amazon building out AWS in 2012 wasn’t allocating capital inefficiently — it was laying the infrastructure for a fundamentally different and far larger business that the income statement couldn’t see yet.

This isn’t a CFA problem per se. It’s a framework problem. Financial modelling is fundamentally backward-looking. It anchors to what the business has done. Forecasting what a platform company will do requires a different kind of reasoning: pattern recognition, technological intuition, and what you might call structural imagination — the ability to envision what a market could look like once a new capability is fully absorbed.

Sacerdote makes this point sharply. The market, he argues, has a “linear bias.” When unit growth is exponential and the business model is highly scalable, long-term earnings power is systematically underestimated by consensus. Not because analysts are lazy. Because their tools are built for a different kind of business.

The Dishwasher and The Browser

Here’s a distinction that crystallises a lot of this quite elegantly.

Traditional enterprise software adoption moves like a dishwasher. It takes years — sometimes decades. It requires infrastructure integration, legacy system migration, procurement cycles, change management programmes, and an entire ecosystem of implementation consultants to grease the wheels. A CIO doesn’t just flip a switch. They run a multi-year project.

This is a world that financial models handle well. You can model the sales cycle, the implementation lag, the renewal economics. The numbers have texture. The timeline is predictable. A skilled analyst can build a credible five-year revenue bridge.

AI behaves like opening a browser tab. A company can go from zero to a thousand knowledge workers using an AI tool in a week. There’s no infrastructure integration, no implementation team, no lengthy procurement cycle. The adoption curve doesn’t slope gently upward — it goes vertical. Or, as Sacerdote puts it, it looks less like an S-curve and more like an “L-curve.”

The implication for investors is profound. The standard playbook for sizing a technology investment — analysing the current market, projecting a penetration rate, modelling the financials — was built for dishwashers. It doesn’t translate cleanly to browsers. When adoption can go from 0.1% of knowledge workers to 15% in four years, the model isn’t off by a little. It’s off by an order of magnitude.

It’s Not Just About Tech

At this point, the fund manager reading this (there are a few of you, presumably) will want to push back: “We hire technology analysts. We do primary research. We visit companies. We’re not just staring at spreadsheets.”

Fair. And I don’t want to caricature an entire industry. There are genuinely great active managers who do understand technology — Sacerdote being a prime example.

But here’s the more pointed challenge: how many active managers truly understood, in 2016, that Amazon Web Services was going to become the most important capital allocation decision in corporate history? How many understood, in 2018, that Microsoft’s cloud transition under Satya Nadella wasn’t just a product refresh but a fundamental restructuring of the company’s economics? How many recognised, as recently as 2022, that Nvidia’s market position wasn’t just about gaming GPUs but about the physical constraints of running AI workloads at scale?

The issue isn’t just knowledge of technology. It’s the willingness and ability to hold a conviction that looks wrong — sometimes for years — because your model says the valuation is stretched, while your qualitative judgment says the business is compounding value that the model can’t see yet. That requires a different temperament, and a different framework, than the one most professional investors are trained toward.

The New Moats Are Harder To Model

Sacerdote’s investment framework — what he calls the “three-legged stool” — is useful here. You need to understand the S-curve (where are we in adoption?), the moat (what structural barriers prevent competition?), and the underappreciated earnings power (what is the market missing?).

The first two legs have always been part of good investing. But the nature of moats has changed materially. For most of the 20th century, competitive advantages were things you could see clearly in the financial statements: scale economies, brand premium, switching costs from proprietary infrastructure. These are quantifiable. An analyst can model them.

The moats in modern technology businesses are often different in character. Network effects are hard to model because their value is convex — they’re worth very little below a critical threshold and disproportionately valuable above it. The value of a foundation model doesn’t just compound linearly with compute investment; there appear to be non-linear capability jumps at certain scale thresholds. The capital barriers to competing at the frontier of AI are so enormous that the industry has, in just a few years, consolidated into something resembling an oligopoly — a dynamic that was effectively impossible to predict from financial statements alone.

This consolidation dynamic is particularly underappreciated. When Sacerdote describes moving a model’s benchmark performance from 80% to 85%, he’s not describing a marginal improvement. He’s describing a phase transition — from useful assistant to autonomous agent. That kind of non-linearity is almost impossible to capture in a model. It requires you to think less like a financial analyst and more like, say, a physicist studying phase transitions in complex systems.

Most portfolio managers are not trained physicists. They’re trained accountants with very good memories.

The Bear Case for Conventional Software

One of the more contrarian — and, I think, correct — insights from this framework is the structural problem facing traditional enterprise software companies.

These are businesses that built durable franchises on seat-based licensing, annual price increases, and the stickiness of workflow integration. For years, the “Rule of 40” — where revenue growth rate plus operating margin should exceed 40% — was the benchmark for evaluating their quality.

But the same AI wave that’s creating these extraordinary opportunities in infrastructure is simultaneously putting a fiscal squeeze on the traditional software vendors. Enterprise IT budgets aren’t infinite. If a CIO is reallocating spend toward AI infrastructure — which delivers immediate, measurable ROI — that budget has to come from somewhere. It’s coming from legacy software renewals.

Worse, the seat-based licensing model faces a structural headwind: companies are beginning to automate roles rather than fill them. If your software is priced per knowledge worker, and the number of knowledge workers is declining, the underlying unit of your business model is eroding. This is precisely the kind of structural shift that doesn’t show up clearly in near-term financial results — and therefore gets systematically underpriced by models anchored to recent growth rates.

Sacerdote’s updated framework for evaluating AI infrastructure businesses is instructive: (% of revenue derived from AI) + (market share within that AI category). Score above 60 and you have a highly efficient vehicle for capturing structural growth. It’s a framework built for the current regime — not the one that produced the fund industry’s existing analytical toolkit.

The Dark Matter on the Balance Sheet

Everything above is really an argument about growth being hard to model. There’s a second, quieter problem that’s arguably more damning, because it doesn’t even require a fast-growing business to bite. It’s an argument about whether the inputs to the model are measuring the right thing at all.

Kai Wu, founder of Sparkline Capital, has built a research practice around this exact question. His starting observation is almost embarrassingly simple: accounting rules treat tangible and intangible investment completely differently. When an industrial company builds a factory, that spend is capitalised — it sits on the balance sheet as an asset, depreciating over time. When a technology company spends a similar amount on R&D, software, or brand-building, accounting rules force it to expense the whole amount immediately, against current-year earnings.

This isn’t a rounding error anymore. Wu’s research puts intangible capital at roughly 42% of the total capital stock of US public companies today, up from a negligible share in the 1980s. And because standard value metrics — price-to-book, price-to-earnings — are built on a foundation that simply doesn’t recognise most of that capital, value screens are mechanically excluding some of the most valuable businesses in the index before any human judgement is even applied. The screen isn’t missing the opportunity through poor analysis. It’s missing it by construction.

Wu’s firm tries to repair this by quantifying what the balance sheet won’t: patent filings and citations as a proxy for R&D quality, brand sentiment and pricing power, hiring velocity and skill density as a measure of human capital, and network graphs as a measure of platform stickiness. Whether or not you buy the specific methodology, the underlying point stands on its own: book value and trailing earnings are decreasingly reliable proxies for the asset base of a modern company. An analyst running a traditional value screen isn’t just missing a curve. They’re working from a balance sheet that’s already lying to them.

There’s a nice irony buried in Wu’s research on this point. Warren Buffett — the patron saint of value investing, the man every “value hardliner” claims to be channelling — only bought below book value in roughly 8% of Berkshire’s historical investments. Buffett’s actual career was a long evolution away from Ben Graham’s cigar-butt tangible value, through Coca-Cola’s brand intangibles, and ultimately into Apple — a position built almost entirely on ecosystem lock-in and brand equity, neither of which shows up meaningfully on a balance sheet. The investor most often invoked to justify ignoring intangibles spent the second half of his career doing the opposite.

“It’s Cheap” Is Not An Answer

Here I’ll step out from behind the “we” for a moment, because this part comes from sitting in rooms and listening to calls for twenty-odd years, not from a podcast.

If you spend enough time around South African asset management commentary, you’ll hear a particular phrase more often than almost anything else: “it’s cheap.” Sometimes “it’s cheap relative to history.” Occasionally, if you’re lucky, “it’s cheap relative to peers.” Almost never do you hear the question that should come right after: cheap relative to what, exactly — relative to its growth rate, its capital intensity, the size of the market it’s chasing?

A price-to-earnings multiple of 40 can be expensive. It can also be startlingly cheap, if the underlying business is compounding earnings at 50% a year with margins still expanding. A multiple of 8 can be a bargain, or it can be a business in terminal decline, where the multiple is low precisely because the market has correctly priced in a shrinking earnings base. The number on its own tells you almost nothing. Strip away the growth, the quality, the capital intensity, and “cheap” is just a word, not an analysis.

This is the same failure as Wu’s intangibles point, just observed from a different angle. A suppressed earnings number — because a company is expensing growth investment that arguably belongs on a balance sheet — makes the P/E look elevated and the business look “expensive” by an analyst whose own R&D and brand-building are doing exactly the same thing to its denominator. Self-described value investors who lean hardest on the Buffett brand are frequently the ones most reluctant to ask whether their numerator-over-denominator shortcut still measures what they think it measures. Buffett, as it turns out, mostly stopped relying on it decades ago.