He Quantified 200 Years of Disruption | Kai Wu on Separating Software Survivors from Value Traps

ELI5 / TLDR

Cheap-looking stocks are sometimes cheap for a good reason: their business is quietly dying. Kai Wu fed two centuries of US patent records into a machine and built a system that spots, in real time, which industries are about to get run over by new technology. The punchline: classic “buy cheap stocks” investing has stopped working for the last fifteen years, but only inside the industries facing disruption. The fix is to stop measuring a company by its accounting book value and start measuring its hidden assets — brand, talent, network effects, patents — because those, not the cheapness, are what separate the survivors from the roadkill.

The Full Story

The trap that looks like a bargain

Start with a simple, painful pattern. A new technology shows up. Investors see it coming, panic, and sell down the incumbent’s stock fast. But the incumbent’s actual sales and profits keep chugging along for years afterward, falling only slowly. That gap — price collapsing fast, fundamentals collapsing slow — creates a window where the stock looks like a screaming bargain. Low price relative to earnings, sales, book value. A traditional value investor sees a no-brainer buy.

It is the opposite of a bargain. Wu calls it a value trap, and he illustrates with four corpses: Blockbuster, Borders, RadioShack, and a newspaper conglomerate, each killed by Netflix, Amazon, and Google.

It’s sucking you in, bringing you on board a ship just as it’s about to collapse and sink into the sea.

The cruel part is structural. Because stock prices look forward and fundamentals report the past, you will almost always get a stretch where the price-to-sales ratio looks delicious right before the wheels come off.

Two centuries of patents, clustered by machine

Four anecdotes prove nothing — they are cherry-picked. Wu wanted to know if this was a systemic problem, so he built a measurement system in two steps.

Step one: take every US patent ever filed. The dataset goes back to 1790; the first patent was signed by George Washington. Feed them to software that groups similar patents into clusters — think of it as sorting two centuries of inventions into bins of “this is electricity stuff,” “this is internet stuff,” “this is AI stuff” — with no human telling it what the bins should be. Then watch which clusters are trending (growing fast) and pervasive (spreading across many industries, not just one corner). Pervasive matters because the biggest disruptions are what economists call general-purpose technologies — electricity, the internet, now AI — things that touch everything rather than one niche.

Step two: figure out which companies are in the blast radius. The system reads earnings call transcripts, company filings, analyst notes, and patents to score how exposed each firm is to each technology, then rolls those scores up to the industry level (any single company is noisy; an industry average is steadier). The whole thing is automated. As Wu puts it, you could dust off the same code in twenty years and it would surface a totally new set of technologies and threatened companies on its own.

This let him map seven major disruptive waves — internet infrastructure, e-commerce, digital media, social media, AI, and others — and see which sectors sat in the crosshairs at each moment.

Disruption stacks

The key insight from watching retail get pummeled over decades: the waves don’t replace each other, they pile on. First you survive e-commerce. Then digital media. Then social media. Then agentic AI shopping. A retailer today isn’t fighting one threat; it’s fighting all of them at once — a multi-front war. Sum up a sector’s exposure across every wave and the line climbs steadily upward as innovation accelerates and each technology builds on the last. (You can’t have AI without big data, which needed the internet, which needed electricity, which needed fire.)

This is why the mall names — Williams-Sonoma, Abercrombie, Claire’s, Hot Topic — have sat in value screens forever and never left. They look perpetually cheap because they are perpetually getting stacked.

Why “value investing is dead” — and why it isn’t

Buying cheap stocks worked for decades after Ben Graham codified it in the 1930s. Then around 2010 the strategy went into a drawdown it has never recovered from. Cue the obituaries.

Wu’s key experiment: don’t apply the cheap-stocks strategy to the whole market at once. Split the market in two — industries exposed to technological disruption versus industries insulated from it — and run the strategy separately on each.

When you apply the value factor in the insulated sectors, actually the performance has been just fine. You almost see no difference between 2010 on and the beginning period.

Value investing never died. In sleepy, insulated industries — real estate, asset-heavy businesses — buying cheap still works exactly as it always did. It only broke in the exposed industries, and broke so badly that the losses there swamp the gains everywhere else, dragging the whole strategy negative. The baseline gap between exposed and insulated was negative seven percentage points a year. He then beat the finding with a stick — global stocks, emerging markets, the textbook price-to-book measure alone, sector-neutral versions, the popular “only buy cheap and profitable” double-sort — and it survived every test.

The hiding spot is vanishing, too. The slice of US market value sitting in technologically exposed industries has climbed from roughly 40% to about 75% over twenty years. Warren Buffett’s old move — stay inside your circle of competence, skip the tech you don’t understand — works fine until tech is the entire market, at which point staying out means sitting in cash.

What the survivors did

So how do you tell a future Walmart from a future Blockbuster? Wu brings in Walmart and the New York Times — a retailer and a newspaper, two of the most disrupted industries alive — both of which survived and thrived. They did two things. First, eventually, they leaned into the very technology disrupting them (Walmart now runs one of the largest e-commerce operations going). Second, they leaned on assets that had nothing to do with that technology: brand, people, network effects.

Here he reaches for a 1986 paper by economist David Teece, Profiting from Technological Innovation. Teece’s timeless point: the firm that ends up capturing the value from an innovation is frequently not the one that invented it, but the one holding the complementary assets that surround it — manufacturing, distribution, customer service, sales channels.

A UK company, EMI, invented the CAT scanner. But selling to hospitals means enterprise sales cycles, training, servicing. GE had those muscles, didn’t have the scanner, and eventually owned the market.
RC Cola invented diet and canned cola. Coca-Cola and Pepsi had the shelf space and the brand. Guess who won.
IBM was late to the PC but captured it in the 1980s — not on technology but on the ecosystem of software and peripherals it built around its standard. We’d call that network effects today.

The lesson for the software selloff: if a company’s only moat is its code, AI may genuinely threaten it. If the moat is the complementary assets around the code, it may be fine. And the early AI winners — OpenAI, Anthropic, Google — being the innovators does not guarantee they capture the profits.

The four intangible pillars

Wu’s framework measures four kinds of hidden, off-balance-sheet asset:

Intellectual property — patents, proprietary data, know-how
Brand equity — customer loyalty and relationships
Human capital — not just talented people but a culturally aligned workforce
Network effects — an ecosystem of outside producers and consumers (Uber, a stock exchange)

He builds a score for each using cheap, observable proxies — patents and R&D spend per dollar of price for IP, trademarks and social media for brand, job postings and employee profiles for human capital — then combines them into one composite, the same yield-style way you’d build a traditional value score, across roughly 5,000 global companies. Teece’s framework and this one nest neatly: the “focal innovation” Teece talks about is just a slice of the IP pillar, and the complementary assets are the other three.

Swap traditional value for this intangible value score and the broken half of the market heals. The strategy now works in exposed industries too, and works consistently — across time periods and across exposed versus insulated. It becomes, in his word, all-weather, rather than something you have to time correctly.

Seeing it in the data

A 2007-to-2017 scatter plot makes it visual. Plot every large company by traditional cheapness on one axis and intangible cheapness on the other, then color the future winners blue and losers red. The blue dots cluster in the top half — cheap on intangible value predicted the next decade’s winners regardless of traditional score. The disagreements are the tell: stocks cheap on tradition but pricey on intangibles (Macy’s, Wells Fargo) were losers; stocks pricey on tradition but cheap on intangibles (Amazon, Apple) were winners. Traditional value sold the Amazons and bought the Macy’s.

Over the full sample: stocks both metrics call cheap returned +4.2% a year; both call expensive, −5.1%; the “expensive disruptors” intangible value rescued (Apple, Amazon) did +2.8%; the value traps (Macy’s) did −1.6%. When both agree, the signal is strongest — the two metrics are complements, and the real red flag is a stock expensive on both.

Applying it to the software wreck

Software has sold off roughly 30% as an index while the market rose, but with enormous dispersion underneath — names down 50 to 80%. Wu plots the intangible-value scores of the software losers (down 30%+ in a year). The average score is positive, suggesting the group was oversold on a shoot-first-ask-later basis. But there’s a real left tail — companies down 80% that are still expensive on intangibles. Genuine traps.

Two ways for a software company to be okay: own a moat, or genuinely embrace AI. On adoption, software is the standout — the most exposed sector and by far the most aggressively adopting, the highest hirer of AI talent. And for the survivors there’s an upside twist: their biggest cost is producing code, and AI that cuts the labor intensity of coding could fatten margins and ease the stock-based-compensation fight investors have started picking. Survival is the big “if,” but conditional on it, AI may be a boon.

Dispersion as an amplifier

The closing idea is the most elegant. When the market decides a sector is disrupted, returns over the next year don’t have a worse average than normal stocks — the median is the same six or seven percent. What changes is the spread. Normal stocks follow a roughly bell-shaped distribution; disruption-scare stocks have fat tails. Ten percent of them double over the next year (versus 3% of all stocks); 16% lose more than half (versus 7%).

When technology comes around it shuffles the deck and everything’s in play.

This is the grin in Grinold’s “fundamental law”: dispersion amplifies any edge you have. If you’re any good at separating winners from losers, high dispersion lets that skill produce bigger gains and bigger gains on shorts — the same reason people prize venture capital. So for a genuine software stock-picker, this is plausibly the best opportunity set of their career: a real edge meeting a once-in-a-career amount of dispersion.

Key Takeaways

A value trap exists because stock prices fall faster than the lagging fundamentals they reflect, guaranteeing a window where a dying business looks statistically cheap.
Wu clusters every US patent since 1790 with unsupervised machine learning to detect trending and pervasive technologies, then scores company and industry exposure from filings, transcripts, patents, and analyst notes — fully automated.
Disruption stacks rather than rotates: a retailer today fights e-commerce and digital media and social media and AI simultaneously.
Traditional value investing didn’t die in 2010 — it kept working in insulated industries and broke only in disrupted ones, where losses (≈ −7 pts/year) overwhelmed the rest. The finding survives global, EM, price-to-book-only, sector-neutral, and profitability double-sort robustness checks.
The share of US market value in technologically exposed industries rose from ~40% to ~75% in 20 years, so the value investor’s “hide in safe sectors” move is running out of room.
David Teece (1986): the firm that captures an innovation’s value is often not the inventor but the holder of complementary assets — distribution, sales, service, ecosystem (GE vs. EMI’s CAT scanner; Coke/Pepsi vs. RC Cola; IBM’s PC ecosystem).
Intangible value scores four hidden pillars — IP, brand, human capital, network effects — using observable proxies (R&D and patents per price, trademarks, job postings), combined into one composite.
Swapping traditional value for intangible value makes the factor work in exposed industries too, and consistently across time — “all-weather” instead of needing to be timed.
2007–2017 evidence: cheap-on-intangibles predicted winners; the diagnostic cases are Amazon/Apple (pricey on tradition, cheap on intangibles → winners) vs. Macy’s/Wells Fargo (cheap on tradition, pricey on intangibles → losers). When both metrics agree, the signal is strongest.
Software losers show a positive average intangible score (oversold) but a real left tail of genuine traps still expensive despite 80% drawdowns. Enterprise-facing software has wider moats than consumer-facing (GoDaddy, Duolingo) because switching costs and compliance lock-in are higher.
Disruption-scare stocks have the same median forward return as the market but far fatter tails (10% double vs. 3%; 16% halve vs. 7%) — being down on price alone carries almost no information about expected return.
Dispersion amplifies any genuine edge (the “fundamental law” intuition behind VC’s appeal), making a high-dispersion disruption an exceptional environment for a skilled stock-picker.
One-liner from Wu: “The code is not the moat.” For most software firms, complementary intangibles — not cheapness, not even AI adoption alone — best predict who survives a paradigm shift.
Aside worth noting: Wu built the robustness experiments by delegating to Claude Code — generate a baseline script, then have the model re-run it across geographies and sub-industries.

Claude’s Take

This is unusually good for a podcast where someone walks through their own marketing deck. Wu is selling an intangible-value ETF, so the framework conveniently concludes that his framework is the answer — keep that in the back of your mind. But the central claim is genuinely strong and, crucially, falsifiable: value investing didn’t fail uniformly, it failed only in disrupted industries, and the split-the-market test plus the robustness battery (global, EM, price-to-book, sector-neutral, double-sorts) is exactly the discipline you’d want before believing it. That’s more rigor than most “value is dead” takes ever attempt.

The Teece connection is the part that elevates it from a quant curiosity to something you can actually reason with. “The inventor rarely captures the value; the holder of complementary assets does” is a forty-year-old idea that explains both the historical cases and the current AI-versus-software question without any hand-waving. And the dispersion-amplifies-edge point is the most quietly useful thing in the hour — it reframes a scary selloff as a stock-picker’s opportunity without pretending the average outcome is good.

Where to stay skeptical: “intangible value” is partly a relabeling of quality-and-growth, and the line between “I found a durable moat” and “I overfit to the stocks that happened to win” is thin — the 2007–2017 winners are visible with hindsight in a way the live signal may not be. The argument that survivors get fatter margins from AI is the most hand-wavy bit, resting on a survival “if” he acknowledges but glides past. Score is 8: conceptually rich, methodologically honest about its limits, and it gives you a reusable mental model rather than a stock tip. Docked a couple points for the inevitable book-talking-its-own-product framing and a few claims (the margin upside) that lean on narrative more than the data shown.