When the Market is a Complex Adaptive System

Amazon peaked at $113 a share in 1999. By late 2001 it had crashed 94%, trading at $5. Anyone who bought at the peak and held to today is up about 45 times their money. Anyone who bought at the bottom and held is up about a thousand times.

Almost nobody actually did either. The drawdown was hard. The recovery was harder. Every multiple along the way — 5x, 10x, 100x — looked like enough.

The dog walks the man

Imagine a man walking a small dog on a long leash. The dog runs in every direction — chasing scents, barking, doubling back, lunging forward. Short-term, the dog's movement looks chaotic and dominant. But over the walk, the dog goes wherever the man walks.

The market short-term is the dog. Long-term, it follows wherever the underlying business walks.

That image is the framework. The rest of the essay is what the picture means once you take it seriously.

What the picture is

My mental model of financial markets is that they are a complex adaptive system. Three things follow.

Nobody controls these systems at any time-scale. Not marketmakers, not the Fed, not the largest institutional flows. Everyone steers within their reach. Nobody controls.

Different agents steer at different time-scales. The time-scale you choose to invest on determines which agents are your competition and which are noise. The next section unpacks the layers.

Outside influence can disrupt any layer at any moment. A central bank announcement, a geopolitical event, a war, a regulatory change, any of a thousand things can override what was steering before. The March 2020 COVID crash collapsed every layer in days; the underlying businesses that compounded out of it were already in motion before the crash and continued after it.

The leash is short for the dog but the walk is long. The dog's energy doesn't change the man's direction. It changes how the walk looks moment to moment.

Different agents steer at different time-scales

In a complex adaptive system, the question isn't who controls the system. The question is which agent has the most influence over which time-scale. In financial markets:

• Marketmakers steer intraday. They set the spread, run the liquidity, dominate the order flow. Algorithms with co-location and data pipelines dominate this layer. Speed and signal-detection are the load-bearing edges at the intraday scale; the infrastructure required to compete on them is structurally available to a small number of well-capitalized firms.
• News and sentiment steer weeks. A surprise earnings beat, a CEO resignation, a regulatory headline can move prices for days or weeks before fundamentals reassert.
• Earnings steer quarters. Cycle-to-cycle, the realized financial performance of the business pulls the price toward its trajectory.
• The underlying business steers decades. Across multiple cycles, multiple narratives, multiple cohorts of marketmakers, the structural position of the business and its compounding dynamics are what's left.

Marketmakers don't actually control even intraday. They steer. Outside influence can break their steering at any moment. Anyone who has watched a marketmaker's quoting strategy get torn apart by a surprise central bank statement understands this viscerally — the steering layer doesn't hold against shocks from outside the system.

Decades belong to the underlying business. Which only matters if the business actually compounds.

Why some businesses survive (and most don't)

Most don't, over multi-decade time-scales. The dot-com era surfaced this with unusual clarity.

Three archetypes from that era are worth distinguishing.

Pure infrastructure builders — Cisco, Nortel, Lucent, Global Crossing, the dark-fiber companies. Peaked together with the bubble; collectively wiped out around $2 trillion in telecom market value. Cisco peaked at ~$500B market cap in March 2000; its stock price didn't reach a new all-time high until December 2025 — twenty-five years later. Nortel went from peak ~C$400B to bankruptcy by 2009. Lucent down ~99%. Global Crossing accumulated $12.4 billion in debt before going under. Williams Communications, 360networks, dozens of others died from leverage plus demand-timing failure. The fiber they built eventually became the cheap bandwidth that enabled YouTube, AWS, and Netflix streaming — but the original builders mostly didn't compound. Infrastructure overbuild produces durable assets for the next era; the builders themselves can still die.

This is the test that distinguishes Cisco from Amazon ex-ante. Cisco sold routers and operating-system software to the overbuilders. It didn't have its own application engine whose growth forced the infrastructure build regardless of bubble demand — its customers were the demand. When the customers died, Cisco's revenue died with them, and the recovery took 25 years. Amazon's retail business kept growing through and after the crash, which forced the AWS build for its own use, which then became a separate business when the spare capacity was monetized. The application engine drove the infrastructure layer, not the other way around.

Pure application plays — Pets.com, Webvan, etoys, the long list. Most died because the TAM was too small to support the cost structure, the execution wasn't there, or both. A few survived (eBay, Priceline) but as smaller compounders.

The third archetype — application companies that built infrastructure for their own use and monetized the spare capacity. Amazon: bookstore → everything-store buildout → AWS (2006) monetizing the spare server capacity the retail business had built. Google: search engine and advertising business (still ~75% of Alphabet's $400B+ revenue) → data centers built for search and ads → Google Cloud, the spare capacity monetized as a separate fastest-growing segment. Both are still compounding twenty-five-plus years after their IPOs because they straddle two layers — application businesses with infrastructure businesses inside them. (Worth noting where the framework doesn't apply: commodity plays where structural insiders genuinely move outcomes, or opaque markets where deterioration can hide for years. It applies to liquid public markets where you can actually research the business you're investing in.)

The third archetype isn't an accident. It's the structural shape that compounds longest across regime changes because it monetizes both layers. When the application thesis matures, the infrastructure business carries the next era. When new infrastructure emerges (AI), these companies have the cash flow, the customer relationships, and the technical capability to build it before the pure-play challengers.

The practical implication: the conviction-from-deep-knowledge has to be deep enough to identify the third-archetype shape before the infrastructure business is publicly recognized as the real value. Amazon was operationally monetizing AWS in 2006. Wall Street didn't fully price it until around 2015. The nine-year window between operational reality and market recognition is where conviction-from-deep-knowledge separates the 100x+ investors from the 5-10x ones.

Which raises the question of what conviction-from-deep-knowledge actually does for you, and how it's tested.

What holding actually requires

The thing that lets you hold through a 50% drawdown isn't a PE ratio. It's deep enough understanding of the specific business that price moves stop registering as signal.

And it isn't only the drawdown that tests this. Holding through success is harder. Selling at 2x because the price doubled feels prudent. Selling at 10x because that's where most professionals stop feels disciplined. Each exit forecloses the next decade of compounding. Without conviction in the underlying business, you sell either when the intraday agents are loud with fear at the bottom or when they're loud with "lock in the gains" at every multiple along the way. The dog convinces you it's leading; you stop watching where the man walks.

Amazon, in full

The dot-com crash that produced Amazon's 94% drawdown wasn't a normal cycle. It was the worst extended bear market in technology stocks since 1929. The NASDAQ took fifteen years to recover its 2000 peak. The 2008 financial crisis, the 2020 COVID crash, and the 2022 tech correction were all shallower and shorter. Individual high-flyers like Amazon, Cisco, and Sun Microsystems lost 80-95% of their value and stayed underwater for the better part of a decade. If the framework survives that, it survives ordinary cyclical drawdowns by a wide margin.

Most people who bought near the top sold near the bottom, which is what fear does in a complex adaptive system when the loudest steering layers — sentiment, narrative, marketmakers — are all moving in the same direction. Amazon launched AWS in 2006 — the strategic inflection that retrospectively defines the business. The stock didn't recover to its 1999 high until late 2009. About a ten-year arc, peak to pre-crash high.

That's the part of the story most retellings stop at. It's also the part where the harder lesson begins.

By 2010, the peak-buyer was up 50% — back above water, but still nowhere near the upside the bottom-buyer had already realized. By 2012, finally doubled. By 2015, six times the original investment. By 2017, ten times. By 2020, thirty times. As of mid-2026, about forty-five times. Each multiple was a reason to exit by any standard heuristic — book the win, take some off the table, no one ever went broke taking profits. Each exit would have foreclosed everything that came after.

The bottom-buyer's trajectory is more dramatic still. From $5 in 2001 to today's roughly $5,000 (pre-2022-split-equivalent — Amazon split 20-for-1 in June 2022; today's ~$250 share is the post-split version of $5,000 pre-split), they're up about a thousand times. By 2003 they were at 10x. By 2010, 22x. By 2015, roughly 130x. By 2020, 600x. Each of those milestones looked, at the time, like a complete win. Anyone with reasonable discipline would have taken profits. Almost everyone did.

The investor who got the 1,000x outcome from the 2001 bottom — or the 45x from the 1999 peak — didn't just survive the drawdown. They overrode "I've already won, lock it in" repeatedly across the next twenty-five years, through multiple sub-cycles where the price doubled and tripled and then halved again. They held because their conviction was in the decade-scale force — the underlying business — not in the price moves. The price was always going to look like enough at some multiple along the way. The business was always going to be the steering force regardless. The dog ran in every direction for twenty-five years; the man kept walking.

The recovery isn't the payoff. The recovery is one stage. The payoff is the decades of compounding that happen on both sides of the recovery — for the bottom-buyer, mostly during the drawdown and early recovery; for the peak-buyer, mostly after the recovery completes. In both cases, the discipline that produces the 100x+ outcomes is the same: holding through the moments where the loudest steering layers are screaming "this is enough."

The cost of the alternative

The discipline this framework requires costs less than the alternatives. That's the part that's easy to miss when you're sitting at a 2x and the price has finally moved and the relief feels like a signal.

Selling at 2x looks prudent until you see what it would have become. Selling into the 50% drawdown looks like risk management until you see what came back. Re-entering at the inflection point requires recognizing the inflection in advance, which marketmaker noise reliably obscures, which is why almost nobody actually re-enters at the right multiple — they wait for confirmation, and confirmation comes after another 5x.

The investor who holds through both directions isn't braver. They've concluded that the alternatives — re-entering at every multiple, timing the recovery, recognizing the next-era inflection before Wall Street prices it — are structurally more expensive than staying aligned with the decade-scale force.

The decision rule reduces to this: identify a third-archetype business where the application engine forces the infrastructure build regardless of cycle demand. Commit at a position size you can hold to zero without forced selling. Then treat every subsequent signal that isn't a structural-business change as the dog, not the man.