The 2023-2030 window

February 2000. In the same week, two people make a purchase, each of them utterly convinced that “the internet is the future.”

The first buys Pets.com stock at the IPO — a company with a brilliant mascot, an instantly recognizable ad campaign, and a business model built around losing money on every shipment of pet food. Nine months later, the company doesn’t exist. The second buys something far more boring: an HTML textbook — and spends weekends building an online store for a local wholesaler that had been taking orders by fax.

Both of them “believed in the internet.” Both were right about the internet! And yet the first lost everything, and the second built a skill they rode for the next two decades. The difference wasn’t faith, and it wasn’t courage — it was that they bought different layers of the same revolution: one bought the valuation, the other bought the capability.

Chapter 7 ended with a promise: you are standing in the middle of a window. This chapter won’t try to convince you of that — convincing is a preacher’s job. It will hand you a tool instead: 5 tests that tell a window from a bubble before history announces the result. And then it will run AI through them, honestly — including the places where the verdict isn’t clear-cut.

Three windows and what the tide left behind

Every technology wave looks the same up close: euphoria, capital, overheating, the tide going out. The difference between a window and a bubble only shows in what’s left on the beach when the tide goes out.

The most important lesson sits in the first row. The dot-com crash was real: trillions of dollars in valuations evaporated. And yet the decade of 2000-2010 belonged to the internet more than the nineties ever did — because the bubble, as it burst, left behind cheap infrastructure: an overbuild of fiber that successors picked up for pennies, and an army of people who had learned a trade during the euphoria. The valuation died. The capability stayed.

Which leads to the conclusion at the heart of this chapter — the one missing from 9 out of 10 “is AI a bubble” debates: window and bubble are not an either/or. They usually coexist — in different layers of the same wave. The year 1999 was simultaneously a bubble (the valuation layer) and a window (the capability and infrastructure layer). The question is never “window or bubble?” It’s: which layer are you buying right now?

The five tests

Since windows can only be confirmed in retrospect, you need tests that work before history’s verdict. Here they are — each one calibrated on the previous waves:

Test 1: usage vs valuations. In a bubble, asset prices climb while the usage charts go flat. In a window, usage grows regardless of market mood — internet traffic grew in 1999, and in 2001, and in 2003, indifferent to the Nasdaq. The ICOs of 2017 failed this test perfectly: token valuations multiplied while the number of people using those protocols for anything at all stayed cosmetic.

Test 2: existing cost. The hardest windows replace spending that’s already in someone’s budget — because they don’t require faith in new demand, only the relocation of old demand. E-mail replaced the fax and the courier. E-commerce replaced the mail-order catalog. Bubbles live off promised demand: “someday everyone will…”.

Test 3: the residue. Picture the crash and ask what doesn’t disappear. After the dot-coms: fiber, standards, people with a trade. After the ICOs: not much beyond BTC and ETH. The more that’s left on the beach, the more the wave was a window — whatever happens to the valuations.

Test 4: rival adoption. Bubbles build walled gardens and fight format wars to the end. Windows have a TCP/IP moment: rivals who hate each other in the market adopt a shared standard, because the cost of not using it has overtaken the cost of pride. You know this test from chapter 1.

Test 5: unit cost. In a window, the cost of a unit of value falls exponentially (compute, data storage, bandwidth — every great wave had its own falling curve). In a bubble, costs rise along with the euphoria: customer acquisition gets more expensive, people get more expensive, everything gets more expensive except the revenue.

AI before the committee: 4.5 out of 5

Now for the honest part — because tests that always return “buy” aren’t tests, they’re marketing.

Test 2 — passed clearly. AI replaces costs that have been sitting in budgets for years: writing code, customer service, translation, research, content production. This is not promised demand; this is old spending changing its address (chapter 4 mapped it caste by caste).

Test 3 — passed clearly. Run the exercise: tomorrow, AI company valuations fall 80%. What remains? Open-weight models on the world’s hard drives, mountains of compute hardware (which — the fiber lesson — cheaper successors will snap up), protocols held by neutral foundations (MCP can’t vanish from any balance sheet, because it sits on none), and millions of people who learned to work compositely during the euphoria. The residue is thick.

Test 4 — passed by the textbook. The MCP adoption cascade from chapter 1 — competitors adopting a rival’s standard within 12 months — is the cleanest TCP/IP moment any wave has ever seen.

Test 5 — passed with an asterisk. The cost of inference (the price of a unit of model work) is falling by orders of magnitude — that’s a window’s curve. But in parallel, the cost of training frontier models is growing to the size of national budgets — that’s an overheating curve. Two curves at once: consumption getting cheap, peak production getting expensive. The asterisk stays.

Test 1 — ambiguous, and that needs to be said out loud. Usage is growing for real (the fastest-adopted consumer product in history — chapter 2; genuine penetration into companies). But valuations are growing too, and partly on promises. This is exactly the profile of 1999: real usage and overheated valuations at the same time. Test 1’s verdict reads: the valuation layer may be a bubble — and nothing follows from that for the capability layer.

The scorecard: 4.5 out of 5 — for the capability and infrastructure layer. For the valuation layer, this book has no forecast and doesn’t intend to have one. The operational conclusion, though, is unambiguous, and you’ve known it since February 2000: buy the HTML-textbook layer, not the Pets.com layer. Skills, infrastructure you own, positions that survive the tide — because even if valuations collapse, then (the lesson of all three windows) after the crash, the wave accelerates on cheaper infrastructure.

The best precedent for this asymmetry was written, fittingly, in the middle of the previous disappointment. In 2011, when the post-financial-crisis mainstream had written technology off as spent, Marc Andreessen announced, in the Wall Street Journal, a thesis that half the commentariat laughed at:

In short, software is eating the world.

Software is eating the world. He wrote that 9 years after the crash — at exactly the moment when the dot-com residue (cheap infrastructure, mature standards, trained people) was beginning to bear fruit: that was the decade in which the internet ate retail, media, and taxis. Windows don’t end with their bubbles. That’s when they begin in earnest.

The window seen from Poland

Everything above is visible from San Francisco. But you are probably reading this from Poland — and from here, the window looks different. In a few ways, better.

Geographical arbitrage has flipped. The classic arrangement went: a Polish specialist works for a fraction of the rate for a Western company. The post-cognitive era changes the equation: the composite stack gives frontier productivity to anyone who can assemble it — while the cost of living stays local. An operator in Gliwice and an operator in Palo Alto work on the same models today, for the same dollars a month. The difference: for one it covers an office; for the other, a life.

Language and context are a moat, not a limitation. Global players optimize for English and continent-sized markets; Polish niches — local law, institutions, custom, the long tail of Polish-language needs — are too small for them to bend down for. For a solo operator they are exactly the opposite: too big to eat alone. “Too small a market” from a corporation’s perspective is “a market with no competition” from the perspective of one person with a stack.

Adoption lag extends the window locally. Waves reach Polish small business with a delay — what counts as obvious in the Valley will still be a novelty in the average Polish company for years. For an operator, that means one thing: your local window is longer than the global headlines suggest.

The honest flip side: capital is shallower, customers are more cautious, and the exit market is smaller. The Polish premium is real, but it pays out in the calm of building, not in spectacular funding rounds. For an operator — possibly the better currency.

Four paths — conclusions, not coaching

If the tests came out the way they came out, then positions in the capability layer are taken along one of 4 paths, in order of rising risk: build (products and automations for niches invisible from San Francisco), advise (the compression from chapter 2 makes you the frontier’s translator for companies that have money but no time to learn), teach (with one hygiene rule: teach what you yourself deploy — the market of gurus without practice is crowded already), allocate (the time positions from chapter 7 and the fifth lever — with the decision journal from chapter 5, not with hope). Each of these paths passes the residue test: even if the wave fails, you are left with a skill, a reputation, or infrastructure.

And how much time is left? The honest answer: we don’t know — windows run 5-15 years and are dated in hindsight. We know it’s year four of the wave and the TCP/IP moment is behind us, which historically means the middle of a window, not the end. But this book won’t scare you with a clock, because fear is the bubble’s advisor, not the window’s. Instead of a clock you have the 5 tests — and they work both ways: when they stop coming up positive, you’ll be the first to see that too.

The window doesn’t need your faith. It needs a decision — and decisions are made on criteria, not headlines. February 2000 taught us that you can believe in the right revolution and buy the wrong layer of it. You have the tests. You know the layer. The rest is allocation.

---

The post-cognitive era — the period in which cognition stops being an exclusively individual resource and becomes composite: human + thinking model + AI + data + external memory. An extension of the Extended Mind thesis (Clark & Chalmers, 1998) into the age of LLMs.

Methodological disclosure: this book is written with AI as a co-author — this chapter was written by Claude Fable 5 (June 2026) from the author’s conceptual framework, with facts and quotations verified at the source; this English edition was translated from the Polish original (June 2026). This is not a gimmick but consistency with the thesis: a text about composite cognition is written by composite cognition — and thinking is versioned the way code is.