Educational content from the Criptospace desk. Not investment advice, and not a solicitation to invest.
Every industrial economy inherited the same unit of account: the hour. Wages, project estimates, actuarial tables, even how growth itself gets described — all of it assumes that turning an input into an output takes a roughly predictable, roughly linear amount of human time. That assumption survived mechanization, computers and the internet largely intact. It does not survive generative AI, and the break isn't cosmetic — it changes how value gets priced, how capital gets allocated, and how fast an advantage compounds once it exists.
The man-hour was always a proxy, not a law
Man-hours became the default unit of value because, for most of industrial history, output really did scale with labor-time — one more welder welded roughly one more unit per shift, one more analyst modeled roughly one more scenario per day. Software already cracked that logic open: a program written once runs infinitely at near-zero marginal cost. But writing it still took roughly proportional developer-hours. The bottleneck moved from execution to creation, and creation stayed human, and stayed linear.
Generative AI removes that remaining bottleneck. A working prototype, a first draft, a full pass of test coverage, a market model — the things that used to gate a project on a person's calendar — now happen in the time it takes to describe them. The hour hasn't gotten more productive at the margin; for a growing share of knowledge work, it has stopped being the unit that matters at all.
Why this feels exponential, not linear
It's tempting to treat this as one curve — models get better, so output gets better, in a straight line extending the last two years. That undersells what's happening. At least three separate curves are compounding at the same time, not one curve improving on its own.
Each of these would be significant in isolation. Stacked together, in the middle of the same adoption cycle, they don't add — they multiply:
- Model capability. Reasoning, context length and tool-use have each improved on their own trajectory, and gains in one compound with gains in the others rather than substituting for them.
- Unit economics. The cost per token has fallen by orders of magnitude in a few years, which changes not just how much gets built but what becomes worth building at all — tasks that were uneconomical to automate at last year's price are routine at this year's.
- Tooling and distribution. Agents, APIs and integration layers now turn raw model capability into a deployable system in weeks instead of years, so the gap between a lab result and a product on someone's desk keeps shrinking.
None of these three curves is exotic on its own. It's the fact that they're compounding on top of each other, mid-cycle, that makes the aggregate look exponential rather than linear — and makes extrapolating from last quarter's pace a reliably wrong way to plan the next one.
The internet comparison — and where it breaks down
The internet is the obvious reference point, and the pace comparison alone is stark: it took roughly a decade for the internet to reach a quarter of the U.S. population; television took about as long; radio took longer still. The most widely used generative AI products crossed comparable adoption thresholds in months. On pure speed of diffusion, this cycle is running several times faster than the one before it.
But speed isn't the structural difference that matters most. The internet was, overwhelmingly, a distribution and coordination technology — it made moving information, connecting buyers to sellers and coordinating remote work dramatically cheaper, while the production of that information and labor stayed human. AI acts one layer deeper: it doesn't just move the output of an hour of work faster, it reduces how many hours the output required in the first place. That's a different category of technology, and comparing its adoption curve to the internet's — while directionally useful — understates what changes once it's fully priced in.
What Kondratieff would say
Kondratieff's long-wave model reads economic history as a sequence of roughly 40-to-60-year cycles, each organized around a cluster of technologies that arrive together and take decades to fully build out: steam and textiles, rail and steel, electricity and chemicals, oil and mass production, then computing and telecom. Each wave has a build-out phase where growth compounds quickly, followed by a maturity phase where the same technology becomes infrastructure — necessary, but no longer the source of the marginal return.
The current AI wave reads as a build-out phase inside that model, not a one-off shock — it's the technology cluster this cycle is organized around, which is why growth is compounding rather than merely accumulating. But long-wave theory makes a less comfortable point too: the technologies that define the next wave are usually seeded well before the current one matures, quietly, while all the attention is still on the wave that's compounding.
The next wave is already seeded
None of the following is speculative. Each is an active, funded research and engineering program today, at a stage roughly comparable to where computing was in the 1960s. What Kondratieff's model suggests is that they matter now, not just eventually — because build-out cycles overlap:
- Quantum computing and quantum simulation, which attack a category of optimization and modeling problems that stay intractable for classical hardware no matter how much AI compute is thrown at them.
- Post-quantum cryptography, made urgent by that same quantum hardware — the encryption underpinning digital signatures, custody and blockchain security today is exactly what a sufficiently capable quantum computer breaks first.
- Nanorobotics and atomic-scale manufacturing, extending precision engineering from the chip down to the molecule, with the earliest real applications already showing up in targeted medicine and materials science.
- Next-generation silicon — photonic and neuromorphic chip architectures designed for a post-Moore's-Law world, built specifically to keep feeding the compute appetite this AI wave has created.
- AI itself, redeployed as the connective layer across the other three — the tool researchers now use to design better quantum error-correction, better post-quantum algorithms and better nanoscale materials.
Read that way, this isn't "AI now, quantum later" as two separate stories. It's one long wave whose build-out technology is also the tool accelerating the wave that follows it — which is why, for anyone allocating capital or managing risk on a multi-year horizon, pricing in only the current wave is already a form of being unhedged.