AI Is Different. The Last Wave Still Has a Job.
· 9 min read · By Jacques Jean
Most of us have learned to wait out the hype cycle.
For ten years, every eighteen months, somebody has walked in with a deck and told me this is the technology that changes everything. Web3. Blockchain. DAOs. Agentic coding pitched as autonomous startups. Real engineers, real money, real conferences, real failures. By the third or fourth round, you stop arguing and start waiting. You let the cycle run. You see what survives.
AI is the one I stopped waiting on.
Not because the marketing got better. Because the constraint actually moved. AI is the first technology in this run that earns the exception, and the reason is specific: it does not just route around gatekeepers, it does the work. The thesis of this post is simple. AI is a capability shift, not an access shift. Compute is no longer the bottleneck. We are. And the wave I just spent ten years being skeptical of still has a job to do.
Subtractive vs. additive technology
Look at what previous cycles did to compute.
Bitcoin and proof-of-work consumed it. ASICs got built and burned through electricity to write hashes nobody was going to read. GPUs got bought up by miners and pulled out of the market for everything else. The technology was subtractive. It took compute that could have been doing general-purpose work and converted it into a settlement guarantee. Useful in a narrow sense. But every dollar of demand made compute scarcer, more expensive, and harder to access for anyone trying to do something else with it.
AI is the inverse.
Every dollar of demand for AI pulls more compute into the world. TSMC books another fab. NVIDIA ships another generation. Hyperscalers commit hundreds of billions to data centers nobody had the financial appetite to fund three years ago. Power utilities are restarting nuclear plants for AI loads. Memphis, Northern Virginia, Phoenix, Abilene -- the AI buildout is reshaping the physical map of compute the way the railroads reshaped the physical map of the United States.
That is not a small shift. It is the difference between a technology that draws down the commons and one that expands it.
Bitcoin made GPUs more expensive for everyone. AI made compute the most-funded asset class in the world. The 2026 AI capex cycle is the largest infrastructure build in modern history outside of national grids. Every previous wave promised abundance and delivered scarcity. AI is the first one that flipped the polarity.
When the bottleneck moves, the conversation has to move with it.
Time is the new bottleneck
Once compute stops being the limit, the limit becomes us.
Everyone has the same 24 hours. The reason that mattered less in previous cycles is that the technology in question still needed something humans had less of -- compute, capital, coordination, regulatory clearance. AI flips it. Capable models are cheap and getting cheaper. Inference is moving to the edge. The marginal cost of asking a frontier model another question is approaching zero. What it costs is a person to ask it, evaluate the answer, and decide what to do with it.
That is not productivity hype. That is a structural change in what scales and what doesn't.
I notice it in my own work. The bottleneck used to be how fast I could write the code, the proposal, the schematic, the pitch. Now the bottleneck is how fast I can decide what is worth writing in the first place. The work compresses. The judgement does not. The technology is no longer holding me back. I am.
This is the part that quietly disqualifies most of the previous wave from comparison. Web3 promised to remove gatekeepers. AI removes the work. Those are different sizes of problem. One reroutes the river. The other drains it.
Capability problems vs. access problems
Two words worth being precise about: access and capability.
Web3 and blockchain solved access problems. Permissionless coordination. Sending money across borders without asking a bank or a country. Connecting people to tools previously locked behind corporations or governments. Banking the unbanked. Routing around capital controls. Those are real problems. The technology that solved them was real. The deployments were real. They just got drowned out by speculation, because speculation paid better than impact in the short run.
AI solves a different kind of problem. It does not unlock a door. It compresses what is on the other side. A doctor in a small clinic can now read imaging the way a specialist would. A solo founder can now ship the product of a small team. A teacher can now build curriculum the way a department would. A farmer can now plan irrigation the way an agronomist would. The work itself gets smaller.
That is why AI matters more right now than the previous wave did. Capability is closer to the root than access. If the work disappears, the access problem changes shape.
But the access problem does not disappear. It gets sharper. And that is the part the previous wave was actually built for.
DAOs decentralized the wrong thing
A short detour, because it matters.
DAOs tried to decentralize governance. The pitch was that strangers on the internet, holding a token, could collectively run a treasury, a protocol, a project, a community. The technology to do it shipped. The legal wrappers got built. The voting tools got built. Most of them are gone now, and the few that survived spend more energy on procedure than on output.
The miss was target selection. People did not need decentralized governance. Governance was already cheap in the places that had it and not the binding constraint in the places that didn't. What people needed was decentralized access. Access to tools. Access to capital. Access to networks. Access to the thing being governed in the first place.
Web3 had the right machinery aimed at the wrong layer. Decentralized voting on a treasury nobody could meaningfully contribute to is theater. Decentralized access to a tool that compresses your work is infrastructure.
The technology was right. The target was wrong. That is a fixable mistake.
The new access problem is AI itself
Here is the access problem nobody is treating like an access problem yet.
AI is becoming a subscription. $20 a month for the consumer tier. $100 a month for the developer tier. $200 a month for the heavy users. Trivial for a US founder. A rounding error for an Austin engineer. Impossible for the average person in Haiti, in much of sub-Saharan Africa, in parts of South and Southeast Asia.
I have been to Haiti. I saw what most of the world actually looks like before I had a frame for any of this. No consistent power. No consistent water. No consistent telecom. The idea that someone there is going to put $100 a month on a credit card to talk to a frontier model is a category error. The credit card is not there. The connectivity is not there. The disposable income is not there.
If AI is the most capability-amplifying technology of the century, and it ships behind a subscription wall, then the world's first true capability technology becomes the world's sharpest inequality vector.
That is not a hypothetical. That is the trajectory it is on right now.
We solved the wrong access problems for ten years and now we have a real one. The capability gap is becoming a wealth gap, in real time, and the rails to fix it are sitting in a folder labeled "previous wave."
A speculative synthesis
This is speculation, not prescription. I want it on the record because I think it is worth talking about, not because I think it is the answer.
The shape of it.
People in compute-rich countries already leave hardware idle most of the day. Mac minis on shelves. Gaming rigs at night. Workstations between meetings. The hardware can run useful local models -- Ollama, LM Studio, llama.cpp, MLX. The compute is there. The bandwidth is there. What is missing is the coordination and incentive layer that turns idle inference capacity into a global pool.
Web3 is built for exactly that job.
Token compensation for sharing spare compute. Settlement and metering at the protocol level, not at a corporate gateway. Routing requests to the nearest available node. Open standards for what counts as a unit of work. None of that is technically novel. The tooling has been waiting for a use case worth coordinating around. AI is that use case.
On the other side of the pipe, people in low-income regions tap that distributed pool through public infrastructure. Free or ad-supported. Schools. Libraries. Community centers. NGOs. Government-funded access points the way public phones used to be funded. The deliverable is concrete: something on the order of an hour of frontier-equivalent inference per person per day, no subscription required, no gatekeeper at the door.
Web3 does not need to merge with AI technologically. The model weights stay where they are. The training stays where it is. What Web3 is good at is exactly what AI distribution needs: an open settlement layer for a resource that is otherwise locked behind a payment wall most of the world cannot clear.
That is the job the last wave was actually built for. We just had it pointed in the wrong direction.
The previous wave was not a graveyard. It was a toolbox we picked up too early, before we had the thing worth distributing.
AI is the thing.
Web3 might still be how it gets to the rest of the world. I am not betting the next decade on it. I am paying attention. The technology that earns the exception this time is the one that compresses the work itself, and the technology nobody could quite point at the right problem last time might still be how that gets shared.
I would rather plant a flag here, early, and be wrong out loud, than wait another cycle and miss the version that works.
All posts · Home