AI Industrial Revolution: Where Are We Now?

Over the past year, I’ve attended some AI-themed industry conferences. On stage, speakers take turns showcasing the tricks and “magic” of AI. Off stage, people hold up their phones to capture the screens, post to social media, and then keep scrolling. But when you get back to the office, it’s still the same weekly meetings, the same approvals, the same weekly reports. Big companies have already baked token consumption into their KPIs. Someone can become an “employee of the year” just by using scripts to inflate volume. The people on social media—today it’s the Claude revolution, tomorrow it’s “Codex is amazing,” the day after that it’s “Gemini forever”—are they embracing a revolution, or just sprinting to catch the next train?

All of this is noise. It’s not the answer I’m looking for.

The real question isn’t whether AI is strong enough—steam engines have already been built. The question is who will be the first to tear down the old factory floor.

The day the Industrial Revolution truly began wasn’t when Watt improved the steam engine—it was when factory owners in Lancashire decided to leave the river, and rebuild their workshops around the steam engine. The same is true for AI’s most important moment—not the day a large model was invented, but the day the first organization decided to dismantle the old processes and rebuild the way production works around AI.

That day hasn’t come yet. But it’s already on its way.

Two people saw this early. Notion CEO Ivan Zhao wrote a cold assessment at the end of 2025, in “Steam, Steel, and Infinite Minds,” judging that we’re still in the “water wheel replacement” phase—adding AI chatbots to existing tools, but nobody redesigning the factory itself. A former OpenAI employee, Leopold Aschenbrenner, took a different path: he wrote a 165-page “Situational Awareness,” then built a fund that grew from $225 million to $13.68 billion, betting entirely on AI infrastructure. One looks inward, the other bets outward.

This article isn’t about them. It’s about us—where we are now, and which part of history we’re repeating.

( Power-loom weaving, engraving by J. Tingle after Thomas Allom, 1835 / Wikimedia Commons )

1. The workshop is still the old one

Most people’s day looks like this: in the morning, they use AI to write an email and save ten minutes; then they spend two hours on a weekly meeting that could have been skipped; in the afternoon, they copy-paste the same set of data across three tools; at night, they post a social media update saying, “AI really is great.” The ten minutes saved are simply eaten right back by the old process, unchanged from before.

Likewise, when the steam engine first appeared, factory owners initially just replaced water wheels with steam engines—everything else stayed the same: the factory was still built by the river, still multi-story buildings, still one central drive shaft powering the entire production line. We put ChatGPT into Slack, add Copilot to Office, and embed AI chat windows into the workflow—doing the same thing. The tools were upgraded, but the workshop didn’t change.

But switching to a new machine doesn’t mean switching to a new workshop. As McLuhan put it:

We drive into the future through the rearview mirror. Using old processes to accommodate new tools is like early films that are just stage plays filmed on camera. The real breakthrough has to wait until someone fully detaches the steam engine from the river and redesigns the entire production process around new power.

If you line up the timeline of the Industrial Revolution with AI, you can roughly locate where we are on the map:

Today’s timeline has been compressed dramatically. The Industrial Revolution—from steam engine to railway frenzy—took 60 years. AI—from Transformer to the data center construction boom—took just 7 years.

Speed isn’t the issue. The issue is where we’re stuck. In the first four lines, we’re still in the stage of putting new machines into the old workshop: the steam engine is installed, and the railway is being laid, but the production process remains the same. The sixth line is the real turning point. We’re probably stuck between these two steps.

The steam engine is in hand, but the workshop is still the old one.

2. All the money is poured into the layer farthest from the factory

Infrastructure is always overbuilt. The ones who go bankrupt last are investors, not the infrastructure.

In 1846, the UK Parliament passed 263 railway bills, approving the construction of 9,500 miles of new railways. At the peak, railway investment accounted for 13% of the UK’s GDP. Railway stocks required only a 10% down payment to buy, and the middle class rushed in. The bubble burst in 1847. One third of the approved lines were never built, and countless investors lost everything. Darwin lost 60% on railway stocks, though his luck was better than most people’s.

But the railways remained.

Today’s AI infrastructure is following the same route. Goldman Sachs’ latest estimate says that by 2026, global AI infrastructure capital expenditures will reach $765 billion, and by 2031, it’s expected to rise to $1.6 trillion annually. The proportion of capital spending by ultra-large cloud providers relative to operating cash flow has climbed from about 40% in 2023 to nearly 70% in 2025. AI-related investments already account for about one quarter of all investment in the United States. Aschenbrenner’s bet of $13.68 billion is on this layer—he isn’t betting on which application will win; he’s betting on the underlying compute power itself.

This capital cycle is isomorphic to real estate development. Building data centers is like building buildings: land is electricity, construction materials are GPUs and storage, contractors are data center builders, developers are cloud providers, tenants are AI application companies, and rent is API revenue. The business model of cloud providers is to “rent to fund the loan”—using API revenue to cover data center capital expenditures, then waiting for AI application breakthroughs to drive valuation jumps.

(Compute real estate: a generation’s infrastructure)

The core risk is the same: will the decline in API unit prices be offset by the growth rate of calls? If rent falls below the debt service line—that’s the nightmare real estate developers know best. The lesson from 2008 wasn’t that they built too many houses; it was that the structures didn’t match real demand. The equivalent risk for AI is: general-purpose compute becomes oversupplied, while the specialized capabilities needed to truly handle high-value scenarios like financial compliance and medical diagnosis remain scarce.

Railways, real estate, AI—three eras of infrastructure investment all share the same rule: overbuilding is the norm; material suppliers always lose pricing power; and long-term returns belong to owners who hold “prime locations.” Just look at Q1 fund holdings on Wall Street—you’ll likely find about 80% concentrated in this infrastructure layer: NVIDIA, data centers, and cloud infrastructure. But what the railway frenzy taught us is that this isn’t the full picture of the AI revolution, and it’s not even the highest-return layer.

What is AI’s “prime location”? It’s unique industry data and deeply embedded workflows. For individuals, the real “prime location” isn’t the stocks you hold—it’s your irreplaceable judgment and your industry knowledge. The prerequisite is that you’ve already rebuilt the way you use those capabilities around AI.

Real returns are in the next layer. But the transition from infrastructure to value creation isn’t seamless. There’s a gap in between—and historically, that gap has swallowed decades.

3. Who is tearing down the workshop

People tearing down workshops and people “using AI to boost efficiency” aren’t doing the same thing.

Ivan Zhao’s co-founder, Simon, used to be a “10-times-speed programmer,” but now he rarely writes code himself. He manages three or four AI coding agents at the same time, achieving 30 to 40 times efficiency. Notion now has 1,000 employees and more than 700 AI agents. The difference isn’t the tools—it’s that Simon tore down his own old workshop, while most people only swapped out a water wheel.

6 hundred million Chinese users have used generative AI tools, up 142% year over year—this is the world’s largest AI demand pool. But almost no Chinese company has rebuilt core workflows around AI. The largest demand side is moving, while the supply side organization barely changes. This contrast itself is a signal: it’s not that tools aren’t enough—it’s that the organization hasn’t caught up. The context of knowledge work is scattered across dozens of tools and the minds of dozens of people; outputs can’t be verified; and nobody knows how to judge whether a strategic memo is effective.

(Labor market impacts of AI: A new measure and early evidence)

Anthropic has already started on a larger scale. They published the Economic Index, using real usage data to show which tasks and industries AI first replaces, and then based on that picture they “build what’s next”: they formed AI-native enterprise services companies in a joint venture with Goldman Sachs, Blackstone, and Hellman & Friedman; they built a global alliance with KPMG, with 276,000 employees connecting to Claude; and Accenture formed a business group, training 30,000 people focused on finance, life sciences, and healthcare.

These consulting firms aren’t playing the role of AI users. They’re acting as the railway engineers of AI. They don’t build steam engines, and they don’t lay tracks. They help companies dismantle old factories and rebuild production lines around new power. Without this role, most factory owners don’t know where to start.

The signals are already flashing. One of the sharpest signals comes from the job market.

Young people aged 22-25 who enter AI high-exposure careers have a 14% lower probability of finding jobs than their peers entering low-exposure careers. Entry-level roles are already being squeezed.

If I were a fresh graduate, this number would directly affect how I search for a job. If I were a manager, the next batch of entry-level hires I recruit might not be people anymore.

Organizations are tearing down workshops—what about individuals? My education, my resume, the industry experience I accumulated over the years—those are my water wheels. They used to drive an entire production line for me, but the steam engine has arrived. 985 and 211 are no longer a moat; they only prove that I once built a decent factory by the river.

The question now is whether we have the ability to leave that river.

Anthropic’s data shows that users who have used AI tools for more than 6 months have a 10% higher task success rate than new users. Those who start half a year earlier are already 10% ahead, and that gap compounds over time.

But so far, no company has gone bankrupt for not using AI—at least my law firm is still singing praises of AI. The winners haven’t been selected by the market yet. The learning curve is real: early movers are accumulating advantages, but most people are still at the starting line.

4. My next career doesn’t have a name yet

Will my current job title still exist ten years from now? Of the tools I used every day five years ago, how many remain today? The answer is probably none. But I don’t know what will replace them—because those things don’t exist yet.

Historically, every shift has been like this. New things aren’t planned—they grow out of the old constraints once those constraints disappear.

Before the railway was built, Britain was a patchwork of isolated local economies. The price of cotton in Manchester could differ from London’s by 30%. Each city had its own time standard, and no one thought that was a problem. In the 20 years after the railway, everything changed. For the first time, a unified national market emerged, and price differences were erased. Standard time was forced by the railways, not invented. Stationmasters, telegraph operators, travel agents—those jobs didn’t exist before the railway.

No one foresaw department stores before laying the railways. No one predicted standard time before building the steam engine.

(The stories of steam, steel, and AI’s infinite intelligence)

The history of cities tells the same story. Centuries ago, cities were on a human scale—walking across Florence in 40 minutes. Steel frames made skyscrapers possible. Rail connected cities to the hinterlands, and elevators, subways, and highways followed. Tokyo, Chongqing, Dallas—these aren’t just “bigger Florence.” They are entirely new ways of living.

Now, knowledge work is also on a human scale. Teams of dozens, with meetings and emails scheduled in rhythm—once you go beyond a few hundred people, it becomes unbearable. We’re building “Florence” with stone and wood. AI makes “Tokyo” possible—organizations made up of thousands of AI agents and people, with workflows running continuously across time zones. Old weekly meetings, quarterly plans, annual reviews—maybe they no longer make sense.

Simon no longer writes code. His job becomes “managing AI agents.” Two years ago, that role didn’t exist. My next career title might not have a name yet either. But someone is already building the future we can’t name.

5. What does a new workshop look like?

After tearing down the old workshops, what do you build? YC’s answer is: let the company improve itself.

Their internal systems now rewrite their own code overnight. An employee runs a query during the day and it fails. A supervising agent reads the failure, infers the cause, writes code to fix it, submits for review, and deploys. The next day, that same query runs successfully. The whole thing is completed while everyone is asleep.

This isn’t AI helping people produce 30% more. It’s the system completing an entire closed loop on its own—figuring out how to get better.

In an internal talk, YC partner Tom Blomfield called this kind of company form a “recursive self-improving AI loop.” His judgment is straightforward: most companies are still like the Roman legions—layer upon layer relaying information up and down, with people acting as conduits. What AI breaks isn’t just the efficiency of a particular link. It breaks the premise that this whole bureaucratic structure exists.

His new logic is: burn tokens, not people. The bottleneck is shifting from manpower to compute. YC’s data shows that companies that make it to Demo Day have per-person revenues about 5x higher than 18 months ago. Middle management roles are being taken over by AI—“collaboration” no longer needs humans to do it. Everyone should be an IC, a builder, an operator—each task has a named owner, not a committee.

And there’s another prerequisite: the company must be “readable” to AI. If it isn’t recorded, it’s as if it never happened. YC is now archiving all partners’ emails, recording all Slack messages, and saving office hour recordings. Using 2,000 hours of recordings accumulated over three months, one partner enabled the AI to regenerate a 150-page internal manual—far better than the original. This manual updates automatically every month, becoming a continuously “fresh” living brain.

Tom left a question:

If you were to start building your company from scratch today, would you build it in this form? And if your company already has a bureaucratic hierarchy in place, then you have to answer an even harder question—would the pain of rebuilding be less than the cost of continuing to run like the Roman legions?

People aren’t at the center of the workshop anymore. They’re on the periphery—responsible for the parts where AI can’t reach yet: offline judgment, entirely new scenarios, and high-stakes, high-emotion moments. The company’s “big brain” is assembled from data, records, and industry knowledge. The software running on top is consumable—if it can be generated, it can be generated again. What’s valuable is in human minds: how the business runs, which steps involve judgment, and these understandings—those are the real assets.

What Ivan Zhao describes in “Steam, Steel, and Infinite Minds” is exactly the other side of this same direction: an organization with 1,000 employees and more than 700 AI agents collaborating, where humans handle judgment and agents handle execution. Aschenbrenner bets on compute infrastructure. Zhao bets on organizational reconstruction. Both paths ultimately lead to the same destination: a new mode of production rebuilt around AI.

6. Conclusion

Between the 1840s and the 1850s—railways were already laid, but factories hadn’t been rebuilt yet.

Where are we? Simon no longer writes code. His water wheel was dismantled by himself.

The question has never been whether the steam engine is good enough. The question is who will be the first to tear down the old workshops.

I’m not trying to predict the future department store. I just want to do my part—make sure I stand along the railway line, not guard a river that’s drying up.

What about you?