The AI Industrial Revolution: Are We Still Using Outdated Workflows?

Article by Will Awang

Over the past year, I’ve attended several industry conferences focused on AI. Onstage, speakers showcased one AI demo after another; downstairs, attendees filmed the screens with their phones, posted to Moments, and went right back to scrolling. But back in the office, it was the same weekly meetings, the same approvals, the same weekly reports. Big tech companies have already incorporated token consumption into their KPIs, and some have become model employees by running scripts to inflate usage. Meanwhile, on Moments, the same crowd declares today’s revolution is Claude, tomorrow’s marvel is Codex, and the day after’s savior is Gemini—are they embracing a revolution, or just rushing from one event to the next?

These are all noise, not the answer I want.

The real issue isn't whether AI is powerful enough—the steam engine has already been built; the question is who will be the first to tear down the old workshop.

The true beginning of the Industrial Revolution wasn’t the day Watt improved the steam engine—it was the day factory owners in Lancashire decided to move away from rivers and rebuild their workshops around steam engines. The same is true for AI’s most pivotal moment—it won’t be the day large models were invented, but the day the first organization decides to tear down old processes and rebuild its production methods around AI. That day hasn’t come yet—but it’s already on its way.

Two people saw this early on. Notion CEO Zhao Yivan wrote a piece in late 2025 titled “Steam, Steel, and Infinite Minds,” offering a sober assessment: we’re still in the “replacing waterwheels” phase—adding AI chatbots to existing tools, but no one is redesigning the factory. Leopold Aschenbrenner, a former OpenAI employee, took a different path: he wrote a 165-page paper titled “Situational Awareness,” then founded a fund that grew from $225 million to $13.68 billion, betting everything on AI infrastructure. One looked inward; the other bet outward.

This article is not about them. It’s about us—where we stand now and which part of history we are repeating.

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

Most people’s day goes like this: In the morning, they use AI to write an email, saving ten minutes; then spend two hours in a weekly meeting that didn’t need to happen; in the afternoon, copy and paste the same set of data across three tools; at night, they post on Moments saying, “AI is really great.” The ten minutes saved are completely swallowed up by the old processes.

Similarly, when the steam engine first appeared, factory owners simply replaced water wheels with steam engines, leaving everything else unchanged—factories were still built beside rivers, still multi-story, still powered by a central drive shaft running the entire production line. Today, we’re doing the same thing: putting ChatGPT into Slack, adding Copilot to Office, embedding AI chat windows into our workflows. The tools have been upgraded, but the workshop remains the same.

But changing machines doesn’t mean changing the workshop. McLuhan said it well:

We drive toward the future looking through the rearview mirror. Using old processes to accommodate new tools is like early films being nothing more than recorded stage plays. True breakthroughs come only when someone completely detaches the steam engine from rivers and redesigns the entire production system around this new power source.

Comparing the timeline of the Industrial Revolution with that of AI can help us pinpoint where we are on the map:

The timeline today has been drastically compressed. The Industrial Revolution took 60 years to go from the steam engine to railway mania, while AI has gone from the Transformer to a data center building boom in just 7 years.

Speed isn’t the issue—the problem is where we’re stuck: the first four lines are still in the phase of installing new machines in old workshops; the steam engine is in place, the rails are being laid, but production methods remain unchanged. The sixth line is the true turning point. We’re most likely stuck between these two steps.

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

Infrastructure is always overbuilt. In the end, it's the investors who go bankrupt, not the infrastructure.

In 1846, the British Parliament passed 263 railway bills, approving the construction of 9,500 miles of new railway. At the peak of railway investment, it accounted for 13% of Britain’s GDP. Railway shares could be purchased with just a 10% deposit, prompting a surge of middle-class investors. The bubble burst in 1847; one-third of approved lines were never built, and countless investors lost everything. Darwin lost 60% on railway stocks, yet he was far luckier than most.

But the railway remained.

Today’s AI infrastructure is following the same path. Goldman Sachs’s latest estimate projects global AI infrastructure capital expenditures to reach $765 billion in 2026 and an estimated $1.6 trillion annually by 2031. The proportion of capital expenditures to operating cash flow for hyperscale cloud providers has risen from approximately 40% in 2023 to nearly 70% in 2025. AI-related investments now account for about a quarter of all U.S. investment. Aschenbrenner’s $13.68 billion bet is placed on this layer—he’s not betting on which application will win, but on the underlying compute power itself.

This capital cycle is structurally identical to real estate development. Building a data center is like building a tower: land is electricity, building 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 cloud providers’ business model is “rent to finance”—using API revenue to cover the capital expenditures of data centers, while waiting for a valuation surge driven by the explosion of AI applications.

Hash power real estate: Each generation has its own infrastructure

The same core risk applies: Is the rate of decline in API unit prices being offset by the growth in usage volume? If rental income falls below the mortgage threshold—this is the nightmare most real estate developers are familiar with. The lesson from 2008 wasn’t that too many houses were built, but that the supply of houses didn’t match the actual demand structure. The equivalent risk in AI is: general-purpose computing capacity is oversupplied, while specialized capabilities capable of handling high-value use cases like financial compliance and medical diagnostics remain scarce.

Railroads, real estate, AI—three generations of infrastructure investments, all governed by the same rule: overbuilding is the norm, material suppliers always lose pricing power, and long-term returns always go to owners of “prime locations.” Look at Wall Street’s Q1 fund holdings—you’ll likely find 80% concentrated in this infrastructure layer: NVIDIA, data centers, cloud infrastructure. But the railroad frenzy teaches us this: it’s not the full picture of the AI revolution, and certainly not the layer with the highest returns.

The core location of AI is unique industry data and deeply embedded workflows. For individuals, the true "core location" is not the stocks they hold, but their irreplaceable judgment and industry knowledge—on the condition that they have already rebuilt how they use these capabilities around AI.

The real rewards lie one level deeper. But between infrastructure and value creation, there is no seamless connection—there is a gap, and historically, this gap has swallowed decades.

The people dismantling the workshop and the people using AI to improve efficiency are not doing the same thing.

Simon, co-founder of Zhao Yivan, was formerly a "10x programmer" but now rarely writes code himself—he simultaneously manages three or four AI coding agents, achieving 30 to 40 times the efficiency. Notion now has 1,000 employees and over 700 AI agents. The gap isn’t in the tools; it’s that Simon tore down his old workshop, while most people have merely swapped in a new waterwheel.

600 million Chinese users have used generative AI tools, a 142% year-over-year increase—the largest AI demand pool in the world. Yet almost no Chinese company has rebuilt its core workflows around AI. The world’s largest demand side, paired with nearly zero organizational change on the supply side. This contrast itself is a signal: it’s not that tools are lacking, but that organizations haven’t kept up. Context for knowledge work is scattered across dozens of tools and the minds of dozens of people; outputs are unverifiable, and no one knows how to judge whether a strategic memo is effective.

(Labor Market Impacts of AI: A New Measure and Early Evidence)

Anthropic has already taken action on a larger scale. They released the Economic Index, which uses real usage data to map out which tasks and industries AI will replace first, and then built upon this roadmap: partnering with Goldman Sachs, Blackstone, and Hellman & Friedman to form an AI-native enterprise services company; establishing a global alliance with KPMG, enabling access to Claude for 276,000 employees; and helping Accenture build a dedicated business unit, training 30,000 staff members with a focus on finance, life sciences, and healthcare.

These consulting firms do not act as users of AI, but as railway engineers—they do not build steam engines or lay tracks; instead, they help companies dismantle old facilities and rebuild production lines around new power sources. Without this role, most factory owners wouldn’t know where to begin.

Signals are already flashing. The sharpest one is coming from the job market.

Young adults aged 22–25 entering AI-exposed occupations are 14% less likely to find employment compared to their peers entering low-exposure occupations. Entry-level positions are already being squeezed.

If I’m a recent graduate, this number directly affects my job search. If I’m a manager, the next batch of entry-level hires I bring on may not be people at all.

The organization is being dismantled—what about me? My education, my resume, the industry experience I’ve accumulated over the years—these are my waterwheels. They once powered my entire production line, but the steam engine has arrived. 985 and 211 are no longer moats; they merely prove 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 six months have a 10% higher task success rate than new users. Those who started six months ago are already 10% ahead, and this gap will compound over time.

But no company has yet gone bankrupt for not using AI—at least my law firm is still pushing hard on AI. The winners haven’t been chosen by the market yet. The learning curve is real—those who started early are already gaining an advantage, but most are still at the starting line.

Will my current job title still exist in ten years? How many of the tools I used daily five years ago are still in use today? The answer to both may be no. But I don’t know what will replace them—because those things don’t exist yet.

This has always been the case throughout history. New things don’t get planned—they emerge on their own once old constraints disappear.

Before the railway was built, Britain consisted of isolated regional economies. The price of cotton cloth in Manchester could differ by 30% from that in London. Each city had its own local time standard, and no one found this problematic. Within twenty years of the railway’s construction, everything changed. The first nationwide market emerged, eliminating price disparities; standard time was not invented but forced into existence by the railway; jobs such as stationmasters, telegraph operators, and travel agents did not exist before the railway.

No one laying railroad tracks anticipated department stores. No one building steam engines anticipated standard time.

Steam, Steel, and AI Infinite Intelligence

Cities have told the same story. Centuries ago, cities were human-scaled—forty minutes of walking across Florence. Steel frames made skyscrapers possible, railways connected cities to their hinterlands, followed by elevators, subways, and highways. Tokyo, Chongqing, Dallas—these are not larger Florences; they are entirely new ways of life.

Today’s knowledge work is still on a human scale: teams of dozens, with rhythms set by meetings and emails—beyond a few hundred people, it becomes unsustainable. We’re building Florence with stone and wood. AI makes “Tokyo” possible—an organization of thousands of AI agents and humans, with workflows running continuously across time zones. Old practices like weekly meetings, quarterly planning, and annual reviews may no longer make sense.

Simon no longer writes code—his job has become "managing AI agents." Two years ago, this position didn't exist. My next job title may not even have a name yet. But someone is already building the future we can't yet name.

After tearing down the old workshop, what should be built? YC’s answer is: Let the company improve itself.

Their internal system now modifies its own code at night. An employee posted a query during the day, and it failed. A monitoring agent detected the failure, traced the root cause, wrote code to fix it, submitted the change for review, and deployed it live. The same query succeeded the next day. The entire process was completed while everyone was asleep.

This isn't AI helping people produce 30% more. This is the system completing an entire闭环 on its own and figuring out how to improve itself.

YC partner Tom Blomfield referred to this business model as a "recursive self-improving AI loop" in an internal talk. His assessment was straightforward: most companies are still like Roman legions—information flows down through layers and up through layers, with people acting as conduits. AI doesn’t just improve efficiency in one环节; it undermines the very foundation upon which this hierarchical structure exists.

The new logic he proposed is: burn tokens, not heads. The bottleneck is shifting from human labor to computational power. YC’s data shows that companies reaching Demo Day now generate about five times the revenue per person compared to 18 months ago. Middle management roles are being taken over by AI—coordination no longer requires human intervention. Everyone should be an IC, builder, or operator, with every task having a clearly named owner, not a committee.

Another prerequisite: the company must be “readable” by AI. Anything not recorded might as well never have happened. YC now archives all partner emails, logs every Slack message, and records all office hours. One partner used 2,000 hours of recordings accumulated over three months to have AI generate a new 150-page internal handbook—far superior to the original. This handbook updates automatically each month, becoming a constantly refreshed “living brain.”

Tom left a question:

If you were starting your company from scratch today, would you build it this way? If your company already has a hierarchical structure in place, you must answer an even harder question: Would the pain of rebuilding be less than the cost of continuing to operate like a Roman legion?

People are not in the center of the workshop—they’re on the periphery, handling the areas where AI cannot yet reach: offline judgment, novel situations, and high-stakes, high-emotion moments. The company’s core is a “company brain” built from data, records, and industry knowledge. The software running on top is disposable—once generated, it can be regenerated. What’s truly valuable resides in people’s minds—the understanding of how the business operates and which steps require judgment; these insights are the real assets.

Zhao Yivan describes the other side of this direction in "Steam, Steel, and Infinite Minds"—an organization where 1,000 employees collaborate with over 700 AI agents, with humans making judgments and agents handling execution. Aschenbrenner bets on compute infrastructure, while Zhao Yivan bets on organizational restructuring. Both paths ultimately lead to the same destination: a new mode of production rebuilt around AI.

Between the 1840s and 1850s—the railways were completed, but the factories had not yet been rebuilt.

Where are we? Simon no longer writes code. His waterwheel is one he took apart himself.

The issue has never been whether the steam engine was good enough; the issue is who was the first to tear down the old workshop.

I’m not trying to predict the future of department stores; I just want to focus on myself—ensuring I’m on the railway line, not beside a river that’s drying up.

What about you?