The Real Economic Threat Is Automation at Scale

Once you accept that we’re not in a speculative bubble (that what we’re watching is big tech finally deploying capital that’s been sitting idle for years), the next fear shows up quickly: if all this capital is being deployed, won’t AI wipe out jobs and crash the economy anyway?

I think that fear is misdirected: physical-world automation is the real threat to labor, not large language models, and it has very little to do with chatbots.

AI Makes Me Faster

I use AI every day. Claude Code is a core part of my workflow. I talk about it across this entire blog series. And I can tell you honestly: it makes me faster, but it doesn’t replace what I do.

AI is faster at writing code than me. That’s true. But I’m constantly correcting or questioning decisions: formatting, styling, making the code modular. The 0→1 is where AI struggles, the 10→100 is where it thrives, and the 1→10 is where it can do well with good guidance, though it still needs guardrails.

The difference is direction. If you tell an AI “build me an app that does X,” you’ve given it too much liberty. It’ll confidently build you something that doesn’t work. But “I need to build feature X using Y in our application code that lives here at Z”: that’s where AI shines. It needs the constraints. It needs the existing codebase, the chosen libraries, the architecture decisions already made. A seasoned engineer with a good environment already set up and clear opinions about their stack turns AI into a force multiplier. A blank canvas with no direction turns it into a confident mess.

Does it reduce the number of people a startup needs? I’m unsure. It lowers the barrier for entry in some cases and allows builders to build faster, but that’s not the same as eliminating roles.

This is Jevons’ Paradox playing out in real-time: when technology makes something more efficient, we don’t use less of it, we use more.

The pattern is everywhere. Watt’s steam engine made coal three times more efficient around 1865, economists expected consumption to drop, and instead it rose tenfold over the next 50 years because steam engines became viable for more applications. VisiCalc and Lotus 1-2-3 did something similar to financial analysts in the 1980s: they made analysis so cheap that every department started doing it, and finance teams grew instead of shrinking. Cloud computing did it again in the 2010s. AWS didn’t reduce total IT spending, it eliminated the barrier to spinning up infrastructure, so companies launched more projects, ran more experiments, and spent more overall. I lived through that last one. The cloud didn’t shrink teams, it gave them more to build.

LLMs will likely follow the same pattern. As AI makes certain tasks faster, organizations won’t fire everyone, they’ll find more tasks worth doing. The bottleneck moves, but the work expands. White collar work shifts. It doesn’t vanish.

The “AI in Everything” Problem

There’s a tension I see from the inside: there’s a clear direction from leadership to use AI and build AI products, AI in everything, even just throwing it in for buzzwords.

I watched a PE make a demo of something that used AI. I rewrote it with regex and parsing rules, made it deterministic, cheap, and faster.

That’s the dirty secret of a lot of AI integration right now. LLMs are wonderful pieces of technology, but the last thing they are is deterministic. You can use the same large prompt multiple times and get different results. For a lot of use cases, data validation, parsing, rule-based processing, that’s a problem, not a feature. The companies I’ve been in pitch rooms with understand this: humans don’t scale, but AI isn’t good enough to replace them yet. The cost to process all this data with humans doesn’t make sense, but LLMs aren’t deterministic enough to do it reliably either.

That gap, between humans being too expensive and AI not being reliable enough, is where a lot of the real opportunity lives right now: building the tooling that bridges it, rather than replacing people.

The Real Shock Is Physical

The fear that AI will eliminate white collar jobs misses what’s actually happening. There’s a movie called Hidden Figures where the women had the job title “Calculator”: that was literally their role. Now we have advanced calculators in our pockets, and nobody mourns the loss of that job title, because the work evolved and the people evolved with it. AI is doing something similar to knowledge work, changing what those roles focus on rather than eliminating them.

But physical automation is a different story entirely. When a robot replaces a warehouse picker, there’s no “evolution” of that role. The role is gone.

If you want to talk about an existential labor shock, stop looking at chatbots and start looking at warehouses.

The trucking industry employs 8.5 million people, including 3.55 million professional drivers. Studies predict anywhere from 400,000 to 60-65% of heavy truck driving jobs could be displaced by full automation. Warehouse roles involving repetitive tasks have already declined 29% since 2023 due to AI-integrated systems.

Amazon is the clearest case study. According to internal documents reported by The New York Times, Amazon plans to automate roughly 75% of warehouse operations by 2033, potentially eliminating over 600,000 jobs. The company already operates more than 1,000,000 robots, approaching parity with its 1.56 million human workforce.

At a pilot facility in Shreveport, Louisiana, a thousand robots allowed Amazon to employ 25% fewer workers than traditional warehouses. By 2026, that number drops to 50%. The company plans to replicate this design across 40 facilities by 2027, starting with a massive warehouse in Virginia Beach.

The projected savings: $12.6 billion from 2025-2027, about 30 cents per item shipped.

These systems don’t “assist” humans, they replace them. And unlike white collar roles, there’s no ambiguity: a truck either drives itself or it doesn’t, a warehouse either needs pickers or it doesn’t.

AWS is roughly 15-17% of Amazon’s revenue but was 57% of the company’s profits. The retail and distribution side is a massive revenue engine, but the margins are thin, so any reduction in labor costs there is welcome: it lowers the overall shipping and storage cost per item.

If Amazon figures out physical automation at scale, they become the most profitable company in the world. That’s arithmetic. The infrastructure spending I talked about in Part I covers AI models and cloud compute, but a lot of it is building the physical automation layer that turns Amazon’s revenue machine into a profit machine.

The autonomous truck market is projected to reach $1.5 trillion by 2034, growing at 16% annually. That’s industrial transformation with a timeline and a spreadsheet attached.

Why AI Gets Blamed Anyway

AI becomes the villain because it’s visible. People interact with models daily. They feel replaced. They feel diminished.

But most of the capital is flowing into machines, buildings, power, and logistics, the same places automation has always lived, not into replacing humans with chatbots. AI is just the interface layer people see.

And inside these companies, the messaging doesn’t help. Leadership says “ruthless prioritization” and “use AI for everything” in the same breath. Engineers go through layoff after layoff (I’ve been through three rounds at Amazon) and then get told to be more efficient. The layoffs aren’t because of AI. But the narrative conflates them, and it’s easy to see why people connect the dots even when the dots don’t actually connect.

Infrastructure spending creates work, but not evenly or immediately. The jobs show up as construction, electrical work, hardware manufacturing, networking, power engineering, and data center operations. Capital deployment created those jobs. I think there will be more of a focus on servicing the robots, maintaining the hardware, and humans building things: furniture, art, anything handmade. When machines handle the repetitive physical work, the value of human craftsmanship goes up, not down.

But the mismatch is temporal and geographic, not existential. The uncomfortable truth is that the workers displaced from trucking and warehousing aren’t automatically qualified for robotics maintenance roles. The net job count might be positive while individual lives are disrupted. Aggregate statistics can hide a lot of pain.

Where This Argument Could Be Wrong

I could be wrong about this in a few ways. Jevons’ Paradox assumes the technology is a tool that amplifies human capability, but what if LLMs become capable enough to fully substitute for knowledge workers instead of just assisting them? I’m skeptical this happens in the near term, based on what I see daily in my own workflow, but I can’t rule it out over longer time horizons.

White collar resilience might also be temporary. The jobs that feel safe today, strategy, judgment, relationship management, might be exactly the ones AI targets next. If models get significantly better at reasoning and context, the ambiguity buffer protecting white collar work could shrink faster than I expect.

The transition costs are real even if I’m right about the rest of it. Even if Jevons’ Paradox holds and new jobs emerge somewhere else, that’s cold comfort if you’re a truck driver in your 50s.

There’s also the policy problem. I’m arguing that physical automation is the bigger risk, but that assumes policymakers focus on the right problem, and history suggests they usually don’t. We might end up regulating chatbots while autonomous trucks reshape logistics unimpeded.

What to Watch

If you want to track whether physical automation is actually reshaping the economy, these are the signals that matter.

In the near term, 2025 to 2027, watch whether Amazon’s Virginia Beach facility actually hits its 50% labor reduction target when it goes live, how far Waymo and Cruise expand beyond Phoenix and San Francisco (regulatory approval velocity is the thing to track), and whether warehouse robotics density at major retailers passes a 1:1 robot-to-human ratio.

In the medium term, 2027 to 2030, the first autonomous trucking corridors should start operating at scale, probably in Texas and Arizona first. Port automation should keep climbing past 80% at major hubs, Rotterdam is already there while LA and Long Beach lag behind, and last-mile delivery economics should start flipping as robotic delivery gets cheaper than human delivery in dense metros.

On the policy side, watch whether state-level autonomous vehicle legislation leans permissive or restrictive, whether federal infrastructure bills include automation transition funding, and how union negotiations in logistics and transportation play out.

The infrastructure spending happening now will determine the pace of these transitions. The money is already allocated. The question is execution.

I don’t think AI replaces everyone. The real risk is that automation concentrates gains too quickly, labor transitions lag capital transitions, policy moves slower than machines, and we end up regulating the wrong thing.

AI will change how work feels. Automation will change whether some work exists at all. Those are very different conversations, and we keep mixing them up.

The calculator didn’t disappear when the job title did, it became an app, and the people who held that title became analysts, engineers, programmers. That’s the pattern AI is following for knowledge work. But the warehouse picker, the long-haul trucker, the port operator, those roles don’t evolve when a robot can do them faster, cheaper, and around the clock. They just end.

The money is already being spent. I’d rather spend the next few years watching the warehouses, the ports and the trucking corridors than arguing about chatbots.