“AI Will Replace You” Is Probably the Greatest Lie Ever Told

“Just 6 months to go” yeah right

CEOs keep saying AI will replace all white collar jobs within 18 months.

And an electrical engineer just did the actual math on why that’s physically impossible.

Not pessimistic. Not anti-AI.

Just math.

Start with what AI actually is

A large language model is a mathematical function with adjustable parameters.

You feed it input. It runs it through a massive number of equations. It produces output.

The reason AI exploded recently isn’t because someone suddenly got smarter. Two things happened to align at the exact right moment.

Transformers — a 2017 architecture that lets every word look at every other word simultaneously instead of reading one at a time — turned out to scale perfectly with GPUs.

GPUs are designed exactly for parallel math operations. Transformers need parallel math operations.

That alignment produced the steep climb everyone felt. The feeling that AI was improving exponentially and would keep doing so forever.

It won’t. Because that’s not how any technology works.

The S-curve that nobody talks about

Every technology in history follows the same pattern.

Early phase — slow progress, mostly ignored. Inflection — a key idea unlocks scaling. Steep climb — everything feels explosive. Upper bend — constraints start dominating, improvement continues but costs more per gain.

Semiconductors. Batteries. Networking. Imaging sensors. Every single one.

AI is not special. It follows the same curve.

OpenAI’s own researchers published a paper in 2020 called Scaling Laws for Neural Language Models. Key finding — model performance improves predictably as you add compute, data, and parameters. But it follows a power law. Diminishing returns.

The jump from GPT-2 to GPT-3 felt enormous. The jump from GPT-4 to GPT-5 cost vastly more and delivered a much smaller perceptible improvement.

That’s not a bug. That’s math.

The memory problem nobody explains

Most people think AI’s constraint is GPUs.

The real constraint is memory.

Specifically two things — capacity, how much data fits on one chip, and bandwidth, how fast data moves between memory and compute.

The Nvidia H100 — the most widely deployed AI chip right now — has 80GB of high bandwidth memory and 3.35 terabytes per second of bandwidth.

A 100 billion parameter model at 2 bytes per parameter requires 200GB just to store the weights.

That’s more than double what fits on a single H100. Before a single input has been processed.

So you have to split it across multiple GPUs. That means fast interconnects. More hardware. More cost. More complexity.

And we’re talking about a model that’s considered modestly sized for what these CEOs are describing.

The KV cache problem

When you’re in a conversation with AI it stores key and value vectors for every previous token.

The longer the conversation the bigger this stored memory gets. More memory per user means fewer simultaneous users per GPU.

If you want millions of AI agents running all day — maintaining long conversation histories, handling complex multi-step tasks, replacing human workers — every single one of those agents is a KV cache sitting in GPU memory.

Running a tab. You pay for that tab in hardware.

Multiply by millions of agents and you have an infrastructure buildout that makes current AI spending look tiny.

Adding more GPUs doesn’t solve it

Amdahl’s Law.

If 95% of your work is parallelizable and you throw 1,000 GPUs at it — your system doesn’t get 1,000 times faster.

It gets 19.6 times faster.

Because 5% of the work — synchronization, communication between GPUs, data dependencies — can never run in parallel. That 5% becomes a hard wall. You can throw a million GPUs at it and the wall doesn’t move.

The doomer narrative requires civilization-scale deployment. Amdahl’s Law proves that just adding more GPUs doesn’t linearly solve that problem.

The manufacturing reality

Even if you solve memory. Even if you solve bandwidth. Even if Amdahl’s Law didn’t exist.

You cannot manufacture AI chips fast enough.

Modern AI GPUs require three things simultaneously.

Advanced lithography machines — made by exactly one company on earth. ASML in the Netherlands. Their EUV machines cost $200 to $370 million each. Weigh 180 tons. Require three Boeing 747s to ship. Contain over 100,000 individual parts. Lead times measured in years.

High bandwidth memory — made by SK Hynix and Samsung. Production is already booked out.

Advanced packaging — the technology that bonds the GPU die to the memory stacks in one unit.

You need all three at the same time. At a scale massively larger than current production. Which is physically impossible to achieve in 18 months.

The power math

An engineer did the actual calculation on what replacing 100 million white collar workers would require.

100 million workers. One H100 per worker — being extremely generous. H100 draws 700 watts under load.

100 million times 700 watts equals 70 gigawatts.

Add 30 to 50% cooling overhead.

Total — roughly 90 to 105 gigawatts.

The entire US data center industry right now — for everything, not just AI — consumes 20 to 30 gigawatts total.

Replacing white collar workers requires tripling to quadrupling the entire US data center infrastructure.

In 18 months.

Building a new data center takes 6 to 24 months with everything going perfectly. Large power transformers have 18-month lead times. You can’t run these steps in parallel like a software sprint. Each depends on the previous step finishing first.

The math isn’t pessimistic. It’s just arithmetic.

Hallucination isn’t a bug they can fix

People assume hallucination is a software problem that will get patched out.

It isn’t. It’s structural.

Transformers are trained to predict the next token — to generate text that sounds plausible. Not to check facts against a database. Pattern matching at massive scale.

You can get the hallucination rate down from 15% to 3%. But a probabilistic system always has a probability of being confidently wrong.

In high-stakes domains — legal filings, medical decisions, financial compliance, engineering sign-offs — a 1% error rate is disqualifying. One wrong answer isn’t “oops try again.” It’s a lawsuit. Someone dying. A bridge falling.

Until reliability reaches the threshold each industry requires, a human has to check the work.

A human checking the work means the job still exists.

The economic argument that kills the doomer narrative

Even if you ignore all the engineering. Even if you give them unlimited compute tomorrow.

Their model of how displacement works is still wrong.

The lump of labor fallacy — the assumption that there’s a fixed amount of work in the economy and if a machine takes some, humans lose some.

This fallacy has been proven wrong for 200 years in a row.

When ATMs rolled out in the 70s and 80s everyone predicted bank tellers would disappear. ATMs made branches cheaper to run. Banks opened more branches. Total bank teller headcount increased over the following decades.

Jevons’ Paradox — when steam engines got more fuel efficient, total coal consumption went up. Efficiency made steam power cheaper which made it accessible to more industries which expanded usage dramatically.

If AI makes financial analysis five times cheaper, companies won’t just fire analysts and do the same analysis. They’ll do ten times more analysis. They’ll explore scenarios they couldn’t afford to explore before.

Efficiency almost never shrinks demand. Historically it reliably expands it.

What Anthropic’s own data shows

Anthropic published a chart in March 2026.

Blue area — theoretical AI capability across job categories. Computer and math jobs — 94% of tasks theoretically handleable. Office and admin — around 90%.

Red area — actual observed AI usage in real professional settings measured from Claude usage data.

The red is tiny.

Computer and math workers — 94% theoretical capability. 33% actual coverage. And that’s the highest adoption category. Everything else is even smaller.

The gap between blue and red isn’t people not knowing about AI yet. The gap is memory bandwidth ceilings. Power grids. Enterprise security reviews. Hallucination rates. Amdahl’s Law. Deployment realities.

Anthropic’s own researchers then looked at unemployment data for workers in the most AI-exposed occupations.

Finding — AI has not increased unemployment in the most exposed jobs after two-plus years.

One real signal — hiring of workers aged 22 to 25 has slowed in the most exposed occupations. Entry-level and routine work getting pressured first. Real. Worth paying attention to.

But hiring slowing for some entry-level roles is completely different from all white collar jobs disappearing in 18 months.

The one wildcard that could actually change this

A fundamentally new AI architecture.

Not transformers with more parameters. Something completely different that scales in a way that doesn’t hit current hardware constraints.

Nobody knows if that exists. Nobody knows when it would arrive.

Building career decisions around an unpredictable future breakthrough isn’t strategy. It’s panic.

The actual timeline

AI will transform how work gets done. That’s real.

The people pushing 18-month extinction timelines are going against engineering, against economics, against manufacturing reality, against their own research teams, and against the actual employment data.

The honest timeline based on how every transformative technology has actually deployed — not how fast demos run on dedicated GPU clusters — is 7 to 15 years for broad workforce transformation.

Faster than electricity. Faster than the internet. Still not 18 months.

Prepare for change. Don’t spiral into decisions based on fear.

The physics doesn’t care about the pitch deck.

Want more stuff like this? Even more detailed? I go into deep dives every single day on my 📌Substack📌

At the price of a coffee I can give you enough knowledge to replace a bachelor’s degree in computer science or maybe even a masters, who knows.

💜 If you like my work and would like to support to me financially. Even a small donation would help a lot!

My PayPal — Please Support