Transcription
We have tech CEOs say AI will replace your job in a matter of months. Listen to enough of these CEOs push this narrative and you actually start to believe it. We at Meta are going to have an AI that can effectively be a sort of mid-level engineer that you have at your company that can write code, entry-level right work. I worry that those things are going to be first augmented, but before long replaced by AI systems. And 18 months, two years, every white-collar job in the USA just gone to be replaced by software that never sleeps.
Here's the other thing that these CEOs are saying just on a different call that you are not privy to. The gas turbines that they need to power all of this is a 5 to 7-year wait. The grid connection that they need for this power, it's a queue that doesn't clear until the 2030s. There are power plants that have already been built that the utility won't switch on because there's just nowhere to put the power. Both of those statements cannot be true at the same time. If the replacement timeline is 18 months and the power infrastructure timeline is 7 years, somebody's lying. One of those timelines is based on engineering. The other is just plain marketing. I'm going to show you which is which. And by the end of this video, I will have provided you with the ability to fact-check any one of these headlines by yourselves in about 30 seconds using nothing but your power bill and a calendar.
So, let me tell you why I'm qualified to be the one telling you this. I'm a chemical process engineer. I spent years in the energy sector designing process systems, heat exchangers, gas turbines, cooling loops. Part of my job was to look at a facility and work out what it can actually produce, not just what the brochure says it can. Also, I have a PhD in sustainability and energy access, which is the other half of this picture. My research was about who actually gets power, who doesn't, and what it actually takes to build energy systems at scale, and of course, who ends up paying. And so, when a tech CEO talks about replacing 100 million workers, what I hear isn't a forecast about jobs. Instead, I hear a claim about power and infrastructure. And because of my background, that's a claim I can check with relative ease.
Now, let me be clear about what I think is actually going on here. This is a narrative that's being pushed by people who have a direct financial interest in you being afraid. Because the scarier AI sounds, the more money flows into their companies. It's that simple. Fear is the product. And when you are afraid, you begin to make decisions that benefit whoever frightened you. You perhaps take a lower salary offer because you've been told your whole field is about to disappear. None of this helps you. All of it helps the company that needs you cheap, compliant, and convinced that resistance is pointless. So that's what this video is. I'm not here to tell you that AI won't change anything because it will. I am here to take the claims that they make, you know, their timelines, their numbers and test them against real engineering constraints: power, heat, water, grid, infrastructure, basic economics. I will let the engineering tell you how worried you should be.
Now, I want to be fair because there is a serious case for the defense. And so, before I demolish this, I'm going to state the claim at full strength. There are roughly 100 million white-collar workers in the US. Globally, closer to 400 million or so. The claim, and I'm quoting the CEOs directly here, is that a significant share of these jobs could be replaced, so not assisted, but entirely replaced by AI within the next year to 18 months. So let's be clear about what they mean by replace, because these guys are not talking about, you know, a quick ChatGPT query where you ask your chatbot one question and you close the tab. They are talking about always-on AI agents, systems that run persistently. They handle complex multi-step tasks, maintain long context, and do the work that a human would do at a desk all day, every day. That's what they mean when they say replace. And this is important because the moment that you say 100 million always-on agents, you've said something that an engineer like me can measure because each agent needs a chip, and each chip pulls power. 100 million of them running continuously. That isn't an app that you download. It's a number in watts. And that number is where this entire story runs into the physical world.
You see, each of those AI agents running persistent, long-context, always-on work draws roughly 700 watts of GPU power. 100 million agents at 700 watts each is 70 gigawatts, just the chips, by the way. So now, 70 gigawatts, let me give you some context. The entire United Kingdom's power grid, so every power station, every wind farm, every nuclear plant, every solar panel in the country, has an installed capacity of about 76 gigawatts. And so replacing US white-collar workers with AI requires almost an entire United Kingdom's worth of new power generation. And this is just the chips. And this is, of course, before you even cool a single server, as otherwise you damage the chips. But we'll get to that later. And that's why these claims are straightforward for an engineer to dissect. A claim about replacing labor at scale is underneath a claim about building power infrastructure. You are not just replacing a person with software. You are replacing them with a power plant, with a cooling system, and a grid connection. So let's find out if they can build this.
Before we get into the numbers, I want to give you something that you can keep, a portable tool, if you will, because the next time you see a headline about AI replacing jobs, I want you to be able to check it yourself in about 10 seconds or so. You see, at its core, an AI agent is basically electricity that's turned into tokens. That's it. You put power into the data centers, the GPU does computation, and heat comes out every single time. That's not an opinion. That is the first law of thermodynamics. You cannot destroy energy. You can only convert it. And in the case of AI, you convert it from electricity into computation and heat. So here's the tool I mentioned earlier. Anytime someone tells you AI is going to replace X million workers at a certain date in the future, don't argue about whether the model is clever. Just do the power balance. Take the number of workers, multiply by the power each AI agent needs, and ask them: How many watts does this future need? How many watts exist today? Can they generate a difference in the time specified? And can they cool it? If any one of those four answers is no, the claim is wrong. Full stop.
So let's take their 100 million agents and run the power balance against two walls. The first wall is generation and delivery. All right. So, we've established that 70 gigawatts of GPU power is needed just for the chips. 100 million agents at 700 watts each. Now, let's be generous. Let's say the engineers are brilliant and the cooling overhead is only 30 to 50% on top of the compute load. That puts you at roughly 90 to 105 gigawatts of continuous power draw. For the purpose of this video, let's call it 100 gigawatts, just a round number. Now, this number on its own might not mean anything to you. So, I'll put it in perspective. The US data center fleet, the entire infrastructure that runs every cloud service, every streaming platform, every AI chatbot that you've ever used, draws about 20 gigawatts on average. That's not peak capacity. It's just the average continuous draw. And that amounts to about 4.5% of all US electricity. So replacing 100 million white-collar workers with AI isn't adding a little bit more load. It's a three and a half to five times build on top of the fleet that already exists. You would need to quadruple the entire US data center power footprint just for the replacement scenario before you've even built a single new hospital, school, or housing development that also needs electricity. The Department of Energy and Berkeley Labs project data centers will reach somewhere between 325 and 580 terawatt-hours by 2028. So if we take the very top of that range and call it 66 gigawatts on average, that's still below the 70 gigawatts that you'd need just for the agents. And it's a 2028 figure built on every optimistic assumption that they had. Where does that power come from? And how fast can you actually get it? I'll give you an honest engineering answer: slowly, very, very slowly.
I'll start with the generation part. Gas turbines. So these are the workhorses of American power. The fastest conventional generation that you can build. They now have a lead time of 5 to 7 years. That's just the machine. It doesn't include the permitting, not the sites, not the grid connection, just the turbine. And so you order one today. It arrives in 2031 at the earliest. And the price has jumped roughly 50% in 6 months from about $2,000 per kilowatt to $3,000 per kilowatt. Supply chains are absolutely jammed. And that's the fast option. Nuclear, well, new builds are running 10 to 15 years. Renewables, they're faster to construct, but you need storage and transmission. And transmission is where the wall is because here's the delivery problem. The US interconnection queue. So that's the line of power projects that are waiting for permission to connect to the grid, is currently 2.6 terawatts long. And you don't just build a power plant and you just plug it in because the median time from applying to actually generating power is now about 5 years. For recent projects, it's getting worse, not better.
Now, I know what some of you are going to say. But wait, what about behind the meter where they bypass the grid entirely? There are companies building their own power plants on site. Elon Musk is doing it in Memphis. Now, this is true, and it's the most revealing thing in this entire story. XAI's Colossus facility in Memphis is running roughly 1.5 gigawatts of behind-the-meter gas, private turbines, off-grid, no queue. Microsoft is doing the same in West Virginia. These companies are so desperate for power that they are building private gas plants because the grid literally cannot connect them fast enough. But here's what they're not telling you. Total behind-the-meter gas capacity online right now, as of mid-2026, is about 2 gigawatts, just two. We need 100 for this scenario. And those private gas plants take roughly 18 months to build. The turbines still have the 5 to 7-year lead time and 50% price jump. So going off-grid doesn't make the turbine queue disappear. And so the math tells us what the CEOs want. Generation says no. Delivery says no. And the escape hack, the private behind-the-meter plants, also say no. That's wall one. We've established the power isn't there and it can't be delivered fast enough.
But let's pretend it could be. Let's pretend in this fantasy world that the grid queue vanishes overnight and the turbines arrive tomorrow and you've got magic 100 gigawatts of electricity flowing into your data centers just ready to go. You still have a cooling problem. Do you remember the tool I gave you earlier? Electricity in, computation out, heat as waste, the first law of thermodynamics. Well, practically every single watt of power that goes into a GPU comes out as heat. You put 100 gigawatts of electricity in, you get, for all intents and purposes, 100 gigawatts of heat out, bar a minuscule amount of light. This is basic first-year chemical engineering. And here's the part the AI industry doesn't talk about in their pitch decks. You now have to get rid of all that heat. All of it. Because if you don't, the quite expensive GPUs will literally cook themselves and your AI replacement agent dies. And so the question is, what does it actually take to reject 100 gigawatts of heat? Let me walk you through it the way I would walk you through a chemical plant. In process engineering, when you've got a reactor producing heat, and a GPU rack can be seen as just a very dense reactor, you design a cooling system. The cooling system has a coefficient of performance, also known as a COP. That's the ratio of heat removed to energy input. A typical data center cooling system using chilled water units runs at a COP of about 3 to 4. And that means for every 3 to 4 watts of heat that you remove, you spend 1 watt of electricity running the cooling system. And so that 100 gigawatts of compute, let's add another 33 gigawatts just to run the cooling. So we are now at 133 gigawatts. And remember, we started this calculation pretending that the grid problem was solved. It wasn't. We've just made it worse.
But wait, because the COP I just quoted you, the 3 to 4 figure, that's at standard conditions. Moderate ambient temperature, clean water, well-maintained equipment. And here's the thing about data centers. They are not built in moderate conditions. They are built wherever land and power is cheap, which often means hot, dry climates. So, Virginia, Texas, Arizona. And when ambient temperature rises, COP degrades because the hotter it is outside, the harder your cooling system has to work to reject heat. The harder it works, the more power it draws. And the more power it draws, the more heat the cooling system itself generates, which you then have to cool. It's a feedback loop. And in places like Arizona in the summer, where ambient temperatures can hit 45 degrees Celsius, your COP can drop by 30 to 40%. So your cooling overhead can go from 33 gigawatts to nearly 50 gigawatts. And we've not even included things like the pumping power needed, but that's a story for perhaps a whole different video. There's also a thermodynamic limit to air cooling. You cannot air cool a GPU rack beyond a certain power density. It's not a money problem, and it's not an engineering effort problem. It's just a physical limit. The heat transfer coefficient of air is roughly 25 to 50 watts per square meter per degree. If you compare that to a liquid, about a thousand to 10,000. So two to perhaps three orders of magnitude. This is why the industry is scrambling towards direct-to-chip liquid cooling, putting servers in baths of dielectric fluid. And to be fair, liquid cooling does work. It's a real solution to the heat density problem. But here's what nobody mentions when they pick liquid cooling. It doesn't eliminate heat. It just moves it. You still need to reject 100 gigawatts of thermal energy into the environment. The liquid transports it more efficiently, but the heat still has to go somewhere: into a cooling tower, into a river, into the atmosphere through evaporation. And evaporation needs water. Researchers at UC Riverside and Caltech estimate that the buildout needs somewhere between 700 million and 1.5 billion gallons a day of additional peak water capacity by 2030. The top of that range is roughly the daily water use of New York City. One whole New York City in new peak demand on top of what is already used. The new water works supplied run 10 to 58 billion. Some single facilities are already asking for up to 8 million gallons a day on their own. So now you've got both walls. You can't generate and deliver the power on the timeline. And even if you could, you can't cool it without a city's worth of water that isn't there.
Which raises the obvious question: If it can't actually be built, why does the buildout keep going? And who's paying for everyone trying? To the question of who's paying? Well, it's you. In Virginia, the largest data center market on the earth, the State Corporation Commission approved a Dominion Energy rate increase that adds roughly $16 a month to a typical residential bill starting in 2026. That's not a tax that you voted on, but instead the cost of building power infrastructure for AI companies that's passed directly to the consumer. And it's not just Virginia. Virginia's data center boom is raising West Virginia's electricity bills, too. But it's not just a power bill, it's your retirement. The data center buildout is being financed partly through special-purpose vehicles, also known as SPVs. These are debt structures that package up the risk and sell it on. Roughly $120 billion in SPV debt is out there right now. And that debt doesn't sit in a vault. It gets sliced into tranches and sold into funds, pension funds, 401(k)s, you know, the kind of default option retirement savings that you are automatically enrolled in. Oracle is currently facing a securities class action over an $18 billion bond that's tied to AI infrastructure. Analysts are raising the risk that if AI demand comes in below the hype, some of this infrastructure becomes a stranded asset, meaning the infrastructure gets built, the power demand doesn't materialize on the timeline promised, and the bonds lose value. Now, I want to be careful here because I haven't been able to source the actual data on exactly which pension funds hold this debt. But if you have a 401(k) or a workplace pension in a default fund, you are likely holding the companies that are building this infrastructure. You are funding the power plants that the thermodynamics says they can't finish fast enough. And if the productivity boom doesn't land on schedule, which the engineering says it can't, the correction will come out of your returns. So the physics say no, the grid say no, the water says no.
But let's play the same game where we pretend in this magical land that none of that exists. Let's give the AI industry every physical advantage: unlimited power, unlimited cooling, unlimited water. Let's just wave every constraint that I spent this video building. Even if you hand them everything, the economic model that they're using to predict the job replacement is still wrong. It's called the lump of labor fallacy. The assumption that there's a fixed amount of work in the economy, machines come along, they take a slice, and humans are left with less. I have to admit, it does feel intuitive, but it has been proven wrong every single time that someone has believed it for the last 200 years. Because what happens is when productivity rises, costs fall, and then demand expands. And when demand expands, new work appears that nobody had budgeted for. The classic example is ATMs. In the '70s and the '80s, when ATMs rolled out, everyone predicted that bank tellers would disappear. The machine does the job. So why would anyone hire a human? Here's what actually happened. ATMs made branches cheaper to run. Cheaper branches meant more branches, and more branches meant more tellers. Total headcount went up for two decades. But I want to be fair, as some people misuse this story. Teller numbers peaked around 2007 and have declined since, but not because of ATMs. They declined because of online banking, decades later, for different reasons entirely. And so the paradox is not proof that the job lives forever. Instead, automation reshaped the role and grew it for 30 years before it tapered, which is the opposite of your job will be gone in 18 months.
And here's the second economic error that's sitting right next to the first. It's called the rebound effect or the Jevons paradox. And it was discovered by William Stanley Jevons in 1965. What Jevons realized was that when steam engines got more efficient, total coal consumption went up, not down. Efficiency made steam power cheaper. Cheaper meant more industries used it, and more usage meant more coal. The savings of efficiency were eaten by growth. AI is the purest Jevons case I've ever seen. Let's imagine that AI makes financial analysis five times cheaper. Do you honestly think that companies will suddenly fire 80% of the analysts? No. No, they wouldn't. Instead, they will run 10 times more analysis on scenarios that they couldn't afford before. The market expands, and the pie grows. Both slices get bigger. And this, by the way, is already happening in AI. One analysis of Deep Six's reasoning model found that it could burn several times more energy per prompt than a rival, even while costing less. Deep Six's lower price just ended up driving more queries, more usage, and more total compute, eliminating any gains in efficiency and cost.
And so, I've been making the engineering case and the economy case. But here's what I find the most damning. The strongest evidence against the 18-month claim doesn't come from me. It comes from them. Anthropic, the company that makes Claude, one of the most advanced AI models on Earth, published research in March 2026. And their data shows something remarkable. The blue area on this chart is theoretical capability. So that's what AI could do. Computer and math jobs, 94% of tasks theoretically handleable by large language models. Office and admin, 90%. The blue reaches all the way out, and that's what the doomers point out when they say all jobs are gone. But now let's look at the red. That's what's actually happening today. Observed AI usage in real professional settings, measured from actual cloud data. For computer and math workers, the highest adoption category, it's 33%. Everything else is smaller. The gap between what AI can theoretically do and what AI is actually doing is enormous.
But it gets even worse for the CEO narrative. Anthropic's own researchers, Masanov and Mari, looked at the unemployment data for workers in the most AI-exposed occupations. And what they found was that AI has not increased unemployment in those jobs after two-plus years. The effect, they wrote, was indistinguishable from zero. Those are their words, not mine. Their own research team contradicting their own CEO in the same year. They did find one real signal: that hiring of 22 to 25-year-olds has slowed in the most exposed occupations. If you're a junior trying to break in, this is genuinely harder, and it's a point I will come back to later. But entry-level hiring has slowed is a completely different claim than all white-collar jobs are gone in 18 months. And it's not just Anthropic. MIT's Project NANDA found that 95% of enterprise Gen AI deployments show no measurable ROI, $30 to $40 billion spent. Only about 5% reach real production. And here's the irony. Dario Amodei, so the same man who originally said to anyone who would listen that AI would wipe out half of all entry-level white-collar jobs by May 2026, is already reframing. He's starting to talk about the Jevons paradox. The same mechanism I just explained. He's walking it back.
So, we have three physical walls, two economic fallacies, and the AI is getting more efficient argument as opposed to save it all. Just makes the problem worse by leading to an increased energy use. The engineering verdict is clear. The timeline is impossible. But here's the question I keep coming back to, and it's a very confusing one. If the engineering is this obviously wrong, why are the CEOs saying it? These are smart people, after all. They have armies of engineers who can do the exact math I just did. They know about turbine lead times and interconnection queues and cooling loads. So why the 18-month claim?
Let me be clear about what I think is really going on here. Fear is the product, and it's working exactly as designed. Think about it. Dario Amodei says AI could wipe out half of all entry white-collar jobs. What happens next? Anthropic raises a billion dollars. Sam Altman says most jobs will be automated within a decade. OpenAI's valuation climbs. Zuckerberg says AI agents will do the work of mid-level engineers. Meta's AI capex gets justified. There are four separate payoffs here, and I want you to see all of them.
First, fear funds the capex raise. The scarier AI sounds, the more capital flows in. A billion for Anthropic, tens of billions for the hyperscalers. Nobody writes a check that big for a tool that's kind of helpful. Second, it covers layoffs that are already planned because if you're a tech company that overhired during the pandemic and you need to cut headcount, "AI can do the job now" is a much better headline than, "Hey guys, we made a strategic error and we're firing you to protect our margins." Third, it suppresses wages. If you are a junior analyst and you believe an AI agent is about to take your job, you don't ask for a raise. Heck, you don't even negotiate. You are grateful to be employed. The threat of replacement is cheaper than the replacement itself. And fourth, and this is the one that perhaps I don't really see being talked about, it lifts the multiple. AI company valuations are priced on the expectation of transformative displacement. The more jobs AI is supposed to eliminate, the bigger the total addressable market. The higher the multiple, the richer the founders. The narrative is the valuation. And so I have to give it to them. The narrative is working. It's just not working for me and for you. It's working on your wages. It's working on your energy bill. It's working on your pension. You are paying for the power plants that the thermodynamics says they can't build fast enough.
Now, I want to be fair because there's a serious case for the defense, and I'm not going to dodge it. I'm not telling you AI won't change anything because it will. I'm not telling you that no one loses their job. They will. The real signal in the data, and this comes from Anthropic's only labor market research, is that hiring of 22 to 25-year-olds has slowed in the most AI-exposed occupations. Tech entry-level hiring is down 30 to 50% in 2025. Roughly 55,000 US layoffs have been attributed to AI. Banks are cutting analyst roles. That is real. If you are a junior, if you're trying to break in, this is genuinely harder than it was 2 years ago. And I'm not going to pretend that it isn't happening. But hiring has slowed for some entry-level roles is a completely different claim than all white-collar jobs will be gone in 18 months.
And let me answer the bad-faith objection before it arrives. You know that I'm claiming to know what CEOs really think. I'm not. I don't need to read anyone's mind. I'm looking at structural incentives. Because if you are the CEO of an AI company, a terrifying narrative about job displacement directly increases your valuation, your fundraising, and your pricing power. You don't need to be acting in bad faith to benefit from a story that happens to make you richer. The incentive is enough.
Now, let me be honest here. There are wild cards. Because if someone invents a fundamentally new AI architecture, something completely different that doesn't scale the same way against GPUs, well, that changes everything. And that's something I can't predict. Small modular reactors, perhaps, could accelerate grid capacity. Power breakthroughs can collapse the generation constraints that I've basically built half of my arguments in this video on. Any of those will change the math. Could they arrive in time for the 18-month timeline? Well, I think it's improbable, but I could be wrong. And here's how you know if I am wrong. Don't watch the press releases. Don't watch the podcast appearances. Instead, watch the interconnection queues. Watch the turbine order books. Watch the water permits. If those numbers start dropping, so if grid connection times fall from say 5 years to one, or if turbine lead times halve, if data center permits are being approved faster than they are filed, then the physical constraints are loosening and the timeline might be real. Until then, well, there's just no way around thermodynamics, I'm afraid.
So, I want to leave you with two things that are both true at the same time. AI is real and is genuinely transformative. It will change how people work, what work gets done, and who does it. That's not in dispute. I'm not a luddite and I'm not in denial. And the 18-month total replacement story, the ones that the CEOs are selling on podcasts and at conferences, is a narrative that's carefully built to move capital, discipline workers, and lift their company share price. Believing the first one does not require you to fall for the second. You see, the reason I made this whole long, nerdy video is that the fear only works while it stays abstract. The moment you make it physical, so the moment you ask where the power comes from and when it arrives, the entire narrative and timeline falls apart in your hands. And so, here's the thing I want you to keep. The next time someone tells you a job category is dead, whether it's engineers, analysts, lawyers, anyone, ask the one question that they can't fake: Where the power is coming from and when does it land? If they can't answer that question, then they're just selling you a feeling and not describing the future of the workforce. If this video gave you a sharper way to think, please hit the like button. And if you want to see more detailed breakdowns like this from an engineer's perspective, do subscribe. I'm out. Bye.