This video discusses the escalating costs associated with AI implementation in large tech companies. It argues that the problem is not the AI technology itself, but rather how its usage is measured and incentivized, leading to a focus on consumption metrics (like tokens) over actual value creation. The video uses examples from Uber, Meta, Microsoft, and Nvidia to illustrate how this "cost problem" is manifesting and questions the sustainability of current AI investment strategies.
In April 2026, Uber's chief technology officer made a quiet admission. The company had already burned through its entire annual AI budget, the entire annual AI budget by April, a budget size for 12 months gone in four. The details make it even worse. Uber had rolled out Anthropic's Claude code, an AI coding tool, to its engineering team in December 2025. By March, adoption had jumped from 32% to 84% of the company's 5,000 engineers. 95% of those engineers were using AI tools every month. 70% of committed code was AI generated. Individual engineers were costing between $500 and $2,000 per month in AI usage on top of their salaries. And the thing is, this didn't happen by accident. Uber had actively incentivized it. The company built internal leaderboards that ranked engineering teams by how much AI tooling they used. Essentially, the more you consume, the higher you ranked. The higher you ranked, the better you looked. A 12-month budget incinerated in four because the company built a system that rewarded spending it. This is the corporate equivalent of giving every employee a company credit card with a leaderboard for who spends the most and then acting surprised when the bill arrives. I'm ML, I have a PhD in computer science, and I analyze AI developments to understand what's actually happening beneath the hype. And what's actually happening right now is that the AI industry has developed a cost problem so significant that even the companies selling the technology are admitting it just doesn't add up. But Uber is just the beginning. In this video, I'm going to show you how Meta turned AI usage into a competitive sport, why Microsoft killed the tool its own engineers preferred, and why the CEO of Nvidia and his own vice president are actively making opposite arguments. And then, I'm going to tell you where I think all of this compute should actually be going because right now the entire industry is pointing it in the wrong direction. It is the nine circles of absurdity, my friends, and we're here for it.
Before I get into what happened, I need to explain one concept that makes the rest of the story legible super quickly, just like 15 seconds tops. A token is the basic unit of data that an AI model processes. It's roughly one word, sometimes part of a word. Every time you use an AI tool, you ask it a question, have it write code, run an AI agent, you're consuming tokens, and tokens cost money. The more complex the task, the more tokens it consumes. A simple chatbot exchange might use a few thousand tokens. An AI agent performing a multi-step coding task can burn millions.
Here's where it gets interesting. The cost of individual tokens is falling, fast. A recent report from Gartner found that by 2030, the cost of actually using a frontier AI model, sending it a prompt, getting a result back, also called running inference by the way, will cost providers over 90% less than it did in 2025. That sounds like good news. It's not because agentic AI models, the ones that don't just answer a question, but actually go and do things autonomously, like writing code across multiple files, booking flights, or managing workflows without a human guiding every step, require between five and 30 times more tokens per task than a standard chatbot. So, the unit price drops, but the number of units explodes. Goldman Sachs projects that agentic AI could drive a 24-fold increase in token consumption by 2030, reaching 120 quadrillion tokens processed per month. This is a well-known economic pattern called the Jevons paradox. When a resource becomes more efficient to use, people don't use less of it, they use dramatically more. I covered this concept at length in a previous video where I applied it to the relationship between AI deployment and human employment. The same paradox applies here except instead of jobs, the resource being consumed is compute. Cheaper tokens don't mean cheaper AI, they mean more tokens. As Gartner senior director analyst Will Summer put it, chief product officers should not confuse the deflation of commodity tokens with the democratization of frontier reasoning. That sentence is doing an extraordinary amount of work and I suspect most of the people who need to hear it are not reading Gartner reports. Oh, by the way, I've just recently launched channel memberships. The videos stay free always. Memberships and Coffee are just a way to support the work if you want to and are able to. No pressure, no payroll, no missing out. The link is down below.
So, with all of that context, what happens when you take a resource that's getting cheaper per unit but more expensive in aggregate and then build a corporate culture that rewards consuming as much of it as possible. You get Meta's Clodonomics. In early April 2026, The Information reported on an internal Meta leaderboard built voluntarily by an employee on the company's intranet called Clodonomics. It tracked AI token consumption across more than 85,000 Meta employees. It ranked the top 250 power users. It awarded gamified titles like token legend for the highest consumers, session immortal for extraordinary session duration, cash wizard for efficiency in reuse. In a single 30-day period, Meta employees collectively consumed 60 trillion tokens. The top individual user burned through 281 billion tokens averaging 9.36 billion tokens per day. At Anthropic's public API prices, the total would have cost approximately $900 million. Even at steep enterprise discounts, the figure is still staggering. And here is the detail that tells you everything you need to know about whether this consumption was producing proportional value. Some employees were leaving AI agents running idle for hours doing nothing in order to inflate their position on the leaderboard. They invented a way to get a robot to do nothing on their behalf. Is that automation or delegation of idleness? An AI version of dolce far niente, maybe? The leaderboard was taken down after press coverage, but Meta CTO Andrew Bosworth had publicly endorsed the underlying logic. He said his best engineer was spending the equivalent of his salary in tokens, but was five to 10 times more productive as a result. "It's like this is easy money," Bosworth told Forbes. "Keep doing it. No limit." And Meta is not alone in this. Amazon has reportedly pushed its employees to token max, a term that has entered Silicon Valley vocabulary in 2026, meaning use as many AI tokens as possible. According to the Financial Times, Amazon employees have been running the company's internal AI tool on trivial tasks to inflate their token counts and climb internal rankings. The industry has invented a term for competitive AI waste, and the largest technology companies on Earth are participating in it. Somewhere an economy is writing a paper about this. The AI will probably summarize it wrong.
I want to pause here and name what this is because I think the absurdity of it can obscure the analytical point. These companies are measuring input, how many tokens are you consuming, as a proxy for output, how much value are you creating, and those two things have no guaranteed relationship. This is like rewarding your sales team not for how many houses they sold, but for how much petrol they burned driving to meetings. The top performer isn't the one who closed the most deals, is the one who drove the most miles. The leaderboard tracks fuel consumption, but nobody's striking revenue.
As somebody with a PhD in computer science who has worked in this field for over a decade now, I can tell you this pattern is not exactly new. Engineers, and I say this with a lot of love because I am one, have a tendency to get so enamored with the complexity of a system that they often lose sight of what the system is supposed to do. It's almost like being an artist. You want the code to be beautiful, you want the architecture to be elegant, you want the documentation to be pristine, and all of those things have value. I'm not saying they don't before somebody comes at me, but they are secondary to the fundamental purpose of software, which is to solve a problem, to do the thing, to work. Token maxing is the organizational version of that same impulse. It's optimizing for a metric that feels important, consumption, activity, usage, while completely decoupling from the question of whether any of that consumption is producing proportional value.
Meanwhile, Microsoft has been dealing with its own version of this problem, and the details reveal something slightly different but equally instructive. In December 2025, Microsoft opened up access to Anthropic's Claude Code alongside its own GitHub Copilot CLI. Both are AI coding tools that run in a developer's terminal and can write, edit, and manage code. For roughly six months, Microsoft engineers ran both tools side by side. The engineers preferred Claude Code. Internal comparisons reportedly showed it outperforming Copilot CLI on real engineering tasks. It became, as The Verge's Tom Warren put it, very popular, perhaps a little too popular. In May, Microsoft began canceling Claude Code licenses. Engineers in the Experiences and Devices division, the teams behind Windows, Microsoft 365, Outlook, Teams, and Surface were given until June 30th to remove Claude Code from their workflows and switch to Copilot CLI. The official justification was tool chain unification. The timing, June 30th, the last day of Microsoft's fiscal year, suggests cost-cutting was at least equally important. The specifics of the billing difference are very technical. If you'd like me to do a full breakdown of seat-based versus API-based pricing and why the distinction matters, I can make a separate video on that. But the short version is essentially, running Claude code means paying Anthropic both a per-seat license fee and API rate token cost on top. Routing Claude's models through Copilot CLI, which Microsoft owns via GitHub, eliminates the licensing layer and brings the billing in-house. It's cheaper and the money stays inside Microsoft's ecosystem. But the cost story obscures the more revealing detail. Microsoft ran a six-month experiment. Its own engineers chose the competitor's product, and Microsoft responded by killing the competitor's product rather than improving its own. That is a company optimizing for platform control over engineering quality. Whether that's the right strategic decision is debatable. That it's a cost-driven decision is not. And consider the timing more broadly. GitHub Copilot is simultaneously transitioning to usage-based billing on June 1st, moving from flat premium request counts to a token consumption model using AI credits. So, Microsoft is consolidating onto a single platform at exactly the moment that platform is switching to the same consumption-based pricing model that blew up Uber's budget. The cost problem is not being solved. It is being reorganized under a different logo. This is what the AI cost crisis actually looks like at the enterprise level. It is not a single dramatic failure. It is a thousand procurement decisions, each individually rational, that collectively reveal an industry spending faster than it can measure the returns.
And then, there is Nvidia, where the contradiction is so clear it almost doesn't need commentary. Jensen Huang, Nvidia's CEO, said in March 2026 that he would be deeply alarmed if a $500,000 engineer at his company was not consuming at least $250,000 in AI tokens. He framed this as a productivity imperative. The tokens make the engineer more valuable. Bryan Catanzaro, Nvidia's vice president of applied deep learning, told Axios, "For my team, the cost of compute is far beyond the cost of the employees." Same company, one executive says spend more on tokens, the other says the tokens already cost more than the people. Neither of them is wrong, by the way, and I think that's the part that most people miss when they look at contradiction like this and assume somebody must be lying or confused or stupid. They're not. They have different incentive structures. Huang is the CEO of a company that manufactures the GPUs that process every token consumed everywhere in the world. More tokens consumed means more GPUs needed means more Nvidia revenue. His incentive is to maximize demand. Of course, he wants engineers to burn through tokens. Catanzaro manages an actual engineering budget. His incentive is to deliver results within a cost envelope. Of course, he notices when the compute bill exceeds the payroll. In life, there are very few good and bad actors in situations like this. There are incentives, objectives, and paths to reach them. The problem isn't that Huang is wrong or that Catanzaro is wrong. The problem is that the use more signal from the top of the supply chain, from the company that profits most from consumption, cascades downwards through the entire industry. It reaches Meta, where it becomes Cloudonomics. It reaches Amazon, where it becomes token maxing on trivial tasks. It reaches Uber, where it becomes leaderboards that incinerate a year's budget in four months. The signal to consume originates from the people who sell what's being consumed. And at no point in this chain is anyone systematically measuring whether the consumption is producing proportional value.
At this point, the question that every comment section will ask is, of course, is this a bubble? I want to be very honest about this rather than satisfying for everyone. Nobody other than maybe God can call a bubble in real time. That is a near universal truth in economics and financial markets. You only know it was a bubble after it pops. People who confidently tell you this is definitely a bubble or this is definitely not a bubble are offering you certainty, which is comfortable and almost certainly wrong. What I can do is describe what the numbers look like. Combined 2026 capital expenditure from Amazon, Microsoft, Alphabet, and Meta is now pushing $740 billion, a 69% increase from 2025. There have been over 92,000 tech layoffs in 2026 so far, already outpacing last year's total. A 2024 MIT study found that AI automation is economically viable in only 23% of roles where visual tasks are primary. In the remaining 77% keeping human workers is still cheaper. But, and this is very important, AI generally works. The underlying technology produces real output. Claude code actually writes functional code. AI agents actually complete tasks. This is not a speculative technology with no demonstrated capability. The question is not whether AI works. The question is whether the rate of spending is sustainable given the current rate of return. And the honest answer is, I don't know. I don't think anyone does. These companies are investing at this scale because they see potential, and that is not exactly irrational. The last time humanity encountered a technology with this kind of transformative promise was the Industrial Revolution, and for all the human suffering it caused, which I covered in a separate video, it was from a purely technological standpoint one of the most productive periods in human history. I can understand why CEOs are going all in. Their investment thesis is a combination of fundamentals and human judgment. The fundamentals are real. AI can deliver a certain set of things, and that set is expanding. The judgment part carries a great deal of hope. Hope that this technology will reshape industries that we steam and electricity did. Would I personally be more conservative? Probably, but everyone has their own risk appetite, right? And I'm not going to pretend I have better judgment than people who are running these companies. What I will say is that hope is not a deployment strategy. And right now, the gap between the hope and the operational reality is being filled with leaderboards.
And there is another thing that nobody seems willing to say. Nobody's forcing these companies to spend $740 billion in a single year. They could invest a fraction of that, say 100 bill, measure what actually produces returns, scale the things that work, and ramp up from there. The technology is not going to vanish next quarter. The models are not going to magically untrain themselves overnight. There is no external deadline, but every company is spending because every other company is spending, and the fear of being the only one that under invested, the one that missed AI, is more powerful than the absence of proven returns at this scale. That is not a technology-driven investment strategy. That is FOMO herd behavior at $740 billion. And the history of herd behavior at scale is not exactly encouraging, regardless of whether the underlying technology is real or not.
But what I do generally believe is that the current deployment strategy, squeezing more consumption out of the same use case, which is software engineering, is not the path that justifies this level of investment in my eyes at least. I covered the other side of the cost equation in a previous video, companies that fired their human workforce, replaced them with AI, and watch the results collapse. If you haven't seen that one, I definitely recommend watching it next because it's the mirror image of what I've shown today. That video was about companies removing humans and AI failing to do the job. This one is about companies keeping the humans, adding AI on top, and discovering the AI costs more than the people it was supposed to augment. Both are deployment failures, and both stem from the same underlying problem. These companies do not have a coherent theory of what AI is supposed to do for them. They are adopting AI because everyone else is adopting AI. They are measuring consumption because consumption is easy to measure, and they are discovering one blown budget at a time that the absence of a strategy is itself the most expensive mistake they can make. The technology is not the problem. The technology works. The deployment of the technology, who decides how it's used, what's being measured, and whether anyone is checking if the output justifies the input, that's where every single one of these stories breaks down.
So, where should the computer actually go? Right now, Jensen Huang wants his engineers to burn $250,000 in tokens each. Meta built a leaderboard to see who could consume the most. Amazon told employees to token max. The entire demand generation strategy of the AI industry appears to be: take the people who already have AI tools and tell them to just use more. That is the wrong axis entirely. If the goal is to generate the kind of demand that actually justifies $740 billion in capital expenditure, the answer is not to squeeze more consumption out of software engineers who are already saturated. The answer is to point the compute at people who have unsolved problems. Universities are the obvious candidate, I think, and I say this as somebody who has worked in academia. Now, before the comments fill up with but this already exists, yes, it does. Nvidia has CUDA research grants, I'm aware. Google has academic partnerships. Microsoft has Azure credits for researchers. These programs are real and they are doing useful work, but they are a fraction of what's possible and they are dwarfed by the resources being poured into internal leaderboards and token maxing incentives. The question again is not whether university partnerships exist, it is whether the industry is serious about scaling them or whether they remain a rounding error next to the budget being spent encouraging Meta employees to compete for the title of session immortal. Give researchers in chemistry, physics, medicine, economics, and climate science free or subsidized access to frontier AI models at a scale that matches the ambition of the investment. Not for homework, I don't mean exams, but I mean for research legitimately. These are people who already know how to code, who already understand computational methods, and who are sitting on problems that generally exceed human cognitive capacity. The chemist trying to model molecular interactions across billions of configurations. The economist trying to simulate non-linear systems that the entire field still models with linear regressions. The medical researcher trying to identify drug interactions across data sets too large for any human team to process. The physicist running simulations that currently take months on existing hardware. I know this world. If somebody had given me a large allocation of free compute when I was working in academia, I would have been significantly more incentivized to experiment more, to try things that felt too computationally expensive to justify on a limited budget. And I'm far from unique in this, by the way. Everybody in academia is incentivized to publish papers, to discover, to push their field forward. Machine learning research is already through the roof. Everybody wants to build a new thing, but the real opportunity isn't just within computer science. It is in empowering the other scientists, the chemists, the physicists, the economists, the medical researchers, and I'm not trying to exclude anyone here, by the way. They are already excellent at coding. They already understand data. They just need the compute to do things that were previously impossible. If you give researchers compute, they will use it, not to climb a leaderboard, but to try to solve something. And the breakthroughs that emerge from that process would create the organic, sustainable demand that the industry desperately needs. Demand rooted in genuine discovery, not in gamified consumption. A new drug target identified through AI-assisted molecular modeling doesn't just use tokens, it creates an entire industry of follow-on research, clinical trials, and applications that need more AI. A breakthrough in climate modeling doesn't just use compute, it generates demand for continuous monitoring, prediction, and simulation at a scale that would keep GPU manufacturers happy and busy for decades. That is how you build sustainable demand, not by telling software engineers to consume harder, but by embedding the technology in as many fields as possible, and letting the problems themselves generate the usage.
And I want to be clear about what I am not saying here. I am not saying this should be done to displace researchers or make their expertise obsolete. I'm saying the exact opposite. The whole point is to harness the technology to empower people who are already doing important work, to supercharge human expertise, not replace it. The technology works best when it's pointed at a problem by somebody who understands the problem. A leaderboard doesn't understand much, but a researcher definitely does. And that distinction matters. If the path to justifying $740 billion in annual capital expenditure is "make engineers climb leaderboards," then yes, the spending probably isn't sustainable. But if the path is "fund breakthroughs in protein folding, climate modeling, drug discovery, and economic simulation that couldn't happen without this technology," that is not a bubble. That is an investment. The difference between those two outcomes is not the technology. The tech is the same. The difference is where we pointed. Right now, the industry is pointing it at leaderboards, at AI token legend badges. It engineers running agents idle to inflate a number that nobody has connected to a business outcome. The people who built these leaderboards sat in a room and decided that measuring consumption was a strategy. Another meeting can happen. A different strategy can emerge. That's not failure, it's just course correction. But it requires the humility to look at the data, recognize that burning a 12-month budget in four months while your employees gain the metrics is not exactly a sign of success and just change direction. Uber CTO is already going back to the drawing board. That's encouraging. It would be even more encouraging if the drawing board had existed before the budget was gone. But hey, whatever, it's cool. We can all move on to live another AI day, right? Anyway, the cost of using AI is only half of the equation. What happens when companies don't just add AI on top of their workforce, but remove the workforce entirely and let AI run the show? The results are not what the brochure promised. That's the one that I would watch next. Thanks so much for watching this one. Subscribe and I'll see you all in the next one.
