AI will not replace human labor, and the reason comes down to simple economics \u2014 human labor is cheaper than the AI tokens needed to replace it.<\/p>\n\n\n\n
We already have proof of this. Microsoft made its employees stop using Claude Code because it got too expensive. And, if AI is too expensive for Microsoft, who can afford it?<\/p>\n\n\n\n
Companies are blowing through their annual AI budgets in a matter of months.<\/p>\n\n\n\n
Take Uber for example \u2014 it burned through its entire annual AI budget in just 4 months<\/a>! The company has since put a daily cap on how much AI their employees can use just to manage costs. AI became a threat to their bottom line.<\/p>\n\n\n\n Meta employees burned through 60 trillion tokens in one month<\/a>. We don\u2019t know how much this cost Meta, but at Anthropic\u2019s standard public API list rates, this would amount to $900 million.<\/p>\n\n\n\n However much it actually cost, it was enough for Meta to implement strict spending controls after internal warnings revealed that employee AI usage was on track to cost billions of dollars.<\/p>\n\n\n\n Microsoft provides one of the most telling examples. In December 2025, Microsoft\u2019s Experiences and Devices team got access to Anthropic\u2019s Claude. Six months later, Microsoft began canceling Claude Code licenses<\/a>.<\/p>\n\n\n\n Officially, this was framed as a step toward \u201ctoolchain unification,\u201d but internal sources quickly leaked the real reason: It cost too much, and the company needed to stop the bleeding before fiscal year-end.<\/p>\n\n\n\n Right now, general users can access AI chatbots like ChatGPT for free because AI companies use external investment money to subsidize the cost for users.<\/p>\n\n\n\n These companies are actually operating at a massive loss. And it\u2019s not just their free plans that are costing them \u2014 they also lose money on paid-plan heavy users. A ChatGPT Pro user, who pays $200 for the monthly subscription and gets max usage out of it, gets an amount of compute that would cost around $14,000 using API-equivalent values<\/a>.<\/p>\n\n\n\n To put into perspective just how unprofitable OpenAI is, in 2025, they made $13.07 billion in revenue. This cost them $34 billion, which means OpenAI lost $20.92 billion<\/a>. They can afford this only because external funding lets them keep the company going and subsidize users.<\/p>\n\n\n\n But that external funding is starting to dry out. The AI bubble has already burst in Korea<\/a>, with many suspecting that the US is just lagging behind. On paper, AI investment in the US looks stronger than ever, but the picture changes if you exclude circular financing<\/a>, where the same handful of companies invest back and forth in each other.<\/p>\n\n\n\n Not only that, OpenAI is looking to go public<\/a>, and if it does, it\u2019ll have to answer to shareholders, who won\u2019t let the company keep hemorrhaging money. And OpenAI isn\u2019t the outlier. Everything I\u2019ve written about it \u2014 the unprofitability, the cost subsidizing, and the intention of going public \u2014 applies to Anthropic, as well.<\/p>\n\n\n\n So, even though ChatGPT is free to the general public, the free version will likely be heavily limited, while the paid versions will get more expensive.<\/p>\n\n\n\n AI companies are constantly building newer and more capable frontier models. They do this by training them on larger data sets. This means that training newer models takes more compute, more power, more water, more everything! Costs rise exponentially<\/a>, while performance gains diminish.<\/p>\n\n\n\n And, even though the AI companies are operating at a loss, the cost of using AI for companies that pay for it keeps going up.<\/p>\n\n\n\n Technically, the price per token has gone down dramatically since 2023, but agentic workflows consume 50 to 500 times more tokens<\/a> per response. This means that companies spend much more on tokens per response now than they did a year ago, despite tokens being cheaper.<\/p>\n\n\n\n The end of the AI subsidy means AI will cost users and companies more. And we can surmise these subsidies are nearing their end because of compounding factors that drive up both the costs of developing AI and the profitability demands for AI companies \u2014 unprofitability, going public, increased costs of frontier models, and AI funding slowing down (once you exclude circular financing).<\/p>\n\n\n\n When AI companies completely ditch their subsidies, 2 things will likely happen: AI companies will switch to token-based billing, and the costs of tokens will go up.<\/p>\n\n\n\n Let\u2019s take a closer look at both of these possibilities. <\/p>\n\n\n\n AI companies are subsidizing flat-rate users more than users who pay per token, which is just unsustainable. Just remember the example of how a ChatGPT Pro user can get $14,000 worth of compute for a $200 subscription.<\/p>\n\n\n\n Moving all services to usage-based billing will let AI companies align costs directly with consumption. <\/p>\n\n\n\n Per-token prices will go up by whatever amount the companies are subsidizing to reflect true operational costs. <\/p>\n\n\n\n At the same time, the effective cost for companies will go up even more because agentic workflows simply require more tokens per response. As a result, companies will both pay more per token and spend more tokens per response. <\/p>\n\n\n\n And, since it\u2019s literally impossible for LLMs to completely stop hallucinating<\/a>, not every response will be correct or acceptable, compounding the costs even further.<\/p>\n\n\n\n Comparing raw API token prices to hourly wages for human labor just isn\u2019t the right way to calculate ROI. That would be like looking at the cost of bricks and using only this metric to calculate how much it would cost to build a house. Just like there\u2019s more to building a house than buying bricks, there\u2019s more to using AI in business than looking at the cost of API tokens. <\/p>\n\n\n\n For a truly economically relevant comparison, researchers argue that you would have to look at total cost of ownership (TCO) per acceptable outcome<\/a> (not all outcomes will be acceptable because of hallucinations), which includes things like:<\/p>\n\n\n\n I could go on, but I don\u2019t want to belabor the point: AI costs companies far more than the API listing would indicate.<\/p>\n\n\n\n A human employee brings their own eyes, hands, and common sense to the job, and they can be held responsible for their mistakes \u2014 for AI agents, you have to spend money to replicate and enable these faculties, and even then, your company will be responsible for its inevitable hallucinations.<\/p>\n\n\n\n An MIT FutureTech Study<\/a> from January 2024 \u2014 when AI cost users less than it does now \u2014 found that human workers were more economically attractive to companies. For vision tasks specifically, it showed that human employees remained substantially more affordable for 77% of tasks.<\/p>\n\n\n\n In his paper The Sample Macroeconomics of AI<\/em><\/a>, MIT economist Daron Acemoglu demonstrated that the costs of AI deployment far outweigh the savings for hard-to-learn or context-heavy tasks. This once again shows that human labor is cheaper than AI.<\/p>\n\n\n\n Of course, there are jobs where AI can and does cost less. A fintech company replaced the workload equivalent of 700 full-time support agents with AI assistants to great effect, dropping operational costs per transaction by 40%<\/a> and reducing response times to under 2 minutes on average.\u00a0<\/p>\n\n\n\n A noticeable trend across all positions is that AI can \u2014 and has \u2014 successfully replaced only entry-level positions responsible for simple tasks like clerical claims processing and data extraction.<\/p>\n\n\n\n But even AI support chatbots aren\u2019t foolproof, and your company will be legally responsible for what they tell customers on your website, including hallucinations. This has been the case since Moffatt v. Air Canada<\/a> in 2024, when the airline\u2019s AI chatbot invented a bereavement travel fare policy, and the company tried to deny the refund outlined by that policy, claiming the AI was a \u201cseparate legal entity responsible for its own actions<\/em>.\u201d The judge disagreed, and the airline did indeed pay Moffatt what the AI told him when it hallucinated.\u00a0<\/p>\n\n\n\n Because of cases like this, insurance companies have been making AI insurance exclusions<\/a>, with some carriers making absolute AI exclusions that disclaim coverage for any claim, loss, or liability tied to AI. All of this makes the effective cost of using AI higher, since you need to account for higher insurance costs and the human labor required for oversight to prevent these risks.<\/p>\n\n\n\n FURTHER READING<\/strong><\/p>\n\n\n\n For an idea of what careers AI won\u2019t be able to replace, read this article:<\/p>\n\n\n\n Companies never wanted to replace human labor with AI because they hated humans or cared for AI. They wanted to do it because OpenAI and Anthropic promised that replacing human labor with AI would be cheaper<\/a>. But the business case for this simply doesn\u2019t hold under financial scrutiny. Even Sam Altman has admitted that AI has had a smaller impact on job replacement than he thought<\/a>.<\/p>\n\n\n\n Not only is AI more expensive than human labor (and expected to get more expensive), but it\u2019s also unreliable and requires human oversight. <\/p>\n\n\n\n Despite all of this, AI is still a valuable workplace tool. The solution isn\u2019t to scrap it, but to stop attempting to replace human labor at scale. Companies should identify specific, high-value use cases where AI is demonstrably cost-effective. <\/p>\n\n\n\n If you don\u2019t need an AI agent, use an AI chatbot \u2014 their cost scales linearly and predictably, and here the drop in raw token prices actually matters. Text summarization, information retrieval, content drafting and editing \u2014 these are tasks that chatbots can do just fine. <\/p>\n\n\n\n When applied judiciously, AI can serve as an excellent tool for targeted augmentation of human work. This way, companies can improve operational efficiency without harming their long-term fiscal health.<\/p>\n","protected":false},"excerpt":{"rendered":" Learn what hidden costs AI has to understand why it\u2019s more expensive than human labor, and why the gap between them is only expected to widen.<\/p>\n","protected":false},"author":10,"featured_media":5209,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[12],"tags":[],"class_list":["post-5203","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technology-corner"],"acf":[],"yoast_head":"\nIf AI is so expensive to use, how come it\u2019s free?<\/h2>\n\n\n\n
AI is already getting more expensive for everyone<\/h2>\n\n\n\n
What the end of the AI subsidy means for companies<\/h2>\n\n\n\n
Possibility #1: AI companies will switch to token-based billing<\/h3>\n\n\n\n
Possibility #2: The cost of tokens will go up<\/h3>\n\n\n\n
The hidden costs of AI<\/h2>\n\n\n\n
\n
Human labor is both cheaper than AI (and needed for AI)<\/h2>\n\n\n\n
\n
AI as an enabler, not a replacement<\/h2>\n\n\n\n