Cheap Tokens, Expensive Consequences

It's Saturday, May twenty third. On Wednesday, Mark Zuckerberg fired 8,000 Meta employees and moved 7,000 more into AI-focused roles. His memo said "success isn't a given" in the AI race. The same week, Meta reported record quarterly revenue of over fifty six billion dollars and raised its AI spending to a hundred and forty five billion. Intuit cut seventeen percent of its workforce the same day. Cisco cut 4,000 last week. Over a hundred and ten thousand tech workers have been laid off in 2026 so far, many with AI cited as the reason. This got me thinking about a pattern I can't stop seeing. In 2012, an Uber ride from Manhattan to JFK cost about twenty five dollars. The actual cost of that ride was closer to fifty. Venture capital covered the difference. For nearly a decade, investors paid for roughly half of every Uber ride in America. The goal wasn't to run a profitable business. It was to make the price so low that taxi companies, which actually had to charge what rides cost, couldn't compete. Thousands of cab companies went bankrupt. Millions of riders became dependent on the app. Then Uber went public, the subsidies stopped, and fares rose more than eighty percent. The playbook has a name: predatory pricing. Subsidize a product until the competition dies, then raise prices once consumers have nowhere else to go. The same playbook is now running in AI. Companies are firing engineers because AI tokens are cheap. The tokens are cheap because AI companies are losing billions to win market share. What happens when the subsidies stop and the engineers are gone? Uber's strategy was never a secret. Between 2010 and 2018, the company raised over twenty four billion dollars in venture capital and used it to subsidize rides below cost. Passengers paid about forty to fifty percent of what each ride actually cost. The rest came from investors. The math was simple: keep prices artificially low long enough to bankrupt the competition, build network effects that make it difficult for competitors to enter, then raise prices once you own the market. It worked. Between 2010 and 2020, over a hundred taxi companies in the United States filed for bankruptcy or ceased operations. Medallion values in New York City, which had traded for over a million dollars each, crashed to under two hundred thousand. Drivers who had mortgaged their homes to buy medallions lost everything. Several took their own lives. The public didn't notice because their rides were cheap and convenient. The destruction happened to people they never met. Once Uber went public in 2019, investors demanded profitability. The subsidies ended. Fares surged. Uber also introduced algorithmic pricing that charges you based on how badly you need the ride: time of day, weather, how far you are from a transportation alternative, and reportedly even your phone's battery level. The platform that won customers with impossibly low prices now extracts maximum revenue from customers who have no other option. The taxi companies that could have provided competition are gone. The AI industry is running the same playbook right now, and the numbers are strikingly similar. OpenAI is not profitable. The company doesn't expect to break even until 2029 or 2030. Anthropic has raised over thirty billion dollars and is targeting breakeven around the same timeframe. Both companies have been aggressively cutting API prices to win market share. Token costs have dropped by over ninety percent since GPT-3 launched. This isn't happening because the technology got ninety percent cheaper. It's happening because AI companies are using investor money to price below cost, exactly the way Uber priced rides below cost. The layoffs have already started. Klarna went from 5,000 employees to roughly 3,000, with the CEO publicly targeting 2,000. Their AI chatbot handles the equivalent workload of 700 customer service agents. Duolingo declared itself "AI-first" and stopped using contractors for work AI could handle. Microsoft cut roughly 23,000 positions across 2025 and 2026 while spending over a hundred billion dollars on AI. Salesforce laid off 4,000 customer support roles, saying AI could do fifty percent of the work. The World Economic Forum's 2025 report found that forty one percent of employers expect to downsize by 2030 due to technology. An outplacement firm estimated 55,000 job cuts in 2025 were tied directly to AI adoption. Companies are making these cuts based on current AI pricing. A customer service agent costs forty to sixty thousand dollars a year. An AI chatbot handling the same volume costs a fraction of that at today's token prices. The math looks obvious. Fire the humans. Keep the bots. But today's token prices are subsidized. And the subsidies are already ending. In April 2026, OpenAI increased the cost of its flagship GPT-5.2 model from $1.25 per million input tokens to $5.75. That's a 360 percent increase in a single pricing update. Anthropic moved its enterprise Claude offering from fixed pricing to dynamic usage-based billing, which industry analysts estimate could double or triple costs for heavy users. GitHub Copilot, the AI coding assistant owned by Microsoft, announced it would switch from flat-rate subscriptions to usage-based billing starting June 2026. One developer reported their projected monthly cost jumping from about sixty seven euros to nine hundred and sixty six euros under the new model. These aren't anomalies. They're the beginning of phase two. The AI companies subsidized tokens to get you hooked, the same way Uber subsidized rides. Now that companies have restructured their operations around AI, fired human workers, and built workflows dependent on specific AI platforms, the pricing power shifts to the provider. Switching to a different AI model isn't as simple as changing an app on your phone. Companies have fine-tuned models on their proprietary data, built custom integrations, trained their remaining employees on specific tools, and written internal processes around particular AI capabilities. The switching costs are enormous. And the AI providers know it. The contrarian case deserves honest treatment. There are real differences between Uber and AI. Open-source models like Meta's Llama and Mistral provide alternatives that didn't exist in the taxi market. Competition between OpenAI, Anthropic, Google, and a growing number of Chinese providers like DeepSeek may keep prices lower than a true monopoly would. Hardware improvements from Nvidia's next-generation chips are expected to reduce inference costs significantly. And the efficiency gains from AI are real: companies that use AI well are measurably more productive, not just cheaper. But those differences may not matter as much as they appear. Open-source models require significant engineering talent to deploy and maintain, and companies that have fired their engineers may not have the people left to run them. Competition between providers helps on price but doesn't solve the dependency problem: if your entire customer service operation runs on one provider's API and that provider raises prices, you can't switch overnight. The engineers who understood your old systems aren't waiting by the phone to come back. The Klarna story is the one to watch. The company went from 5,000 employees to 3,000 and bragged about AI replacing 700 agents. Then reports emerged that Klarna was reassigning engineers and marketers to customer support because the AI wasn't handling complex cases well enough. The company that had been the poster child for AI-driven headcount reduction quietly discovered that some of the work it had eliminated still needed to be done by humans. By then, many of those humans had moved on. So if this comes up in conversation, here's how to think about it. Uber subsidized rides below cost for a decade, bankrupted the taxi industry, and then raised prices eighty percent once consumers had no alternative. AI companies are running the same playbook with tokens: pricing below cost to win market share while companies fire employees based on those artificially low prices. The price increases have already begun. OpenAI raised flagship model prices by 360 percent. Anthropic moved to dynamic pricing. GitHub Copilot went from flat-rate to usage-based and one developer's cost jumped from sixty seven to nine hundred sixty six euros per month. Companies that fired their engineers to save money on AI may find themselves paying more for AI with fewer people who know how to manage the transition. The playbook worked for Uber. Whether it works for AI depends on whether the competition and open-source alternatives are strong enough to prevent the same kind of lock-in. If they are, the price increases will be contained. If they aren't, a lot of companies just traded a workforce they controlled for a platform they don't. Stay informed, stay curious, and we'll see you tomorrow.