Something doesn't add up. AI companies spent years promising that intelligence would become infinitely cheap as it scaled. Now, at the moment adoption should be accelerating, the industry is retrenching—and the signs are everywhere.
GitHub paused new Copilot signups this week, tightened usage limits, and removed access to its most expensive models. Anthropic restricted Claude Code to its $100-per-month tier, eliminating access for subscribers of its $20 plan. OpenAI's CFO Sarah Friar has spent months publicly discussing compute shortages as justification for product decisions including shutting down Sora. These aren't isolated technical hiccups—they're the visible fractures of an economic model built on sand.
The core contradiction is stark. The entire AI industry oriented its growth strategy around a single assumption: compute costs would fall forever. GPU performance per dollar would improve, data centers would get cheaper to build, and electricity would remain abundant. Every pitch deck, every pricing model, every "costs will decrease at scale" promise rested on this foundation. The foundation is crumbling.
Meta announced this week it is laying off 10 percent of its workforce—roughly 15,000 people—in part to fund the AI infrastructure buildout it cannot otherwise afford. "We need to allow us to offset the other investments we're making," the company told remaining staff. The investments are data centers and the equipment to run them. This is the equation the industry has been hiding: AI growth requires infrastructure that costs more than the revenue it's generating.
The hardware crunch is already bleeding into consumer markets. The same 2TB external SSD that cost $159 in late 2024 now costs $575—a tripling in under eighteen months. RAM and graphics cards have followed the same trajectory. Apple's manufacturing partners are reportedly struggling to secure chipmaking capacity because AI clients have cornered the available supply. Consumer electronics will get more expensive, and the scarcity won't end soon.
Data centers are consuming so much electricity that residential power bills are spiking in regions with high concentrations. Some municipalities and states have begun rejecting new data center applications entirely. Water consumption—critical for cooling—is creating similar conflicts. Communities are pushing back against infrastructure they didn't choose but must fund.
The price increases already visible—20 to 37 percent for software with embedded AI tools across Microsoft 365, Notion, Salesforce, and Google Workspace—represent the opening salvo. GitHub Copilot charges $19 per month while consuming roughly $90 per month in compute per user. That $71 monthly loss per subscriber was acceptable when growth was the priority and capital was cheap. It is not acceptable when growth stalls and capital costs rise.
The subsidy era is ending because the subsidy model was always fiction. GPU manufacturing timelines cannot compress on demand. Power grid expansion cannot be accelerated by venture funding. The industry promised abundance and delivered scarcity. Now it must charge what the product actually costs—or stop selling it.
The next six months will separate companies with viable economics from those that built empires on a lie. The companies that survive will be those that can price accordingly. The industry-wide reckoning is not coming. It is here.