Analysts warn an 'AI compute crunch' is reshaping labour markets, device economics and energy demand
- GitHub paused new Copilot Pro, Pro+ and Student signups on April 20 as AI coding demand pushed it to tighten usage limits, a concrete sign that cheap, flat-rate compute is getting harder to sustain. - Anthropic expanded its Google and Broadcom compute pact on April 7, with a Broadcom filing pointing to 3.5 gigawatts and Anthropic calling it its biggest compute commitment yet. - The squeeze is spreading from software into power grids and capital spending, with new U.S. data centers needing 44 gigawatts by 2028 against about 25 gigawatts of likely supply. (ft.com)
Artificial intelligence runs on specialized chips in data centers, and those chips now look less like abundant cloud capacity and more like scarce industrial equipment. (404media.co) (morganstanley.com) That scarcity showed up in a consumer product on April 20, when GitHub said it would pause new signups for Copilot Pro, Pro+ and Student and impose tighter usage limits on individual plans. (github.blog) It showed up again on April 7, when Anthropic said it had expanded its compute agreement with Google and Broadcom to power Claude, with TechCrunch reporting a Broadcom filing tied the deal to 3.5 gigawatts of capacity. (techcrunch.com) The basic problem is simple: model makers spent years selling access below the full cost of running the systems, then usage jumped as coding agents and other heavy workloads moved from demos into daily work. (404media.co) (techbrew.com) When compute gets expensive, companies change behavior before they change the models. They cap plans, move premium features into higher tiers, steer workloads to cheaper chips, and measure cost per task instead of just model quality. (github.blog) (404media.co) (techcrunch.com) That is why techniques like batching, caching and routing matter. In plain terms, batching groups many requests together, caching reuses answers that were already computed, and routing sends only the hardest jobs to the most expensive models. (404media.co) The crunch is also changing labor math. 404 Media reported that some startups are already spending on tokens for tools from Anthropic and OpenAI instead of adding headcount, tying payroll decisions directly to chip and inference costs. (404media.co) It is changing device economics too. 404 Media reported sharp price increases for consumer storage and graphics hardware, and linked those moves to manufacturers shifting capacity toward artificial-intelligence demand. (404media.co) And it is colliding with the electric grid. The Financial Times reported that five U.S. facilities due from 2026 would each draw at least 1 gigawatt, while S&P Global Energy estimated new data centers would need 44 gigawatts by 2028 against about 25 gigawatts of available new power capacity. (ft.com) Morgan Stanley said in March that about $2.9 trillion in global data-center construction costs are projected through 2028, with demand for compute still far above supply. The compute crunch is no longer just a chip story; it is now showing up in software pricing, hiring plans and electricity forecasts. (morganstanley.com)
