The AI bubble didn’t explode — it fizzled. Reports say data centers are only using about 30% of their AI hardware, hype is cooling, and more companies are moving to smaller local models while efficient systems out of China keep cutting the need for endless new data centers. In this video I break down idle GPUs, pointless AI-in-everything gadgets, why my own AI use crashed to about 5%, and why the power grid and the trades still decide how far automation can go.
Idle hardware and the build bottleneck
Recent reports put AI hardware utilization around 30%, which means roughly 70% of GPUs and RAM are sitting in warehouses, idle, or not even powered up. Chips can go obsolete before a data center finishes installing them. These buildings don’t appear overnight — they need land, plans, and skilled people — and the labor shortage is slowing the buildout companies wanted. I covered that worker gap earlier in The Real AI Data Center Problem: No Workers.
Demand didn’t keep up with supply
Usage isn’t collapsing overnight, but capacity is being added so fast that flat or slightly soft demand looks like a crash. Double your compute while the market stays flat and a 1% dip starts to look like you lost half your demand. Meanwhile a lot of businesses already found the AI they need. They don’t need another wave of megawatt-scale clusters for AGI theater.
Local models beat token bills
More companies are deciding big-tech APIs are a risk and a cost center. A capable local box — think on the order of 64GB of RAM and a couple of GPUs — can pay for itself in a few months versus token spend on OpenAI, Anthropic, or Grok. Most small firms never need image, video, or audio generation, so a smaller task-focused model is enough and uses far less compute.
AI in your toaster isn’t progress
Stuffing AI into cars, washers, toasters, and fridges doesn’t make daily life better when the features aren’t necessary. As tools get good enough to finish a job and get out of the way, people use chat less. My own Grok chat usage fell from heavy daily use to roughly 5%, while automated bot tasks handle website checks, plugins, security, and simple health tracking — then I’m done for the day.
Trades, the grid, and efficient China models
AI can crush script writing and ad copy; it can’t rebuild an engine or wire a house. Self-driving fleets and robot labor hit the same wall: grids that can’t charge fleets of trucks without sucking down city-scale power, and not enough trades people with real experience. At the same time, efficient models — including China’s Qwen line — keep doing more with less, which shrinks the case for blanketing the map in new data centers. Local on-device AI from phone and PC makers only accelerates that. For another angle on community AI outrunning vendors, see It Happened Again: Free AI Beat Nvidia Before Release.
Related on Mind of Will: The Real AI Data Center Problem: No Workers · It Happened Again: Free AI Beat Nvidia Before Release
Watch & follow: https://youtube.com/@mowuniverse
