The compute is coming back to your desk
Data centers are being blocked in their own neighborhoods. The alternative is the machine you already own — and the only thing standing in the way is the price of memory.
I've been running Qwen3.8-27B locally, and it's changed how I think about the future of AI infrastructure.
The current plan for scaling AI is to build more data centers. That plan is running into communities that don't want them. In Q1 2026 alone, local opposition blocked or delayed 75 data center projects worth roughly $130 billion. The concerns are concrete. Backup diesel generators hurt air quality. Cooling draws on local water. Power bills go up.
So where else can the compute go? Onto machines we already own.
The architecture I'd bet on: frontier models in the cloud handle complex planning, and local models execute the agent work on your own hardware. The heavy reasoning happens once in the data center. The repetitive steps run on your desk. Data centers still matter in that world. They just stop carrying every token.
Getting there takes model builders, hardware manufacturers, and in some cases regulators working together. AI is powerful enough to be misused, so the balance has to cover privacy, safety, cost, and business needs at the same time.
The barrier is cost
For now, cost is the barrier. Qwen3.8-27B fits in about 16GB at 4-bit, so a solid machine can run it. Bigger models or parallel agents need hardware like an M5 Max MacBook Pro (from $3,599) or an NVIDIA DGX Spark. The 128GB Spark launched at $3,999 last year. It now lists at $6,950. Memory shortages are pricing people out of local AI right as the case for it gets stronger.
Privacy, sustainability, and community impact all point toward local. Hardware pricing is the piece that has to catch up.
Anyone running agent workloads locally yet?