The base Mac Mini costs $599. On eBay, they sell for $1,800. That 3x markup is not a supply glitch—it is a preference poll.
For the past six months, a quiet but determined cohort of developers has been building AI infrastructure at home. They are not experimenting. They are shipping products. And they are increasingly choosing local inference over cloud APIs, even when it costs more upfront.
The Mac Mini is their weapon of choice. Apple's M4 Pro chip delivers 24 GPU cores and up to 48GB of unified memory—a configuration that can run a 70B parameter model at reasonable speeds without thermal throttling or external GPUs. The unified memory architecture treats all memory as VRAM, meaning no slow CPU-GPU data transfers. For developers running open-source models like Llama 3 or Mistral, this is a practical setup that works.
The eBay premium reveals how much developers value this capability. Sellers list Mac Minis at $1,800, well above the $1,400 retail for the 48GB variant. Buyers pay it because cloud inference has become unpredictable. Rate limits on GPT-4o. Quotas on Claude. Unexpected price changes from API providers that force mid-sprint rewrites.
Cloud providers built the infrastructure that spawned this backlash. OpenAI, Anthropic, and Google made AI accessible. Then they made it expensive, then rationed. Developers who built on these APIs woke up to find their costs had tripled or their access throttled during peak usage. The response is rational: build locally and own the compute.
This is not purely a cost calculation. Privacy plays a role. Running prompts through a third-party API means that data leaves the developer's control. For applications handling sensitive information—medical records, legal documents, financial data—local inference is not a preference but a requirement. Prompt injection attacks on cloud endpoints have made headlines, reinforcing the perception that public APIs carry hidden risks.
The developer tooling reflects this shift. Ollama, LM Studio, and similar local inference platforms have seen explosive adoption. What began as a hobbyist movement has matured into production infrastructure. Teams that once routed every user query through OpenAI now run smaller models locally for routine tasks, reserving cloud APIs for tasks requiring frontier capability.
The Mac Mini shortage exposes a real infrastructure preference, not an artificial supply crunch. Apple is ramping production. The question is whether demand will sustain at these levels or normalize as cloud providers respond.
Cloud AI is not going away. Frontier models still require infrastructure no desktop can provide. But the market is fragmenting in a way that favors local. Developers have learned that renting inference is renting uncertainty. For teams that can run their workloads on hardware they own, the premium is worth paying. The Mac Mini shortage is the symptom. Cloud dependency fatigue is the diagnosis. And the prescription—own your inference stack—is gaining converts.