Self-hosting an 8x H200 LLM node costs about $10.5K monthly, undercutting APIs only at high utilization

IBM's CEO recently claimed open-weight models cost 10 to 25 percent as much as frontier models while being 90 percent as effective. A cost analysis using public GPU prices, power rates, throughput benchmarks and current API price sheets estimates an 8x H200 server runs about $10,500 a month before colocation, networking, and staffing. Against Claude Opus 5.5 the node breaks even at roughly 6 percent utilization, while against DeepSeek's off-peak API it needs about 54 percent. The analysis concludes that if an API bill is under roughly $10,500 a month, the API is cheaper regardless of setup. It also notes that the performance figures IBM cited do not clearly support the claimed 90 percent effectiveness.