From Mac Studios to H200 clusters — what you actually need.
Self Hosting Open Models Local dev (7B 30B models): M3 Ultra Mac Studio with 192GB. $7K. Runs Llama 3.1 70B quantized. Production small (70B): Single H100 80GB on a serverless platform like Modal. $4/hr on demand. Production large (Llama 4 400B): 8x H200 cluster. $80K to buy, or $40/hr to rent. The break even: If you're inferring more than 100M tokens/month, self hosting beats API pricing.