Self-Hosting LLMs May No Longer Save Money as GPU Prices Surge and API Costs Fall

The traditional cost argument for self-hosting large language models has weakened in 2025, as two opposing trends have reshaped the math. API pricing from providers like OpenAI and Anthropic has continued falling, with GPT model prices cut multiple times over summer 2025 and Claude Haiku 4.5 offering up to 90% discounts on cached inputs. Meanwhile, GPU hardware costs have risen sharply, with the NVIDIA RTX PRO 6000 Blackwell climbing roughly 87% above its March 2025 launch price by August, driven by a memory shortage. Analysis of real production workloads also reveals that tool and web-search call costs can outweigh token costs, and that prompt caching optimizations are often overlooked before hardware decisions are even considered. For narrow tasks like document classification or PII redaction, smaller fine-tuned open-weight models running on consumer-grade 24GB GPUs may offer a more practical and cost-effective path than expensive workstation hardware.