The Economics of Open-Weight Inference

Ornn's analysis argues that open-weight model demand extends the economic usefulness of older NVIDIA GPU families by routing price-elastic workloads to the cheapest compatible hardware. Across eleven open-weight and eight closed models, the cheapest qualifying open-weight model completes tasks at roughly one-fifth the cost of a comparable closed model. Self-hosting on rented hardware lowers compute-only costs to about $0.12-$0.35 per million output tokens at full utilization; for gpt-oss-120b (a sparse 5.1B active-parameter model) the A100 produces output more cheaply than the H100 in spot and multi-year term pricing, with a base-case A100 cost of $0.29 versus $0.64 per million tokens for the H100. Market rental data show the A100’s five-year term price retaining 80.2% of its one-month price (vs. 43.7-59.8% for Hopper/Blackwell), and A100 occupancy rose from 74% to 90% while listed capacity grew 13% and spot price rose 20%.

The analysis uses Ornn’s settled daily rental index, published term-price curves, and throughput measurements to compute compute-only token costs and forward retention. It stresses that latency-tolerant, compute-intensive workloads (long-running agents, batch eval, parts of RL) can migrate to cost-efficient older hardware, preserving multi-year earning life for predecessors. Limitations are explicit: forward marks are analyst indicators, occupancy tracks listed rental supply not installed base, some throughput rows are estimated or combine different third-party setups, and the rental series do not establish causal effects of open-weight demand. Commercial interests and data licensing are disclosed.