Generated from 30 days of usage, seat, and fleet telemetry against live market rates (H100 spot $3.64/hr, as of Sep 13). Savings are monthly run-rate.
A single-vendor price change or rate limit would hit the majority of inference traffic. Frontier workloads on Sonnet-class models have credible fallbacks (GPT-5.6 Sol, Gemini 3.1 Pro, DeepSeek V4 on Vast); wiring a router with a 20% secondary allocation caps exposure.
dgx-h100-02 averaged 25% utilization over the last 30 days, down from 27%. Listing the idle window at the current market rate of $3.64/GPU-hr recovers ~$13,618/mo against a fixed TCO of $12,257.
OpenAI reduced GPT-5.6 Sol list pricing. At current volume (2,979 Mtok/mo) the run-rate is $2,471/mo lower than it would have been.
Kimi (Moonshot) reduced Kimi K3 list pricing. At current volume (398 Mtok/mo) the run-rate is $739/mo lower than it would have been.
25 people are paying metered Anthropic API rates for what looks like interactive, weekday-shaped usage ($26,855/mo combined). A Claude Max 20× seat at $200/mo absorbs up to ~$1,400 of that per person.
hgx-a100-01 averaged 24% utilization over the last 30 days, down from 25%. Listing the idle window at the current market rate of $1.06/GPU-hr recovers ~$3,967/mo against a fixed TCO of $6,129.
Data Platform spends $2,267/mo on small/mid closed models (Haiku, GPT-5.6 Luna, GPT-5.4 mini, Gemini Flash) for classification and summarization workloads. The same 1,120 Mtok/mo on owned H100s costs ~$195 — and 19 owned GPUs are idle on average.
l40s-edge-01 averaged 14% utilization over the last 30 days, down from 14%. Listing the idle window at the current market rate of $0.78/GPU-hr recovers ~$1,664/mo against a fixed TCO of $1,605.
Customer Support spends $1,036/mo on small/mid closed models (Haiku, GPT-5.6 Luna, GPT-5.4 mini, Gemini Flash) for classification and summarization workloads. The same 621 Mtok/mo on owned H100s costs ~$108 — and 19 owned GPUs are idle on average.
Engineering spends $988/mo on small/mid closed models (Haiku, GPT-5.6 Luna, GPT-5.4 mini, Gemini Flash) for classification and summarization workloads. The same 820 Mtok/mo on owned H100s costs ~$143 — and 19 owned GPUs are idle on average.
OpenRouter reduced DeepSeek V4 Pro (via OR) list pricing. At current volume (488 Mtok/mo) the run-rate is $144/mo lower than it would have been.