Based on the last 30 days of usage, seats, and GPU activity, compared against current market prices (H100 at $3.09/hr as of Nov 14). Savings are shown per 30 days; seat and API savings grow with headcount (+20%/yr).
If they raise prices or rate-limit you, most of your AI usage takes the hit. There are good alternatives for the Sonnet-class work (GPT-5.6 Sol, Gemini 3.1 Pro, DeepSeek V4 on Vast). Sending about 20% of traffic to a second provider would limit the damage.
25 people are paying Anthropic API rates for what looks like everyday, weekday use — $26,855/mo between them. A Claude Max 20× seat is $200/mo and covers roughly $1,400 of that kind of usage per person.
OpenAI lowered the price of GPT-5.6 Sol. At your current volume of about 2,979M tokens a month, that's $2,471/mo you're no longer paying.
Kimi (Moonshot) lowered the price of Kimi K3. At your current volume of about 398M tokens a month, that's $739/mo you're no longer paying.
dgx-h100-02 was only 25% busy over the last 30 days, down from 27%. Renting out the idle hours at today's market rate of $3.09/GPU-hr would bring in about $11,555/mo on a machine that costs $12,257/mo whether you use it or not.
hgx-a100-01 was only 24% busy over the last 30 days, down from 25%. Renting out the idle hours at today's market rate of $2.02/GPU-hr would bring in about $7,579/mo on a machine that costs $6,129/mo whether you use it or not.
You're paying AWS $4/GPU-hr in us-east-1 for inference, and it's only 53% busy. The same H100s on Vast go for about $3.40/GPU-hr, allowing 10% extra for a more reliable host.
You're paying Azure $4/GPU-hr in eastus2 for training, and it's only 42% busy. The same H100s on Vast go for about $3.40/GPU-hr, allowing 10% extra for a more reliable host. Since it's under half used, you could drop half the machine entirely.
Data Platform spends $2,267/mo on smaller models (Haiku, GPT-5.6 Luna, GPT-5.4 mini, Gemini Flash) for things like tagging and summarizing. Running that same volume — about 1,120M tokens a month — on Qwen3 on your own H100s would cost around $195, and 19 of those GPUs are sitting idle on average.
l40s-edge-01 was only 14% busy over the last 30 days, down from 14%. Renting out the idle hours at today's market rate of $0.78/GPU-hr would bring in about $1,664/mo on a machine that costs $1,605/mo whether you use it or not.
You're paying GCP $4/GPU-hr in us-central1 for batch, and it's only 27% busy. The same H100s on Vast go for about $3.40/GPU-hr, allowing 10% extra for a more reliable host. Since it's under half used, you could drop half the machine entirely.
Customer Support spends $1,036/mo on smaller models (Haiku, GPT-5.6 Luna, GPT-5.4 mini, Gemini Flash) for things like tagging and summarizing. Running that same volume — about 621M tokens a month — on Qwen3 on your own H100s would cost around $108, and 19 of those GPUs are sitting idle on average.
Engineering spends $988/mo on smaller models (Haiku, GPT-5.6 Luna, GPT-5.4 mini, Gemini Flash) for things like tagging and summarizing. Running that same volume — about 820M tokens a month — on Qwen3 on your own H100s would cost around $143, and 19 of those GPUs are sitting idle on average.
OpenRouter lowered the price of DeepSeek V4 Pro (via OR). At your current volume of about 488M tokens a month, that's $144/mo you're no longer paying.