Published · Sources reviewed . Source publication dates are listed separately below.
Hold the workload and service target fixed
Record the model and version, precision, serving engine, GPU count, input and output lengths, concurrency and latency target. Measure aggregate throughput on that exact setup. Peak hardware FLOPS cannot substitute for a serving measurement, and a memory-fit estimate does not demonstrate usable speed.
Divide the full billed hosting cost by successfully delivered work. Include paid idle time, CPU, memory, storage and network charges where they apply. If throughput already includes interruptions and idle time, do not apply those penalties a second time.
Separate prices, revenue and contribution margin
Falling token prices can coexist with greater usage or more efficient serving. Rising rental prices can concern a different service class from the capacity a model provider already contracted. Two normalized price indices alone do not reveal a provider's margin.
Calculate revenue using the actual input, cache and output mix at the selling prices being tested. Subtract the defined serving costs to obtain contribution margin for that scenario. This is not net profit: research, sales, financing, taxes and other costs may be outside its scope. Anthropic's filing illustrates why term and cost-allocation context must stay beside financial ratios.
Evidence: Anthropic Form S-1/A, November 2, 2026
Measure the cost of accepted output
Artificial Analysis defines cost per task using its benchmark weights, token consumption and endpoint prices. That is a defined evaluation workload. It does not mean that every attempted task is correct, or that a customer's task mix has the same cost.
For your own workload, record the success rule before the run. Divide all included run spending, failed attempts and tool charges by the number of accepted outputs. Keep retries, reasoning settings, quality and latency visible. A cheaper unsuccessful attempt may cost more per accepted result.
Claims and evidence limits
The difference between GPU and token indices is inference margin.
Depends on scope and assumptions
The units differ. Measured workload throughput, utilization, selling-price mix and the stated cost scope are needed to connect them.
Benchmark cost per task means cost per successful task.
Depends on scope and assumptions
The Artificial Analysis metric uses its defined task count and weights. Success-adjusted cost needs an additional acceptance measure.
Gross GPU rental revenue is an asset's net return.
Depends on scope and assumptions
Continuous rental revenue assumes sold hours. Hosting costs, idle time, maintenance and depreciation must also be considered before interpreting returns.
Primary sources
- Anthropic Form S-1/A, November 2, 2026 · published . Reviewed 2026-11-03.
- Artificial Analysis benchmarking methodology. Reviewed 2026-11-03.
Company disclosures describe what the company reported. A review date does not turn a forecast into an observation or establish a private quote's terms.
Apply the explanation
Background reading: Inference Economics on AI Wiki. AI Wiki is run by the same team as AI Compute Tracker.