Artificial Analysis Ties AA-AgentPerf Scores to Daily GPU Rental Prices

Artificial Analysis has connected its AA-AgentPerf datacenter benchmark to daily GPU rental indices from Silicon Data. The leaderboard now reports token throughput per hourly rental dollar alongside raw speed, helping teams compare serving configurations at current market rates.

AA-AgentPerf replays real coding-agent trajectories against inference systems. Each simulated session spans multiple turns, often exceeds 100,000 tokens, includes tool-calling delays, and reuses cached context. This workload tests long-context handling, scheduling, generation speed, and key-value cache performance under conditions closer to agent deployments than short prompt-response tests.

A moving denominator rewrites the score

The benchmark calls its new metric Throughput per $/hr. It follows this calculation: (input tokens/s + output tokens/s) ÷ (GPU count × rental USD/GPU-hour). Input throughput includes tokens served through key-value cache hits, which represent reused computation from previously processed prefixes.

Silicon Data supplies daily rental rates for NVIDIA H100, H200, B200, and B300 GPUs. AA-AgentPerf multiplies the relevant per-GPU rate by the number of accelerators in each tested configuration. AMD’s MI355X currently uses an assumed rate based on recent market quotes because a comparable rental index is unavailable.

Silicon Data’s indices estimate standardized hourly rates from neo-cloud providers and GPU marketplaces. The methodology removes outliers and independently validates daily observations. Actual quotes can still vary by region, contract term, availability, networking, and minimum commitment. H100 also has separate neo-cloud and hyperscaler readings because those provider classes have historically carried different prices.

The rental curve beneath the chart

Approximate rental readings cited in the comparison

GPU	Architecture	USD per GPU-hour	Source
H100, neo-cloud	Hopper	~$2.53	H100
H200	Hopper	~$3.10	H200
B200	Blackwell	~$5.45	B200
B300	Blackwell Ultra	~$6.71	Early October reading

A separate B200 market update placed the accelerator at $5.62 per GPU-hour as September 2026 began, up 27.6% year to date. Its premium over H100 hovered near twofold. Those movements explain why a price-linked leaderboard can change while measured hardware throughput stays fixed.

Blackwell leads at current rates

NVIDIA’s B300 generally delivers the highest throughput per rental dollar among systems with an available price input on the current AA-AgentPerf chart. Its aggregate throughput offsets its higher hourly rate across much of the tested range.

B200 reaches parity with B300 at higher per-request output speeds for some models. At those operating points, B300’s throughput advantage narrows relative to its rental premium, allowing the cheaper Blackwell configuration to produce a comparable per-dollar result.

Where the ranking can bend

• Latency target: Output speed is measured per request at the 25th percentile. Three quarters of observed requests reached at least that rate, preventing the fastest requests from dominating the result.
• Cache accounting: Input throughput includes key-value cache hits. The metric therefore rewards serving stacks that reuse long prefixes efficiently, a common pattern in multi-turn agent sessions.
• Price confidence: The MI355X result depends on an assumed rental rate. Its placement remains directional until observed market data supports an index.
• Provider class: Neo-cloud and hyperscaler rates can differ substantially. A deployment priced through AWS or Azure may produce another ordering.
• Cost coverage: GPU rental rates exclude CPUs, memory, storage, networking, data transfer, orchestration, idle capacity, and engineering overhead.
• Workload specificity: Rankings depend on the model, context length, quantization, serving software, batch policy, cache strategy, and output-speed target.
• Quality scope: Model correctness and completed-task quality sit outside this throughput metric. Teams should evaluate those outcomes separately.

From leaderboard to capacity plan

Silicon Data’s market update says NYMEX has listed B200 rental-index futures under the code GPU2 alongside an H100 contract. Each contract covers 730 GPU-hours and settles financially against the relevant Silicon Data index. Daily indices provide a market reference for capacity costs, while futures add a forward price for budgeting and hedging.

Teams sizing an agent deployment can apply the benchmark through a short validation process:

1. Match the benchmarked model, context profile, cache behavior, and output-speed target to the production workload.
2. Use throughput per rental dollar to shortlist hardware and serving configurations.
3. Replace the index rate with the team’s contracted, all-in infrastructure cost.
4. Replay production traces to verify latency, throughput, reliability, and task completion quality.

The leaderboard will reprice as rental markets move. Easing Blackwell supply could narrow the B200 and B300 premiums over Hopper, potentially changing the per-dollar leader while the underlying benchmark results remain constant.