Every number, and where it came from

Open-weight model size, activation and licence

Total parameters and file sizes were computed by summing safetensors byte counts from the Hugging Face model API. Active parameters are vendor-published where available and derived from tensor shapes where not, marked accordingly. Active parameters, not total, drive serving cost.

Derived figures were computed from tensor shapes in each model config. The method was validated against DeepSeek-V4-Pro, where the derivation gave 50.3B against a published 49B, a 2.7 percent error. Treat every derived value as carrying that bound.

Quality index, same scale

Artificial Analysis Intelligence Index v4.1.1. Reasoning mode is stated because it changes the number materially, and two widely-quoted entries are vendor estimates rather than independent measurements.

Qwen3.8-27B ranks first of 135 open-weight models in the 4B to 40B class, where the class median is 9. Claims that it outranks Opus 4.6 are literally true on this table and rest on an Opus figure that was never independently measured, in non-reasoning mode.

Datacentre node rental, 24/7

Monthly figures are hourly rate multiplied by eight GPUs and 730 hours. The fit column states whether the node holds a roughly 1T-parameter mixture of experts at FP8.

Consumer and prosumer card rental

Marketplace rates, which run roughly an order of magnitude below enterprise clouds for the same silicon. The agent column is how many concurrent sessions the card holds at 50k context each, after weights and workspace.

Concurrency figures account for a fixed per-sequence recurrent state in hybrid linear-attention models, roughly 159 MB per session for Qwen3.8-27B, which is independent of context length. Linear attention makes one long context cheap and taxes many short ones.

Buying, with breakeven against rental

Retail prices are inflated well above launch pricing. Breakeven is months of typical marketplace rental equal to the purchase price, before electricity.

The RTX 5090 launched at $1,999 and retails around 2.3 times that. Buying at this point in the cycle is a bet that the shortage persists. European electricity at 0.27 EUR per kWh adds roughly 30 to 70 EUR a month per card at 24/7 and 70 percent of rated power.

Measured serving throughput

Aggregate output tokens per second on a single eight-GPU node. Peak figures are quoted at low per-user interactivity and do not transfer to interactive work.

Interactivity trades against aggregate throughput steeply. On identical H200 hardware running DeepSeek R1 at FP8, serving 65 tok/s per user yields 1,035 tok/s per GPU; 99 tok/s per user yields 488; 133 tok/s per user yields 264. Making the model twice as fast for one person costs almost four times the total throughput.

Cost per million output tokens, self-hosted

Computed as 277.7778 multiplied by the node hourly rate, divided by aggregate output tokens per second. The constant is one million divided by 3,600.

Kimi K3 self-hosted at $20.07 sits above Moonshot list pricing of $15.00 per million output tokens. The throughput input to that row is published rather than independently measured, so treat the exact figure as indicative; the direction holds because K3 needs ten to fourteen GPUs.

Per-token API pricing

Blended figures assume three input tokens per output token. The cache-read column matters more than either headline rate on a cache-heavy agentic workload, where it can be 95 percent of the bill.

Provider spread on Kimi K3 across thirteen hosts is only about 15 percent, so shopping providers gains little. Prompt caching gains an order of magnitude, since a cache hit bills at one tenth of a miss. On DeepSeek V4 Flash the throughput spread across providers reaches 5.8x at near-identical prices, so on the cheap tier the right variable to optimise is speed, not price.

Break-even grid

Monthly output tokens required before a rented node beats buying the same tokens. Rent divided by API price. Capacity is what the node can physically produce at full utilisation.

Read the last two columns together. An 8x H200 serving an agentic coding workload at 2,469 tok/s produces at most 6.5 billion output tokens a month, which is below its own 8.8 billion break-even against a $2.00 API. That node cannot beat the API at any utilisation. It only wins in batch mode at 4.9 tokens per second per user, which no interactive user would accept.

Reliability and hidden costs

Costs that do not appear in an hourly rate: RunPod volume storage doubles to $0.20 per GB-month while a pod is stopped, so a 1 TB checkpoint parked idle runs about $200 a month. Hyperstack bills virtual machines in the shutoff state. Vast.ai bills storage continuously even at a negative balance and schedules instances, volumes and data for deletion once the balance reaches zero. Hyperscaler egress runs $0.087 to $0.12 per GB, so moving a terabyte out of one costs more than eight H100-hours.

Preemption notice ranges from two minutes on one hyperscaler to thirty seconds on another, five seconds on one marketplace, and nothing documented at all on another. A node loss costs five to twenty-five minutes of cold start before it serves again. A reliable endpoint therefore needs a warm standby, which roughly doubles the effective hourly cost and removes most of the marketplace saving.

Claims removed by audit

The research agents produced these. A separate reviewer, instructed only to refute and to default to failure on anything it could not verify against a primary source, removed them.

Sources

Every figure above traces to one of these. Provider pricing pages were read on 2026-08-19; model files and configuration were read from the Hugging Face API on the same day.

Outbound links carry rel="nofollow" and open in a new tab. Prices and leaderboard positions on those pages change without notice, so a figure quoted here should be treated as a reading taken on 2026-08-19 rather than a current quote.

Link check, run at publish time: 50 of 52 outbound targets resolved. The Nebius and Crusoe pricing pages returned a connection failure from the checking location while their DNS resolved normally and every other provider responded, which points at a network-level block rather than an outage. Their figures were captured earlier in the research and are retained; verify them yourself before relying on either.

Method and limits

Model sizes were computed by summing safetensors byte counts rather than quoting a card. Architecture figures came from each model configuration file. Cost arithmetic is shown inline throughout so a reader can check it rather than trust it.

Three limits are worth stating. Active parameter counts marked as derived carry roughly a 3 percent error bound, validated against the one model that publishes both. Throughput figures are published measurements from third parties, not reproductions, and peak numbers are quoted at interactivity levels that suit batch work rather than interactive use. And the workload shape used to weight these costs came from one heavy agentic setup, so the 94 percent cache-read figure is characteristic of that pattern rather than universal.

Return to the study