For Qwen3-30B-A3B, start investigating self-hosting around 4.30 million requests per 30 days, at 1,024 input and 256 output tokens per request. That is 1.10 billion output tokens, plus their associated input. At that volume, the modeled API bill equals the 1,080.00 USD allocation budget. This is a reason to benchmark a deployment with your traffic, latency target and quality requirements.
Qwen3-30B-A3B activates only a subset of its experts for each token, but the complete checkpoint still needs memory. The concrete candidate here is the publisher's FP8 model served by SGLang on one L40S, the setup behind the reviewed aggregate reference. Use your own API usage split between input and generated output to interpret the live crossover. Before migration, test the candidate against the same prompts, thinking setting, arrival pattern and answer-quality checks; a short synthetic run cannot establish long-context capacity or production latency.
Sparse experts save compute, not resident weights
The publisher describes 128 experts with eight activated per token, and distinguishes 30.5 billion total parameters from 3.3 billion active. The active count helps explain the possible serving-speed advantage; it is not the amount of weight data that must be present. The pinned original BF16 checkpoint is 56.87 GiB of tensor payload. That exceeds one L40S's nominal 48 GB before cache or runtime memory. The tested candidate uses the publisher's separate FP8 checkpoint. Check whether its outputs meet your quality bar before treating its throughput as a substitute for an original-BF16 deployment.
Evidence: architecture source; model source; serving source; benchmark source.
Short requests leave context and cache untested
The pinned original configuration has 48 layers, four KV heads and a 128-element head dimension. A BF16 whole-model KV estimate adds 96 KiB per resident token. Four simultaneous 4,096-token sequences bring original weights plus modeled cache to 58.37 GiB before serving overhead; four full 40,960-token sequences bring that subtotal to 71.87 GiB. The model card discusses longer YaRN contexts, but that is a separate deployment setting. The reviewed FP8 benchmark sent short fixed requests, so it cannot tell you the output rate or tail latency for retrieval-heavy prompts, long reasoning traces or the extended-context setting. Measure those shapes before using the live crossover for them.
Evidence: architecture source; model source; architecture source; benchmark source; benchmark source.
One L40S is a testable candidate, with a bounded measurement
The author's AIPerf run used one L40S, SGLang, 32 concurrent streams and 300 measured synthetic chat requests after 30 warmups. Its fixed request shape was 1,024 input and 256 output tokens. The raw report gives about 897.39 output tokens per second for the whole service; its separate per-user speed is not the capacity input. The benchmark also reports a p99 time to first token of about 1.303 seconds on that short workload. Those results justify a focused rental test when the live spending comparison points to one, but they do not establish sustained rental capacity, quality equivalence or your latency target. The exact benchmarked checkpoint revision and SGLang build were not recorded, so pin both when reproducing it. NVIDIA also reports much higher figures for an FP4 B200 variant of this model; that different precision and GPU cannot replace the FP8 L40S measurement in this cost comparison.
Evidence: benchmark source; benchmark source; benchmark source; serving source; benchmark source.
The current worked traffic calculation
The comparison uses DeepInfra's exact endpoint Qwen/Qwen3-30B-A3B (FP8), at 0.12 USD/million input tokens and 0.5 USD/million output tokens. The rental candidate is Crusoe Cloud, l40s-48gb.1x, us-east1-a: 1 × L40S 48 GB, at a whole-instance 1.500000 USD/hour. This is the least expensive qualifying comparison in our catalog, not a whole-market claim. Confirm current stock and final terms.
Period budget = 1.500000 USD/hour × 720 hours × 1 replica(s) + 0.00 USD
= 1,080.00 USD
API USD/million output, including input = 0.5 + (1024 / 256) × 0.12
= 0.98000000
Output-token crossover = 1,080.00 / 0.98000000 × 1,000,000
≈ 1,102,040,816 output tokens
Requests = output tokens / 256 ≈ 4,304,847| Requests per period | API bill, USD | Allocation budget, USD | Bill and throughput evidence |
|---|---|---|---|
| 100,000 | 25.09 | 1,080.00 | API lower bill |
| 1,000,000 | 250.88 | 1,080.00 | API lower bill |
| 3,443,877 | 864.00 | 1,080.00 | API lower bill |
| 5,165,816 | 1,296.00 | 1,080.00 | Allocation bill lower; within reference |
| 10,000,000 | 2,508.80 | 1,080.00 | Allocation bill lower; above reference |
The crossover needs 425.17 output tokens/second averaged over the calendar period. The published FP8 reference is 897.39 output tokens/second per 1-GPU allocation, on One NVIDIA L40S 48 GB GPU; author-run SGLang service, using Qwen/Qwen3-30B-A3B-FP8; measured revision undisclosed and SGLang. The required rate is below the reference, which identifies a candidate to test. This is not measured rental throughput, a latency guarantee, a peak ceiling, or proof that the API serves identical weights.
API price checked 2026-10-01T06:16:12.941138+00:00; rental quote checked 2026-09-30T07:53:09.362914+00:00; published benchmark. Calculation generated 2026-10-01T19:58:50.663979+00:00. Listed compute excludes additional storage, traffic and operations unless included in the assumptions.
What changing request length would require
Only the 1,024-input / 256-output-token shape has a reviewed aggregate output reference here. A different input/output ratio needs a matching throughput measurement before this deployment comparison can claim capacity or savings for it.
Include the service you actually need
| Scenario | Period budget, USD | Break-even requests, millions |
|---|---|---|
| Current assumptions | 1,080.00 | 4.30 |
| Illustrative 1,000 USD operating budget | 2,080.00 | 8.29 |
| Twice the allocated replicas | 2,160.00 | 8.61 |
The operating allowance is illustrative. Doubling replicas is a cost scenario, not a tested availability design. Below the applicable crossover the API has the lower modeled bill; near or above it, test the deployment under your arrival pattern and quality/latency requirements.
Self-hosting terms and evidence
The original checkpoint and publisher FP8 checkpoint each carry Apache License 2.0. That license permits operating a self-hosted service. If distributing copies or modified files, preserve the license and applicable notices and mark changed files. Review the separate serving-software terms for your chosen stack.
Publisher terms. Analysis reviewed 2026-10-01T14:05:09.754646+00:00. Prices refresh independently; an unchanged source check does not become a new editorial review. This analysis uses published reference measurements. Noach Ark did not rent this allocation or run inference or quality tests for this article.