Step 1 · Build or not
Build, rent or call an API?
Start here before buying anything. Most teams should begin with an API, move steady workloads to rented GPUs, and buy hardware only when utilisation is high and predictable – or when data cannot leave.
| API | Cloud GPU rental | Self-hosted | |
|---|---|---|---|
| Up-front cost | None | None | Hardware + facility |
| Marginal cost | Per token | Per GPU-hour (pay when idle too) | Power + ops; idle hardware still depreciates |
| Data control | Data leaves to the provider | Your VM, provider’s data centre | Fully in-house |
| Model choice | Whatever the provider offers | Any open-weight model | Any open-weight model |
| Fine-tuning | Limited (vendor fine-tune APIs) | Full control | Full control |
| Scaling | Instant (rate limits apply) | Minutes–days (capacity varies) | Weeks–months (procurement, power) |
| Ops burden | None | Software only | Hardware, network, facility, software |
| Best for | Low or spiky volume; frontier closed models | Training bursts, experiments, uncertain demand | High steady utilisation, data that must stay in-house, long-term fine-tuning |
Rules of thumb
- Utilisation decides everything: hardware bought for 20% utilisation is usually more expensive than renting or the API. Measure your real token volume first (API bills are a good proxy).
- Compare like with like: an open-weight model you can self-host (e.g. DeepSeek, Qwen, GLM, Kimi) against the same model’s API price – not against a closed frontier model.
- Throughput must be measured, not taken from datasheets: tokens/s depend on the engine, batch size, context length and quantisation.
- Hidden costs: staff time, spare parts, network gear, rack space, cooling, and the risk that a better GPU halves your hardware’s value within 2 years.
- Hybrid is normal: self-host the steady base load, burst to the API or rented GPUs for peaks.
Break-even calculator
How it works: API = tokens × blended price. Cloud = $/GPU·h × GPUs × hours. Self-hosted = hardware ÷ depreciation months + kW × PUE × 730h × electricity + hosting/ops, multiplied by the number of machines needed for your volume at the measured throughput and utilisation. API presets come from our price table (official prices of open-weight models, updated 2026-10-22); the cloud default ($2.69/h) is RunPod’s H100 community price as quoted on our Services page. Hardware cost, power and throughput defaults are only examples – replace them with your quote and your own benchmark. Not included: staff time, financing, resale value, tax. API price table → · GPU rental services →
Where to rent GPUs or hosted inference
From our Services directory. Prices there are as quoted by the providers’ pages or third parties – check before you commit.
| Provider | Type | Price note |
|---|---|---|
| RunPod | GPU 算力租赁 | H100 SXM $2.69（社区）/$3.29（安全）每小时（官网定价页，第三方转述） |
| Lambda | GPU 算力租赁 | 单卡 H100 SXM $4.29/h，8 卡 $3.99/GPU·h（第三方对比） |
| Vast.ai | GPU 算力租赁 | 竞价；H100 SXM 约 $1.73 起、中位 $2.16/h（2026-09 第三方） |
| CoreWeave | GPU 算力租赁 | 企业报价 |
| Modal | GPU 算力租赁 | 按秒计费；$30/月免费额度（首页） |
| Hyperbolic | GPU 算力租赁 | 按量 |
| AutoDL | GPU 算力租赁 | 4090 约 ¥1.88/h（2026-10-09 第三方监测） |
| 恒源云 | GPU 算力租赁 | 4090 约 ¥1.85/h（第三方监测） |
| 矩池云 | GPU 算力租赁 | 4090 ¥1.54/GPU·h 起（首页） |
| Featurize | GPU 算力租赁 | 4090 ¥1.87/h 起（首页） |
| GpuGeek | GPU 算力租赁 | 4090 约 ¥2–3/h（第三方） |
| Together AI | 推理部署托管 | 按 token/按 GPU |
| Fireworks AI | 推理部署托管 | 按 token（首页示例 $3/M 入、$15/M 出） |
| Replicate（已被 Cloudflare 收购） | 推理部署托管 | 按秒计费（T4 $0.000225/s 等，首页） |
| Baseten | 推理部署托管 | 按用量/专属 |
| Groq | 推理部署托管 | 按 token |
| Cerebras | 推理部署托管 | 按 token |
| DeepInfra | 推理部署托管 | 按 token（首页示例 $0.834/M 入） |
| Novita AI | 推理部署托管 | 按 token（首页示例 DeepSeek V4 Pro $1.74/$3.48 每百万） |
| Hugging Face Inference Providers | 推理部署托管 | 按上游计费 |
| 硅基流动 SiliconFlow | 推理部署托管 | 按 token |
| PPIO 派欧云 | 推理部署托管 | 4090 Spot ¥1.16/h、H100 ¥11.88/h（首页） |
| 无问芯穹 | 推理部署托管 | 企业报价 |