Comparison
FastMetal vs self-hosted LLMs (Ollama and friends)
Running open-weight models on your own GPUs with Ollama or vLLM solves a slightly different problem than an API gateway like FastMetal.
Local inference wins on data never leaving your infrastructure and on fixed costs. A gateway wins on access to the strongest closed models with no hardware, billed per token. Many teams use both.
At a glance
| Aspect | FastMetal | Self-hosted LLMs |
|---|---|---|
| Billing and invoices | Metered billing in JPY. No upfront cost | GPU purchase or rental. No per-token charge for inference |
| API compatibility | OpenAI- and Anthropic-compatible | Ollama offers an OpenAI-compatible endpoint |
| Models and providers | 200+, including closed models | Open-weight models only |
| Operational burden | Nothing to operate | GPUs, model updates, and scaling, all yours |
| Language and support | Japanese support | OSS community |
| Data flow | Requests are relayed to each provider | Data stays inside your infrastructure |
When self-hosted LLMs fit
- Your data must never leave your infrastructure. With local inference, requests stay inside.
- Your volume is steady and high enough that fixed GPU costs beat metered billing.
- You need offline operation, or latency physically close to your workload.
When FastMetal fits
- You want the top closed models such as Claude, GPT, and Gemini. There is a quality tier open weights do not reach.
- You do not want to procure and operate GPUs. Trying FastMetal takes an API key and nothing else.
- You want image and video generation alongside text, on one API.
The actual cost: where self-hosted GPUs break even against hosted APIs
Open-weight models can also be used through hosted APIs, often the very same model. So whether running locally is cheaper comes down to hosted per-token rates versus how continuously you can keep your own GPU busy.
The charts below compare running a 100B-class open-weight model on a single-H100 server you own against using the same class of model through a hosted API. The assumptions are rough August 2026 market figures, and labor is excluded.
Assumptions (rough figures, August 2026)
- Own server: a single-H100 build bought for about JPY 5M, straight-lined over 3 years, plus power, about JPY 170k a month. Labor excluded
- Throughput: about 8,000 tokens/second aggregate with batched inference, an effective ceiling of about 20 billion tokens a month at 100% utilization
- Hosted rates: blended input and output for the same class of open-weight model, about JPY 30 per million tokens standard and about JPY 8 at the cheapest tier
- Exchange rate: USD 1 = JPY 150
- The self-hosted side additionally carries operations and maintenance labor. Published analyses put real total cost at 3 to 5 times the raw GPU cost
In short, choosing local inference on cost alone requires confidence the GPU will stay busy at high utilization. For bursty development use or internal tools that run only during working hours, hosted is cheaper almost every time.
If your reasons are data that cannot leave, offline operation, or a fixed budget, local inference is worth it regardless of utilization. As above, this is not an either-or choice.
A common and sensible setup is local small models for development and gateway access to stronger models for production and quality checks. It is not an either-or choice.
This comparison reflects August 2026.
Try it by changing one line: the base URL
Create a free account and use 200+ models with one API key. No subscription, no minimum.