# DeepInfra

DeepInfra provides hosted machine-learning model inference and on-demand GPU instances for developers and teams. Its catalog covers text, image, audio, video, embedding, reranking, and other model workloads with pay-as-you-go pricing.

- Findings24: <https://findings24.com/ja/tools/deepinfra>
- ウェブサイト: <https://deepinfra.com/>
- 料金: Freemium
- カテゴリー: AI推論プラットフォーム, GPUクラウド

## DeepInfraとは？

DeepInfra is a machine-learning inference and infrastructure platform for running hosted models and renting GPUs on demand. Its model catalog spans automatic speech recognition, embeddings, reranking, text generation, text-to-image, text-to-music, text-to-speech, text-to-video, world models, and zero-shot image classification.

The platform is designed around usage-based billing. Language models can use per-token pricing, while most other models are billed by inference execution time. DeepInfra also offers on-demand GPU instances for training, fine-tuning, and inference, with container setup, selectable base images, and hourly billing.

## DeepInfraでできること

### Hosted model catalog

Browse more than 100 models across language, vision, speech, media-generation, embedding, reranking, and classification categories.

### Pay-as-you-go inference

Use token pricing for some language models and execution-time billing for most other models, without long-term contracts or upfront costs.

### On-demand GPU instances

Rent NVIDIA B200 GPUs in single-card or multi-GPU configurations, with pricing billed by the minute and no egress fees stated on the GPU page.

### Container-based GPU setup

Select a GPU tier, name a container, choose a base image such as TensorFlow, PyTorch, or CUDA, and connect using a generated SSH command.

### Privacy and security commitments

The site states a zero-retention policy for inputs, outputs, and user data, and identifies DeepInfra as SOC 2 and ISO 27001 certified.

### Inference infrastructure

DeepInfra says it operates inference-optimized infrastructure in secure, US-based data centers and can tailor inference around cost, latency, throughput, or scale.

## 利用シーン

### Deploying language-model applications

Teams can select hosted text-generation models from families including DeepSeek, Qwen, Llama, Kimi, Gemini, and others, then pay for token usage.

### Multimodal and media workflows

Developers can evaluate or run workloads involving image, audio, and video inputs or outputs using the corresponding model categories in the catalog.

### Model experimentation and fine-tuning

ML practitioners can launch GPU containers with PyTorch, TensorFlow, or CUDA base images for training, fine-tuning, or inference tasks.

### Scaling inference usage

Applications with changing demand can use usage-based model inference or add GPU capacity without committing to a long-term infrastructure contract.

## よくある質問

### What does DeepInfra provide?

DeepInfra provides hosted machine-learning model inference and on-demand GPU instances. Its catalog includes language, speech, image, video, embedding, reranking, and classification models.

### How is DeepInfra priced?

Some language models use per-token pricing, while most other models are billed according to inference execution time. GPU instances use usage-based hourly pricing, with the GPU page also describing billing by the minute and no egress fees.

### Can I use DeepInfra for training or fine-tuning?

Yes. The GPU instances page describes launching containers for training, fine-tuning, and inference, and lists TensorFlow, PyTorch, and CUDA among the available base-image choices.

### How do I start a GPU instance?

The documented workflow is to choose a GPU tier, name a container, select a base image, copy the generated SSH command, and launch the configured container.

### Does DeepInfra state a data-retention policy?

Yes. The home page states that DeepInfra has a zero-retention policy for inputs, outputs, and user data. It also states that the company is SOC 2 and ISO 27001 certified.

## クイック情報

- **Product type**: Machine-learning inference and GPU infrastructure platform
- **Model categories**: Text generation, speech, image, video, embeddings, reranking, world models, and zero-shot image classification
- **GPU offering**: On-demand NVIDIA B200 instances, from single GPUs to multi-GPU configurations
- **Billing**: Pay-as-you-go; token-based or inference-time pricing depending on the model
- **GPU workflow**: Container launch with selectable TensorFlow, PyTorch, or CUDA base images
- **Source domain**: deepinfra.com