DeepInfra
by DeepInfra
Pay-per-token inference for 100+ open-weight models, plus dedicated GPUs when you outgrow shared endpoints.
DeepInfra is an inference cloud that serves more than 100 open-weight models (LLMs, embeddings, speech, image and video) through an OpenAI-compatible API billed per token. It is for engineering teams that want open models in production without running GPUs themselves, and it also rents dedicated A100 to B300 instances.
DeepInfra is a serverless inference platform built by the team behind the imo messenger app and headquartered in Palo Alto. It hosts a catalog of more than 100 models covering text generation, vision and OCR, embeddings and reranking, text-to-image, text-to-video, speech recognition and text-to-speech, and exposes them through an endpoint at api.deepinfra.com/v1/openai that works as a drop-in replacement for the OpenAI SDK. Language models are billed per million input and output tokens with discounted cached-input rates, and the pricing page lists Standard, Priority (1.5x base price) and Flex (0.8x base price, for non-production work) service tiers. Teams that need isolation can deploy a dedicated autoscaling instance on A100, H100, H200, B200 or B300 GPUs, use private endpoints, or rent a GPU cluster with SSH access; the company markets a DeepCluster option with a 99.982% uptime SLA. The homepage states a zero-retention policy for user data and SOC 2 and ISO 27001 certification, with inference running in US data centers. DeepInfra raised a $107M Series B in May 2026, co-led by 500 Global and Georges Harik with NVIDIA, Samsung Next, Supermicro and Felicis participating, after an $18M Series A in April 2025. SDxCentral reported eight US data centers at the time of the round, and the company says token volume grew 25x since the Series A. It is also a listed provider on Hugging Face Inference Providers and has a first-party package in the Vercel AI SDK, which puts it in direct competition with Together AI, Fireworks AI and Groq for open-model API traffic.
Platform or ML engineering teams moving open-weight models such as DeepSeek, GLM or Kimi into production who want an OpenAI-compatible endpoint instead of operating their own GPU fleet.
Open-model inference billed per token with no minimums or seat fees, with a path to dedicated GPUs on the same vendor when volume justifies it.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Usage-based
- Target Market
- CTOs, Enterprise Developers, ML Engineers, Platform Teams
- Deployment
- API-based, Cloud-only
- Founded
- 2022
- Headquarters
- Palo Alto, USA
Key Features
- ✓OpenAI-compatible API
Point the existing OpenAI SDK at api.deepinfra.com/v1/openai and keep application code unchanged while switching to open-weight models.
- ✓100+ model catalog
LLMs, vision and OCR, embeddings, reranking, image, video, speech-to-text and text-to-speech models are available behind one account and one bill.
- ✓Service tiers
Priority runs at 1.5x base price for faster responses under peak load, while Flex runs at 0.8x for non-production work.
- ✓Dedicated GPU deployments
Autoscaling dedicated instances on A100 through B300 GPUs, priced hourly, for teams that need isolation or a custom model.
- ✓Private endpoints and GPU clusters
Private endpoints for compliance needs and rented GPU clusters with SSH access cover workloads that cannot share serverless capacity.
- ✓Zero data retention
The vendor states it does not retain user data, a term HN commenters cite as a reason to route traffic through DeepInfra.
Capabilities
Use Cases
- •Open-model migration
An application built on the OpenAI SDK moves to DeepSeek or GLM models by changing the base URL, cutting per-token cost without a rewrite.
- •RAG embeddings at scale
A retrieval team generates embeddings and reranks results through the same API, paying roughly $0.005 to $0.01 per million input tokens for embeddings.
- •Non-production batch jobs
Evaluation runs and offline enrichment use the Flex tier at 0.8x the base price because latency during peak demand does not matter.
- •Dedicated capacity for a fine-tuned model
A team deploys its own model on a dedicated H200 or B200 instance with autoscaling, keeping traffic isolated while staying on one provider.
- •Speech pipelines
Transcription and text-to-speech run on per-minute audio models alongside LLM calls, so a voice workflow sits on a single vendor contract.
Ideal For
Best For
- ✓Swapping an OpenAI-SDK application onto open-weight models by changing only the base URL
- ✓Cost-sensitive high-volume workloads such as batch summarisation or classification on the Flex tier
- ✓Embedding and reranking pipelines for retrieval systems at sub-cent per million token rates
- ✓Teams that want serverless endpoints now and dedicated B200 or B300 capacity later from one provider
Not Ideal For
- ✗Agentic workloads that lean heavily on prompt caching, where HN users found DeepInfra's cache-read pricing far above DeepSeek's own API
- ✗Organisations that need inference outside the United States, since the homepage describes US-based data centers
- ✗Buyers who need proprietary frontier models such as GPT or Claude, which the catalog does not centre on
Integrations
Deployment
Market Analysis
Pros
- ✓Published, low per-token pricing; HN users describe it as having good prices and being quite fast
- ✓Drop-in OpenAI compatibility and first-party support in the Vercel AI SDK and Hugging Face Inference Providers
- ✓Zero data retention terms that HN commenters call excellent
- ✓Clear upgrade path from shared serverless endpoints to dedicated GPUs on the same account
Cons
- ✗Cache-read pricing was reported on HN as about 6.5 times DeepSeek's own API, which hurts agent sessions that rely on prompt caching
- ✗Model catalog is narrower than some rivals; one HN user summed it up as less model coverage but good prices
- ✗Quantization labelling (FP4 versus competitors' FP8) has confused users comparing providers on the same model
- ✗A researcher on HN found results through OpenRouter too noisy for reliable conclusions and moved to renting GPUs directly
Pricing
Serverless (Standard)
Usage-based, from $0.06 per 1M input tokens
- ✓Per-token billing, no minimums or seat fees
- ✓Input tokens $0.06 to $5.00 per 1M and output $0.15 to $25.00 per 1M depending on model
- ✓Discounted cached-input rates
Priority / Flex tiers
1.5x / 0.8x base price
- ✓Priority for faster responses during peak demand
- ✓Flex for non-production use
Dedicated GPUs
From $0.89/GPU-hour
- ✓A100 $0.89/h, H100 $2.20/h, H200 $2.69/h, B200 $3.69/h, B300 $4.89/h
- ✓Autoscaling dedicated instances
- ✓Private endpoints
Enterprise / DeepCluster
Contact for pricing
- ✓Dedicated hardware procured and operated by DeepInfra
- ✓99.982% uptime SLA
List prices are published. LLMs bill per million input and output tokens (roughly $0.06 to $5.00 input and $0.15 to $25.00 output per 1M, varying by model), cached input is discounted, audio bills per minute, and dedicated GPUs bill hourly from $0.89 (A100) to $4.89 (B300). Priority costs 1.5x and Flex 0.8x the base rate. HN users flag that cache-read pricing can erase the advantage for caching-heavy agent sessions. Enterprise and DeepCluster contracts are quoted by sales. Prices read from deepinfra.com/pricing on 2026-10-07.
Security & Compliance
Sources
This page was written from 7 sources, 5 on domains other than deepinfra.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Ollama
Run open-weight LLMs locally with one command, then burst to Ollama's cloud on the same API
Inferact Platform
Managed vLLM inference from the creators of vLLM, on your own GPUs or on Inferact Cloud
Perplexity SPACE
Open-source Firecracker microVM sandbox for AI agents, the Apache 2.0 runtime behind Perplexity Computer
Gimlet Cloud
Multi-silicon inference cloud that splits AI agent workloads across GPUs, CPUs and accelerators