Wafer is a San Francisco-based inference provider serving open-source LLMs through a serverless API and dedicated endpoints. Its platform uses AI agents to profile workloads and optimize the model, serving engine, kernel, and hardware combination for each deployment.
What's good about Wafer
- High-throughput serving with results independently benchmarked on Artificial Analysis
- Prompt caching with discounted cached-input rates
- Dedicated endpoints set up in under 24 hours
- Runs across NVIDIA, AMD, and AWS Trainium accelerators rather than a single vendor
Wafer does not publish public hourly pricing.
Contact the provider directly for a quote.
Billing & terms
- Pay-per-token
- Prepaid credits with separate input and output token rates per model and no subscription
- Cached input
- Reduced rates for cached input tokens on supported models
- Dedicated endpoints
- Custom pricing for dedicated deployments with tuned performance targets, arranged with the sales team
- Support
- Documentation at docs.wafer.ai, email support (hi@wafer.ai), and scheduled onboarding calls for enterprise
Watch out for
- Small serverless model catalog centered on the GLM family
- Inference only - no GPU rental, training, or fine-tuning hosting
- Early-stage company with a smaller ecosystem than established inference providers
Features
- Serverless Inference
- Pay-as-you-go API access to hosted open-source models including GLM, Kimi, Qwen, and DeepSeek with no infrastructure management
- Dedicated Endpoints
- Custom-tuned inference deployments with performance guarantees, provisioned in under 24 hours
- Automated Performance Optimization
- Agents profile inference bottlenecks and tune across serving engines (vLLM, SGLang, TensorRT-LLM), custom kernels (CUDA, HIP, Triton, NKI), quantization (FP8/FP4), and decode strategies
- Multi-Accelerator Hardware
- Workloads run on NVIDIA B200/B300, AMD MI350X/MI355X, and AWS Trainium depending on the model and traffic shape
- OpenAI- and Anthropic-Compatible API
- OpenAI-compatible endpoint at pass.wafer.ai/v1 and Anthropic-compatible endpoint at pass.wafer.ai/v1/messages, both using Bearer token authentication
- Prompt Caching
- Cached input tokens are billed at reduced rates on supported models
- Zero Data Retention
- Request-scoped ZDR available on direct API calls via a Wafer-ZDR header for privacy-sensitive workloads
Getting started
Step 1
Create an account
Sign up at app.wafer.ai and load credits for pay-as-you-go usage
Step 2
Generate an API key
Create a key in the console and pass it as a Bearer token
Step 3
Make your first request
Call the OpenAI-compatible endpoint at pass.wafer.ai/v1 with a model from the serverless catalog
Where this page's facts come from
- Profile
- Reviewed Jul 7, 2026
- Method
- How we collect and check prices
Frequently Asked Questions
Which models does Wafer serve?
Check the pricing table above for the models Wafer currently serves and their per-token rates.
How do I get started with Wafer?
Create an account, Generate an API key, Make your first request
What are Wafer's main advantages?
Wafer's main advantages include: High-throughput serving with results independently benchmarked on Artificial Analysis, Prompt caching with discounted cached-input rates, Dedicated endpoints set up in under 24 hours, Runs across NVIDIA, AMD, and AWS Trainium accelerators rather than a single vendor.
What are Wafer's limitations?
Wafer's main limitations include: Small serverless model catalog centered on the GLM family, Inference only - no GPU rental, training, or fine-tuning hosting, Early-stage company with a smaller ecosystem than established inference providers.
The weekly price report
One email every Monday with the week's notable GPU and LLM API price moves, plus occasional product news. No spam.