Wafer

Wafer is an AI inference platform that continuously optimizes how models are served. Its dedicated product takes a model, traffic shape and service-level objective, then searches across the model, serving engine, kernels and hardware; the company says candidate configurations must preserve correctness and meet reliability targets before deployment.¹

For self-serve use, Wafer Pass exposes hosted open-weight models through a single API. Y Combinator's current company profile describes Wafer as providing serverless and dedicated inference and says the API is compatible with both OpenAI and Anthropic interfaces.²

Wafer Pass currently lists a $10/week Starter plan with 1,000 requests every five hours and a $25/week Privacy plan with 2,000 requests every five hours and Zero Data Retention; both include all hosted models according to the current YC launch page.²

The Terms state that Wafer Pass is for a single individual's personal use, limits an account to three concurrent in-flight inference requests, prohibits account sharing or pooling, and directs teams and organizations to dedicated or enterprise arrangements.³ Prices, quotas, supported models and overage terms can change with notice, so buyers should verify current entitlements before committing.³

Wafer's own site publishes strong performance claims and customer testimonials, but those remain first-party or customer-attributed claims rather than independent benchmarks.¹ A current Artificial Analysis provider benchmark gives narrower independent evidence: for DeepSeek V4 Flash 0731, Wafer (FAST) measured 247.4 output tokens per second and 10.77 seconds to the first final-response token, ranking third among 14 providers in the displayed speed and latency comparisons.⁴ That result is useful evidence for the tested configuration, not proof of the same performance across every model, endpoint or traffic pattern.

The Terms say users retain ownership of prompts and completions while granting Wafer a limited processing license for operating, securing and billing the service. They also state that availability, supported models, quotas, rate limits and endpoints may change and that Wafer does not promise uninterrupted service or a specific speed.³

Wafer is best suited to developers and infrastructure teams that care about inference latency, throughput and cost but do not want to hand-tune every serving layer. Wafer Pass is particularly relevant to individual developers using coding agents and other harnesses that can consume OpenAI- or Anthropic-compatible APIs.²

Dedicated inference is the more natural fit for teams with a known model, real production traffic and explicit SLOs, because that surface is built around workload-specific optimization rather than a generic benchmark configuration.¹ Organizations needing contractual guarantees, broad procurement documentation or fixed long-term entitlements should evaluate the enterprise route directly rather than assuming Wafer Pass terms apply.