Open App
Y Combinator

How YC lowered response latency and increased time spent with its AI partners

In YC’s Office Hour Simulator, GLM-5.2 on Wafer delivered up to 44% lower average latency than OpenAI and Cerebras. Conversations on Wafer lasted 2.5 minutes longer on average.

379 ms
Average LLM latency on Wafer
Up to 44%
Lower average LLM latency vs. the configurations tested
+2.5 min
Longer conversations on average, reported by YC
  • GLM-5.2
  • Dedicated Endpoint
  • Voice AI

Helping More People Start Great Companies

Y Combinator helps more people start great companies. Through Startup School, it makes practical startup education available to anyone, for free. At Wafer, we want more people building companies that improve the world. We’re extremely proud to support YC’s work.

A user seated at YC’s orange Office Hour Simulator booth, with a microphone and a screen showing AI versions of YC partners.
Office Hour Simulator

YC’s Office Hour Simulator lets users talk to AI versions of YC partners. The booth works like a video call: users ask questions and get spoken answers.

Wafer provides the LLM inference behind those answers.

Watch the simulator

How YC Landed at the Lowest Latency

YC had been testing lightweight Gemini and OpenAI models for the simulator. It needed useful answers with low enough latency for a spoken conversation. The avatar waits for the LLM before it can answer, so inference latency directly affects how quickly it can respond.

A Dedicated GLM-5.2 Deployment

We worked with YC to deploy GLM-5.2 on a dedicated Wafer endpoint. Wafer agents tuned the serving setup for YC’s request rate, cache usage, and prompt and response lengths.

YC then launched a three-way production test against GPT-4.1 mini on OpenAI and Gemma 4 31B on Cerebras. The evaluation covered answer quality, latency, and conversation duration.

Lower Latency, Longer Conversations

In a production test covering 4,168 LLM turns across the three configurations, GLM-5.2 on Wafer averaged 379 ms of latency: 31% lower than the OpenAI configuration and 44% lower than the Cerebras configuration.

Average LLM latency

Lower is better

  • WaferGLM-5.2
    379 ms
  • OpenAIGPT-4.1 mini
    546 ms
  • CerebrasGemma 4 31B
    674 ms

4,168 LLM turns across all three configurations.

Conversations lasting 10 to 15 minutes

Higher is better

  • WaferGLM-5.2
    41%
  • OpenAIGPT-4.1 mini
    23%
  • CerebrasGemma 4 31B
    9%

Share of sessions in the 10 to 15 minute duration bucket.

Different models and providers; customer measurements.

YC reported that users talked to Wafer-backed avatars for 2.5 minutes longer on average. Sessions lasting 10 to 15 minutes accounted for 41% of conversations on Wafer, compared with 23% on OpenAI and 9% on Cerebras.

For YC, the result was faster responses and more time spent with its AI partners. We’re proud to provide the inference supporting these tools for founders.

Build Your Voice Application With Wafer

Talk to Wafer about a dedicated deployment for your voice application.

Talk to Wafer
The YC PodcastMore on Wafer’s story and our work with YC