Building AI That Optimizes AI: The Wafer Story

The Shift to AI-Driven Inference

Wafer began as a performance engineering tool called Yaka, designed to use AI agents to optimize C++ code and CUDA kernels. The founders, Emilio Andere and Steven Arellano, initially focused on selling these optimization services to other companies. The pivot to becoming an inference cloud occurred when they realized their internal agents could outperform standard NVIDIA libraries by orchestrating optimizations automatically. By applying these techniques to open-source models like GLM 5.2, they achieved performance levels 2-3x faster than existing market providers, triggering a rapid scale-up from zero to $8M ARR in four months.

The Technical Moat: AI Optimizing AI

Wafer’s core value proposition is "AI that optimizes AI." Rather than relying on static libraries, their system ingests specific workload characteristics and uses agents to rewrite kernels, train speculative decoding models, and quantize models on the fly. This creates a hyper-optimized runtime tailored to the user's specific hardware and model requirements. This approach has allowed them to achieve performance parity between different hardware architectures (e.g., AMD vs. NVIDIA) and deliver latency improvements that significantly alter user behavior in real-time applications.

Speed as the Primary Driver for Adoption

While cost reduction is a common argument for switching to open-source models, the founders argue that speed is the true competitive moat. Enterprises are moving away from proprietary commercial models not just to save on token costs, but because latency-sensitive applications—such as voice agents and coding assistants—require faster response times to achieve product-market fit. The founders note that as open-source models have improved, the ability to serve them at high speeds has become the decisive factor for large-scale enterprise adoption.

Founder Philosophy and Scaling

Andere and Arellano emphasize that their success stemmed from building for fun rather than focusing on business models early on. Their transition from university students to leading a high-growth startup was characterized by intense technical experimentation and a willingness to "run out of money" as a signal of product-market fit. They advise aspiring founders to focus on solving deep technical challenges they are personally obsessed with, rather than seeking a business co-founder immediately, and to prioritize speed of iteration over premature scaling.