Building AI That Optimizes AI

Wafer, co-founded by Emilio Andere and Steven Arellano, is an AI inference cloud provider specializing in running open-source AI models at industry-leading speeds. The company, which recently secured a $40 million Series A funding round, achieved remarkable growth by leveraging AI agents to optimize GPUs at every level of the stack, from custom kernels to speculative decoding.

Unprecedented Growth and the GLM 5.2 Launch

Within just four months, Wafer grew from zero to $8 million in Annual Recurring Revenue (ARR). This explosive growth was largely driven by a strategic shift to optimize open-source Large Language Models (LLMs). A pivotal moment was their optimization of the GLM 5.2 model, where their agents, through custom kernel rewriting and full-stack GPU optimization, achieved 2 to 3 times the performance of competitors. This breakthrough generated immense demand, leading to a rapid consumption of GPU resources and necessitating their recent funding round.

Jared Friedman of YC highlighted that this success validated his advice: to launch, let demand outstrip supply, and then use that growth story for fundraising. Steven Arellano described the chaotic but rewarding experience of scaling compute, including onboarding Blackwell nodes from a hotel room during an offsite.

Technology: AI Optimizing AI

Wafer’s core innovation lies in its “AI that optimizes AI” approach. They feed workload characteristics from customers into an AI blackbox. This blackbox then performs various optimizations, such as writing custom kernels, training speculative decoding models, and quantizing models. The output is a hyper-optimized engine or runtime specifically tailored for the user’s workload, yielding 2x, 3x, or more speedups over other inference runtimes. This technology even enabled them to bring AMD hardware to “perf-per-dollar parity” with NVIDIA. The genesis of this technology dates back to their YC days, where they began by using agents to optimize CUDA kernels, effectively acting as a “cursor for CUDA.”

A notable anecdote involved an agent beating “some of these core NVIDIA libraries” overnight, leading the founders to realize the immense value of their optimizations and the potential to verticalize into an inference cloud rather than licensing their optimization tech.

Customer Success and the Importance of Speed

Wafer serves a diverse customer base, including voice agent providers like Neon Health and high-throughput applications like coding assistants for companies like Vercel. Neon Health switched to Wafer due to a need for low latency in their hospital automation agents, achieving 30-50% better performance.

Y Combinator itself is a customer, using Wafer for its “Office Hour Simulator” – an AI clone of a YC partner. Jared Friedman noted that Wafer’s optimized GLM 5.2 was both smarter and faster than proprietary models (Gemini, OpenAI) they had previously benchmarked. An A/B test revealed that users interacting with the Wafer-powered GLM model stayed and talked significantly longer, demonstrating a direct correlation between reduced latency and improved user engagement and product market fit. Jared emphasizes: “Speed is the moat.”

Emilio and Steven underscored that while cost savings are a factor, speed is increasingly the primary driver for enterprises adopting open-source models. For many applications, the core product’s effectiveness hinges on the speed of response, making the ability to deliver “opus-level model that will respond five times faster” a significant value proposition.

Founding Story and Culture

Emilio and Steven met at the University of Chicago. Their journey began with a side project named “Yaka,” developed by Steven. Yaka used an LLM to iteratively optimize C++ code for speed, driven by Steven’s experience in high-frequency trading (HFT) where micro-optimization is crucial. This project, initially pursued for fun and technical interest, evolved into Wafer after the founders realized its immense potential impact and their unique suitability for the challenge. They turned down prestigious job offers to pursue the startup, encouraged by YC’s advice that “you don’t need a business co-founder.”

Advice and Hiring

Their advice for aspiring founders or those joining startups: if offered a seat on a “rocket ship,” don’t overly focus on the specific role, as it will change rapidly. Instead, evaluate the founders and the company culture for its potential for massive growth. Wafer is actively hiring “Members of Technical Staff” and seeks individuals who have demonstrated exceptionalism in any field, regardless of prior GPU or computer science experience. A “day in the life” at Wafer involves being a problem solver, tackling challenges ranging from agent-written kernels to reliability issues with GPUs, all within the domain of ML performance and systems.

• Tags:
• LLM
• Speculative Decoding
• Machine Learning
• Latency Optimization
• Open Source Models
• Y Combinator
• AI Inference
• GPU Optimization
• Startup Growth
• GLM 5.2
• ML Systems
• Custom Kernels