San Francisco-based Wafer closed a $40M Series A, co-led by Marathon and Chemistry, after rapidly scaling its AI driven inference optimization platform for open source models to $8M ARR with a team of under 10.
Wafer raised a $40 million Series A at a valuation of approximately $200 million (or higher), co-led by Marathon and Chemistry. Participating investors included Wing Venture Capital, AMD Ventures, Outset Capital, and existing backers Fifty Years and Y Combinator. A strong group of angels joined, among them Jeff Dean (CEO of DiscoveryLoop), Guillermo Rauch (CEO of Vercel), Andy Fang (CTO of DoorDash), Kyle Vogt, Akshay Kothari (COO of Notion), Matthew Prince (CEO of Cloudflare), and Scott Stephenson (CEO of Deepgram).
What is Wafer.AI?
Wafer was founded in 2025 by University of Chicago alumni Emilio Andere (CEO; background in ML security research published at NeurIPS and weather modeling at Argonne National Lab) and Steven Arellano (experience in high performance computing and AI infrastructure at Google/Bard, Two Sigma, and Sei Labs). The company emerged from Y Combinator’s Summer 2025 batch. It raised a $4 million seed in April 2026 led by Fifty Years (with Liquid2, YC, and early angels including Jeff Dean and OpenAI co-founder Wojciech Zaremba), followed by this Series A roughly five months later. Total funding stands near $44.5 million. The team remains small, around 8–10 people, based in San Francisco and operating fully in-person.
Wafer builds AI agents that act as autonomous performance engineers. These agents continuously profile production workloads, diagnose bottlenecks, test changes across the full stack (model variants, inference engines, custom kernels, quantization, caching, scheduling, and hardware), verify correctness and reliability, and deploy improvements. The core thesis is that inference optimization should be continuous and automated rather than a one-time, manual, expert heavy process performed before deployment. The stated mission centers on maximizing “intelligence per watt.”
The company pivoted from primarily selling optimization tooling/agents to third parties toward operating its own high performance inference cloud focused on open source models. It offers serverless and dedicated endpoints that deliver the cost advantages of open models with latency competitive with (or better than) much smaller or proprietary alternatives. It supports heterogeneous hardware including NVIDIA, AMD, and others.
In approximately three to four months after launching its inference cloud, Wafer reached $8 million in annual recurring revenue (ARR), with reports of 30x revenue growth over three months and service of roughly 2 trillion tokens per month at peak scaling phases with a team of about seven. Demand outstripped GPU capacity, prompting the raise. Named customers and users include Vercel and Inworld AI; testimonials and case studies reference Neon Health (lowest observed latency that holds under load), DigitalOcean (significant speedups on AMD for models such as Kimi), and others. Independent checks, such as Vercel’s tests on GLM-5.2, have shown roughly 2x throughput versus other serverless providers.
Performance claims include running models such as GLM-5.2 on AMD MI355X at ~80% of the throughput of NVIDIA’s B200 at less than half the cost, and delivering industry leading “tokens per second” figures (e.g., thousands of tok/s per node on various open models) while preserving accuracy. The product emphasizes continuous adaptation: when traffic patterns shift, models update, or new hardware appears, the system re-optimizes.
Recommended: Dimension Energy Receives $857 Million In Additional Capital
The round occurred amid rising inference costs, persistent low average GPU utilization (often cited around 20% in production), growing enterprise interest in open source models for cost and customization reasons, and hardware diversification beyond NVIDIA. Wafer’s ability to make alternative accelerators competitive strengthens the case for multi vendor strategies and improves economics for high volume open model serving. Speed (latency and throughput under realistic traffic) has emerged as a stronger driver of adoption than pure cost in many accounts.
The company turned down acquisition offers from multiple cloud and inference providers, preferring to remain independent and scale. Proceeds are directed at automating more of the optimization loop, expanding capacity to meet demand, and growing the team. Immediate hiring focuses on technical staff (kernels, engines, serving infrastructure, customer workload ownership; compensation in the $200k–$300k base range plus equity), growth/marketing, chief of staff, and go to market roles, all requiring full time, five days a week presence in San Francisco.
The $40M Series A at a ~$200M+ valuation represents a roughly 50x step-up from the seed in five months, reflecting exceptional early revenue velocity, technical differentiation, and investor conviction in continuous, agent driven inference optimization as a foundational layer for the next phase of AI infrastructure. AMD Ventures’ participation aligns with demonstrated gains on non NVIDIA hardware. The combination of a tiny team, multi million ARR, hardware agnostic performance leadership on open models, and high profile technical and commercial backers positions Wafer as a standout in the crowded AI infrastructure landscape, with capital now available to convert demand into scaled, continuously improving production systems.
Please email us your feedback and news tips at hello(at)superbcrew.com
