Wafer announced a $40 million Series A on September 1 to expand software that helps developers run AI models more efficiently. Its focus is inference: the work a computer does when a trained model responds to a request. Rather than developing a new general-purpose model, Wafer examines how an existing one runs and changes the surrounding software to improve performance. Its customers are businesses building AI products where response time, computing cost and reliability affect the experience they can offer.
That work happens across several layers. A slowdown might come from requests waiting to be scheduled, data filling the available memory, or the small programs that perform calculations on a graphics processor, known as GPU kernels. Wafer says its agents measure those bottlenecks, test alternative configurations and deploy the best measured result for the target system. The intended advantage is coordination: a change that helps one part of the process can be assessed against what it does to the rest.
The customer workflow starts with a model, a pattern of incoming requests and a performance target. Wafer builds a dedicated endpoint around those requirements—the network service an application calls to get a model's response—and keeps profiling it after launch. The company also describes shared, on-demand model serving and a flat-rate API subscription in its Y Combinator product launch. Those are ways to sell access to optimized computing, not a claim that Wafer manufactures the chips itself.
A company-published case study makes the use case concrete. Neon Health, which builds voice agents for healthcare, moved a model to a dedicated Wafer endpoint. The case study says the customer's measured median wait for the model's first output fell from 800 milliseconds to about 550 milliseconds. That measures one part of the response pipeline, not the complete time from a caller speaking to hearing an answer. It illustrates why infrastructure work can matter to a product whose users notice every pause; it is not an independently audited benchmark or a result guaranteed for other customers.
The financing was co-led by Marathon and Chemistry, with Wing, AMD Ventures and Outset Capital participating alongside existing investors Fifty Years and Y Combinator. Both Wafer's announcement and transaction counsel Wilson Sonsini identify the $40 million Series A. The Information separately reported a valuation above $200 million. That valuation is distinct from the amount raised and remains attributed to the reporting.
Wafer says the capital will help automate more of the optimization cycle. The practical test is whether a configuration that works well on launch day remains useful as traffic changes and new models arrive. Continuing to measure a live deployment could make the service more valuable than a one-off tuning exercise. The business has to deliver that ongoing improvement without turning each customer's setup into a bespoke engineering project.