Gimlet Labs announced a $300 million Series B on September 4 to expand its approach to running artificial intelligence across different kinds of processors. Its customers are organizations operating large AI workloads, including frontier model developers. The company focuses on inference: the computing needed when a trained model produces an answer or takes part in an agent's work.
The premise is that an AI request is not one uniform computing job. Reading a long prompt and generating the next pieces of an answer place different demands on processors and memory. Gimlet's software divides that work into stages and assigns them to suitable hardware, bringing GPUs, CPUs and other accelerators into the same system. The intended benefit is faster responses or more useful output from the available computing capacity.
For a developer, the offering is a managed service rather than a requirement to arrange every processor manually. Gimlet's product page describes agents that combine models with search, custom code and outside data sources. A customer could connect those stages through its inference interface while Gimlet handles scheduling and scaling. Large customers can also deploy the stack in their own data centers. Public materials describe the delivery model but do not disclose standard pricing.
That software has pulled the company into physical infrastructure. CEO Zain Asgar told Bloomberg that customers include AI labs and financial-services companies, without naming them, and that Gimlet now helps configure data centers while developing facilities of its own. Different chip types can need different cooling arrangements and connections. In its announcement, Gimlet also describes billions of dollars in contracted revenue; that is a company claim about contracts, not revenue already recognized.
Andreessen Horowitz led the round, according to Gimlet, with participants including Arm, Microsoft's M12, Sapphire Ventures and Menlo Ventures. Bloomberg, citing an interview with Asgar, reported a $3 billion valuation. The financing follows Gimlet's $80 million Series A in March. The company says it is expanding its engineering, research and infrastructure team as it adds capacity.
The business question is whether customers can get the benefits of specialized chips without inheriting all the complexity of connecting them. Gimlet's mix of software and infrastructure is meant to absorb that work. Its next test is delivering dependable speed and capacity across customer workloads, not only demonstrating a faster model run under selected conditions. Published performance claims remain vendor claims, not independent benchmarks established by Captables.