Gadgets & Hardware

Tendrils Compute Raises $30M to Rethink the CPU for AI Agents

By mohitdeep1122 • • 6 min read • 📷 3 High-Res Visuals
Share Dispatch
𝕏 Post in Share Y Hacker News
⚡ TL;DR — Executive Summary

Tendrils Compute raises $30M in seed funding led by Benchmark to develop a thousands-core parallel processor and its open-source Vine programming language.

Tendrils Compute raises $30 million to develop a new parallel processor architecture designed for the next generation of AI computing.
Tendrils Compute raises $30 million to develop a new parallel processor architecture designed for the next generation of AI computing. 🔍 Click to zoom

Artificial intelligence is driving enormous investment into GPUs, custom accelerators and data center infrastructure, but another part of the computing system is attracting attention: the CPU. AI agents do more than generate text. They execute code, search documents, interact with tools and coordinate multiple operations, creating workloads that can involve thousands of small, unpredictable tasks. Cambridge-based Tendrils Compute wants to rethink the processor responsible for much of this work. The startup has raised $30 million in seed funding led by Benchmark, with participation from Visionaries Club, the UK Sovereign AI Fund, FR8, Antimatter, Cloudberry VC, Conviction, Embassy Ventures and SV Angels. Its ambition is to build a general-purpose processor with thousands of cores operating in parallel, supported by a programming language designed specifically for this new architecture.

Tendrils is building its processor around interaction nets, a graph-based computing model designed to enable thousands of independent operations to execute in parallel.
Tendrils is building its processor around interaction nets, a graph-based computing model designed to enable thousands of independent operations to execute in parallel.🔍 Click image to enlarge

Why Tendrils believes conventional CPU architecture needs a rethink

Modern CPUs have evolved through decades of improvements, including larger caches, branch prediction, multiple cores and sophisticated instruction scheduling. However, Tendrils argues that the conventional computing model still carries structural limitations associated with shared mutable memory and sequential instruction streams. Adding more cores also creates challenges because processors must coordinate access to shared data, maintain cache consistency and manage the order in which instructions execute. GPUs solve a different problem: they are exceptionally effective when thousands of similar mathematical operations can run together, which is why they are so useful for AI model training. But workloads involving changing execution paths, tool calls and irregular operations are not always a natural fit for this approach. Tendrils believes there is room for a third computing architecture designed specifically around fine-grained, irregular parallel workloads.

Interaction nets are the foundation of its new processor

At the centre of Tendrils’ approach is a computing model called interaction nets, originally introduced by mathematician Yves Lafont in 1990. Instead of representing a program primarily as a long sequence of instructions, interaction nets represent computation as a connected graph of small components. These components interact through local rules, changing only the relevant part of the graph. When two independent operations are ready, they can execute simultaneously without needing to coordinate changes to a shared global memory structure. Tendrils wants to translate this mathematical model into physical hardware containing thousands of small processing cores, each working with local memory. The company’s thesis is that reducing the need for complex memory-coordination mechanisms could make it easier to scale parallel computing. It is an ambitious departure from conventional CPU design, not yet a demonstrated replacement for existing processors.

Vine is the software layer designed to make the hardware useful

A new processor architecture cannot succeed through hardware alone. Developers need a practical way to write software for it, which is why Tendrils is building Vine, an open-source programming language developed alongside its processor. Vine is designed to compile programs into interaction nets, allowing the dependencies between operations to remain visible instead of converting everything into a conventional linear instruction stream. Its design draws on ideas associated with Rust, particularly ownership and linearity, while offering a different approach to parallel execution. Tendrils’ own technical demonstration uses a longest common subsequence algorithm to show how independent operations can emerge automatically from an ordinary nested-loop program. Its simulation reports an average parallel breadth of 274 operations for lists of 200 elements and 2,739 for lists of 2,000 elements. These figures describe available parallelism in the computation under an idealised unlimited-parallelism model, not measured performance from a physical Tendrils processor.

Tendrils' open-source Vine programming language is designed to translate software into interaction-net computations for its next-generation processor architecture.
Tendrils' open-source Vine programming language is designed to translate software into interaction-net computations for its next-generation processor architecture.🔍 Click image to enlarge

AI agents could become an important early use case

The timing of this architecture is closely connected to how AI software is changing. An AI agent may need to search several documents, execute code, test the result, call external tools and change its approach depending on what happens next. These operations can involve different execution paths and many independent tasks, rather than one large batch of identical calculations. Tendrils believes its architecture could be suited to this type of computing. Other potential workloads include compilers, simulations, optimisation, sparse graph processing, proof checking and option pricing. This does not mean the processor is designed to replace GPUs used for large-scale model training. Its initial opportunity is more specific: handling the complicated computing tasks that surround AI inference and agent execution. If the architecture proves effective, it could eventually complement GPUs and conventional CPUs in a more specialised computing infrastructure.

Benchmark is backing a hardware and software team

Tendrils’ founding team combines expertise in computer science, programming languages and open-source hardware. Nils Cremer worked on interaction nets while studying computer science at ETH Zurich. Co-founder T6 developed Vine and its supporting toolchain, while Hugo McNally previously worked at lowRISC, the Cambridge organisation focused on open-source chip architecture. Visionaries Club supported Cremer with an early grant and led the company’s pre-seed financing in summer 2025 alongside FR8 and angel investors. The new $30 million round is led by Benchmark, with the UK Sovereign AI Fund and several venture investors participating. Benchmark general partner Jack Altman has also joined the board. The backing reflects investor interest in alternative computing architectures at a time when AI is increasing demand for more diverse forms of processing capacity.

The biggest challenge is moving from a promising model to working silicon

Despite the funding and technical ambition, Tendrils still faces the difficult engineering work required to turn its architecture into a commercially useful processor. The company has not publicly disclosed measured hardware benchmarks, a chip tape-out timetable, a fabrication partner or commercial pricing. It must prove that interaction nets can deliver meaningful improvements in real workloads, not just attractive results in software models. It also needs a reliable compiler, development tools, debugging support and an ecosystem that gives developers a reason to move existing applications to Vine. Established CPU and GPU platforms benefit from decades of optimisation, mature software libraries and enormous developer communities. Competing with them requires considerably more than demonstrating that thousands of small operations can run independently.

Tendrils is betting on a third architecture in the AI computing stack

The significance of Tendrils’ $30 million seed round extends beyond one more semiconductor startup entering the AI infrastructure market. The company is exploring whether general-purpose computing can be built around parallelism from the beginning, rather than adapting architectures designed around older execution models. Its strategy combines a new processor concept, an open-source programming language and a software toolchain intended to make the entire system work together. If the company can translate its architectural ideas into reliable silicon and demonstrate advantages in real workloads, it could establish a specialised position alongside CPUs and GPUs in the next generation of AI infrastructure. For now, Benchmark and its co-investors have funded the attempt. The decisive milestones will be working hardware, independently verifiable performance and developers choosing to build on it.

Sources and further reading

📷 Complete Article Visual Gallery (3 Pictures) Click any photo for full-screen view
Tendrils Compute raises $30 million to develop a new parallel processor architecture designed for the next generation of AI computing.
Tendrils Compute raises $30 million to develop a new parallel processor architecture designed for the next generation of AI computing.
Tendrils is building its processor around interaction nets, a graph-based computing model designed to enable thousands of independent operations to execute in parallel.
Tendrils is building its processor around interaction nets, a graph-based computing model designed to enable thousands of independent operations to execute in parallel.
Tendrils' open-source Vine programming language is designed to translate software into interaction-net computations for its next-generation processor architecture.
Tendrils' open-source Vine programming language is designed to translate software into interaction-net computations for its next-generation processor architecture.
mohitdeep1122 • TechFront360 Editorial Desk

Published by TechFront360.com following our strict fact-checking, benchmark verification, and independent technology journalism standards.

More in Gadgets & Hardware