Today, we're launching something new on The Compute 100 daily newsletter. On Fridays, we'll share single company deep dives, based on fundraises announced that month. It will cover the technology and opportunity that startups entering the compute market are looking to tackle. Give the first one a read today.
Tendrils, a Cambridge, U.K. startup redesigning the CPU, raised $30M on Nov. 10 in a round led by Benchmark. Jack Altman joined the board, and the UK Sovereign AI Fund invested.
From the Tendrils team:
“Tendrils is answering the question: what if we redo the CPU? Instead of starting from a sequential design, we’re building our processor to have thousands of cores operating in parallel. This requires moving beyond the notion of shared mutable memory and linear instruction sequences. To make this possible, we’re working in tandem on the open-source language Vine, which has been designed from the beginning to target a new wave of parallel-first architectures.” (Tendrils)
Why it matters: CPUs spend most of their transistors and energy finding parallelism in code written as one long list of steps, and agent workloads are asking for far more CPU.
What you'll learn: what a "reorder buffer" and "cache coherency" are, and their impact on the CPU, the bet on interaction nets, and how developers will use Tendrils.
How a CPU finds parallelism today
A program is a to-do list the chip works through in order. To go faster, a modern chip peeks ahead for steps that don't depend on each other and does them early.
Keeping all that juggling straight takes a piece of hardware called a reorder buffer. It makes sure the results still come out in the right order. Intel's newest core can juggle up to 576 steps at once (Intel).
Yet all that bookkeeping finishes only about 12 steps per tick of the chip's clock. A big share of the chip goes to working out what it's allowed to do next, not to doing the work.
The buffer problem
Most of a general-purpose chip's energy goes to fetching, decoding, scheduling and moving data, not arithmetic. One study measured arithmetic at under 10% of total energy (CACM).
Adding cores adds a second tax. Cache coherency is the protocol that keeps every core's copy of shared memory consistent. Tendrils says it, and the reorder buffer, bottleneck all existing CPUs (Tendrils).
A graph that runs itself
Tendrils builds on interaction nets, a model where a program is a graph of nodes. Each node has one main port. When two main ports meet, that "active pair" rewrites into a small new piece of graph (Tendrils, Apr. 26, 2026).
Because a node has only one main port, active pairs never overlap. They can all run at once, in any order, with the same result. The dependency graph is the program, so the hardware has nothing to rediscover.
Each value is used exactly once, so data lives in a core's local SRAM, with no coherence needed. Its language, Vine, borrows Rust-style ownership and compiles straight to these nets.
Software comes first
Tendrils is building hardware and software at once. Its language, Vine, is open source and borrows ideas from Rust, a language many developers already know (Tendrils).
The theory behind interaction nets isn't new. A practical way to write everyday programs for it is, and that's what Vine adds. Software built for today's chips will move to Vine over time, which is why Tendrils is hiring compiler and language engineers alongside chip designers (Tendrils).
What's Next
For years, the AI boom was a GPU story. AI agents are changing that, because the work they do between model calls, like running code, calling tools and juggling tasks, lands on the CPU. Tendrils is designing CPUs, set to benefit from the workloads defining an agentic future. We're excited to see what announcements come out next from Tendrils!