Tendrils: Rebuilding the Microprocessor from First Principles

Every big leap in computing power has unlocked things almost nobody saw coming: the personal computer, the App Store, and now AI. It will be just as hard to predict what becomes possible over the next ten years. What is almost certain is that we will put a step change in computing power to use. Yet today's CPU architectures won't scale to deliver it. While most of the attention has gone to the next generation of GPUs, the CPU is increasingly becoming the bottleneck, and AI is simply where it shows first.

Just watch an AI agent at work and a different picture emerges. The GPU thinks, the CPU does. The GPU generates the next step, the CPU runs the Python, compiles the code, executes the tests, searches and filters the documents. In some agentic workloads, that tool execution on the CPU accounts for up to 90% of end-to-end latency. Adding cores helps throughput, but each individual tool call still runs only as fast as one core allows. And per-core speed has hit a wall. Clock speeds stalled in the mid-2000s, and per-core performance gains have slowed to roughly 5–10% a year.

The industry's answer has largely been to bolt ever more cores onto architectures never designed for this level of parallelism, paying for it in cache-coherency protocols, race conditions and a zoo of specialised accelerators papering over the cracks. Underneath, today's chips still carry the DNA of the earliest x86 and Arm designs. And by some estimates, up to 99% of the energy in modern computing architectures is spent moving data back and forth between memory and processors rather than doing the actual computation.

That can't be the future.

It started with a few lunches in Munich

Some of the best introductions in venture arrive with a single sentence attached. This one came from one of my most respected angel investor friends, who passed on something an ETH Zurich professor had said about a student: "Maybe the best computer science student I've seen in a decade."

At ETH, that means a lot!

That student was Nils.

Our first lunches were at a little place called Joy of Madu in Munich, while he was still finishing his Master's thesis. I'll be honest: I spent much of those lunches listening. Following every layer of what Nils described, from the theory all the way down to the transistor, took real effort on my side. But what struck me most was something else. It was a feeling I've only had a handful of times before, with founders like Jonathan Ross when he started Groq and we were lucky enough to back him, Daniel Metzler when he was setting out to build Isar Aerospace, or Julie Boehning when she co-founded PAVE Space.

Not a trace of doubt about why their idea could work. And a path they could see all the way down that seemed almost a no-brainer to them, even when it sounded impossible to everyone else. Seeing that level of conviction in the eyes of a young student is rare. For Jonathan at Groq, it was the idea of a radically simple processor architecture after coming out of Google, with the TPU already behind him. For Nils, it was a completely new processor built on interaction nets, a model of computation he'd been obsessed with since first coming across it in 2024.

Being crazy enough to actually build it

We wanted to lead a pre-seed round right away. But Nils was humble and thoughtful enough to ask us to wait until he had the right co-founder setup in place. That kind of judgement is rare at any age, let alone straight out of university.

So we started supporting him with a small grant, before there was even a round to lead. And when he brought in his co-founders T6 and Hugo, they continued to blow me away. T6 had been working on algebraic problems since the age of three or four. He went on to build Vine, a programming language based on interaction nets, and made fascinating contributions to Parity. Hugo had worked at lowRISC, an independent company out of Cambridge working on open-source chip architectures. The references on all three were unlike almost anything we'd heard before. You rarely hear someone described as being in the "top 0.0000001%". And almost never with that level of consistency.

In the summer of 2025, we were incredibly excited to lead Tendrils' pre-seed round as their first institutional believers, alongside a few angels and our friends at FR8. Tendrils also brought us a wonderful friendship with the FR8 team, whom we deeply admire. We had already had the pleasure of backing more founders like Otso and The Token Company from their very first batch.

Parallel by default

For fifty years, computing has run on the same bargain. Flatten a program into one long line of instructions, then build ever cleverer hardware to guess which of them could have run at the same time. Compilers know exactly what depends on what. They throw that knowledge away, and the chip burns power and silicon trying to win a little of it back.

Tendrils refuses the bargain. Their processor runs on interaction nets: a graph model of computation where every step is a small, local rewrite, independent of all the others. Nothing is shared, so nothing needs to be kept in sync. Everything that can run in parallel does, and the answer comes out the same in any order. That one choice deletes problems the industry has spent decades managing: cache coherency, race conditions, garbage collection, and the out-of-order machinery that eats a modern core. What's left is a simpler core that can scale to thousands. And developers don't have to change how they think. T6’s language, Vine, reads like Rust and compiles straight to the graph. Write an ordinary nested loop and it exposes hundreds of operations ready to run at every step, more as the problem grows. No threads. No CUDA. The parallelism was in the problem all along; Tendrils simply stopped throwing it away.

If it works, this isn't a faster chip. It's a reset of general-purpose computing, and the substrate the next wave of AI will run on. We're talking about the possibility of resetting the foundations of general-purpose computing and building a new substrate for the next wave of AI.

Today, we're thrilled to welcome Benchmark, leading Tendrils' $30 million seed round

When we first met Nils over lunch in Munich, this was still an extraordinarily ambitious idea in the mind of a student. Today, there's a remarkable team turning that idea into reality, with the Benchmark team joining the journey. There's still an enormous amount to prove. Building a fundamentally new computing architecture is about as hard as it gets. But that's also what makes this journey so exciting.

The next generation of computing might look very different from the last. And we can't wait to find out.