Zymtrace Raises $12.2M to Build the Autonomous Optimization Layer of AI Infrastructure

Today, Zymtrace is announcing $12.2M in total funding, including an $8.5M seed round led by Venture Guides. Mango Capital and Fly Ventures doubled down on their pre-seed conviction, with 6 Degrees Capital and Concept Ventures joining the round.

When Meta’s Llama 3.1 dropped on July 23, 2024, something shifted.¹ The model was good. Remarkable even, though it almost feels funny to say that now given how fast things have moved. But what caught our attention was not the model itself. It was the reaction.

Enterprises that had been watching AI from a distance suddenly began asking how to train on their own data, distill their own models, and serve their own inference. GPUs were no longer just the domain of frontier labs. They were coming to every enterprise that took software seriously.

What we kept seeing was the same pattern: powerful GPU clusters running far below their potential. Kernels stall. Memory bandwidth sits underutilized. Synchronization delays hide inside distributed training. Most teams have DCGM utilization dashboards, but those dashboards cannot tell them why workloads are slow or how to fix them.

Most GPU clusters run at 35-40% utilization. The bottleneck is rarely the hardware. It is the code, and the fact that teams cannot see where their workloads are stalling.

Joel and I saw what that meant. GPUs and AI accelerators were about to be everywhere, yet most teams running them had no way to understand why their workloads were slow.

We quit our jobs and started Zymtrace to fix it.

Where we came from

We had seen this problem before. We were part of the team that pioneered the eBPF CPU continuous profiler and donated it to OpenTelemetry, the same technology now running in production at Cisco, Datadog, Grafana, IBM, and more.

Engineers are incentivized to write performant code, but if the tooling is hard to use, they simply will not. Make it hard to instrument, slow to set up, or expensive to run in production, and the insight never lands.

When we looked at the GPU and AI accelerator space, existing tools showed fragments: a kernel here, a utilization metric there. Nobody had the full story. Cluster-wide, from the CPU code paths down through the CUDA runtime to the exact GPU instructions stalling your training or inference workload. Continuous profiling had been a space full of broken promises - vendors either could not go deep enough, could not scale, or both.

What we built

So we built Zymtrace.

A continuous, cluster-wide optimization platform that profiles GPU and CPU workloads across your entire fleet, correlating code execution down to individual CUDA kernels, stall reasons, and the exact lines of code that triggered them.

We took deployability as seriously as the low-overhead profiling itself. By building close to the bare metal with eBPF, Zymtrace requires zero code changes and zero annotations, and imposes none of the performance tax that makes most profilers unusable in production environments.

It is up and running in five minutes, from day one, against your real workloads.

Customers are reducing inference latency, increasing throughput, and avoiding costly overprovisioning not by buying more hardware, but by finally understanding the code running on the hardware they already own.

We also built Profile-Guided AI Optimization. Zymtrace exposes profiling context directly to AI agents via MCP, so the gap between insight and fix collapses. Instead of a dashboard telling you something is slow, AI agents use Zymtrace’s profiling context to write the fix and open the pull request.

What is next

This funding lets us go further. We are building toward fully autonomous optimization: a closed loop that detects a GPU bottleneck and opens a pull request with the fix. Profile-guided. Agentic. Continuous. Built for the complex, heterogeneous compute environments powering the next generation of AI workloads.

We are also growing the team. We are hiring GTM roles in the US and engineering and other roles across the globe. We are a fully distributed team. It does not matter what passport you carry. Come as you are and do great work.

Thank you

We are grateful to every investor who believed in this before it was obvious.

In the pre-seed, Fly Ventures and Mango Capital co-led alongside Entropy Industrial Capital, and Thomas Dullien, Sean Heelan, Ian Livingstone, and Asaf Ezra backed us as angels before we had much to show. That foundation made everything else possible.

In the seed round, Venture Guides led, with 6 Degrees Capital and Concept Ventures joining. Fly and Mango doubled down.

We are joined by a group of strategic angels and advisors who bring rare depth to what we are building: Thomas Wolf, co-founder of Hugging Face; Christian Bach, founder of Netlify; Christopher Fregly, AI systems optimization expert; Jessica (Bartos) Thomas, Partner at Notion Capital; and Reece Chowdhry of Concept Ventures.

Thank you all for backing the mission.

To our customers (including those we cannot name legally), design partners, and early users: you built this with us. Thank you.

The cheapest GPU you can buy is the one you already own. You just cannot see what is wasting it. Zymtrace it!

Join us

We are hiring. If you want to work on hard infrastructure problems at the intersection of performance engineering and AI, check out our open roles.

If you do not see a suitable role, reach out directly at [email protected].

Footnotes

1. Meta, Open source AI is the path forward, July 2024. ↩