Zymtrace Secures $12.2M to Recover Billions in Wasted GPU Spend Through Autonomous Optimization

Zymtrace, a distributed AI infrastructure optimization platform, announced that it has raised $12.2 million to date to help enterprises uncover hidden performance bottlenecks inside their GPU clusters.The funding includes a newly closed $8.5 million seed round led by Venture Guides, an early-stage investor in cloud infrastructure and AI companies, with participation from Mango Capital, Fly Ventures, and 6 Degrees Capital.

The round also includes strategic angel investors Thomas Wolf, co-founder of Hugging Face; Christian Bach, founder of Netlify; AI systems optimization expert Christopher Fregly; Reece Chowdhry of Concept Ventures, and more.

The company also previously raised an unannounced $3.7 million pre-seed round led by Fly Ventures and Mango Capital, with participation from Entropy Industrial Capital. In the latest financing, Mango Capital and Fly Ventures doubled down on their investment, reinforcing their conviction and support in Zymtrace’s mission.

The new funding will support continued product development, expanded enterprise deployments, and growth of the US go-to-market team, while advancing Zymtrace’s move toward profile-guided autonomous AI workload optimization. The company’s Profile Guided AI Optimization approach completes the full agentic optimization loop autonomously, from detecting a GPU bottleneck to opening a pull request with the fix. With MCP integration, customers can wire it directly into their existing pipelines, cutting what once took weeks of manual investigation down to minutes.

The founding team pioneered, open-sourced, and donated the eBPF CPU profiling agent to OpenTelemetry while at Elastic. This technology is now used in production at Cisco, Datadog, Grafana, IBM, and more. At Zymtrace, they are bringing that same engineering excellence to GPUs and AI-accelerated workloads.

Continuous GPU Profiling for Production AI Workloads
As AI adoption increases, infrastructure spending has risen significantly, with the global GPU market expected to reach $326 billion by 2036. Yet, most GPU clusters operate at just 35-40% utilization, wasting billions of dollars in compute capacity.

This inefficiency is a massive economic drain. Underutilized GPUs lead to longer training cycles, costly inference, and wasted energy. When performance bottlenecks arise, identifying the root cause is no simple task. It demands highly specialized expertise and days or weeks of manual investigation across fragmented tools. As a result, many organizations default to a costly stopgap: buying more GPUs.

At the heart of the problem is a lack of fleet-wide, production-grade visibility. Existing solutions are intrusive, fragmented, and blind to the critical interactions between hosts and GPUs. They show utilization%ages. They don’t show why.

Zymtrace was built to close that optimization gap. The platform continuously profiles GPU and CPU workloads across distributed systems, correlating cluster-level activity down to individual lines of code. Engineers can trace GPU kernel stalls, memory bottlenecks, or scheduling inefficiencies back to specific CUDA kernels, Python functions, Rust or C++ routines, without requiring code changes.

“The cheapest GPU you can buy is the one you already own,” said Israel Ogbole, co-founder and CEO, Zymtrace. “The bottleneck is rarely the hardware. It’s the code that runs on it. Every idle GPU cycle is money and energy lost. We are building the autonomous optimization layer for AI infrastructure, improving unit economics with more throughput per GPU, lower cost per inference, and less energy per output.”

Customers have used Zymtrace to reduce inference latency and improve GPU throughput while avoiding costly overprovisioning. To cite an example, “before Zymtrace, we spent so much time hunting down why our GPUs were being used inefficiently,” said Ben Carr, co-founder and CTO, Anam. “Zymtrace pinpointed where our workloads were stalling and showed us how to resolve the issues. We improved inference latency by 2.5x and increased throughput by 90% for our Cara3 model.”

Unlike traditional profiling tools that can introduce significant overhead in production environments, Zymtrace uses an eBPF-based architecture designed for continuous introspection with minimal performance impact. The platform generates actionable optimization recommendations across kernel execution and batch sizing, CPU scheduling, and distributed communication, along with estimated cost and performance gains.

Scaling the Next Layer of Efficient AI Infrastructure
As AI infrastructure costs continue to rise, Zymtrace aims to become a critical efficiency layer for enterprises running large-scale AI workloads. Here’s what some of the investors backing Zymtrace had to say.

“Zymtrace is creating core technology that will underpin the next gen of AI infrastructure. As infrastructure increasingly becomes the limiting factor to growth, performance gains and efficiency aren’t optional, they’re essential,” said Sage Nye, partner and founding team member, Venture Guides. “With a strong focus on customers and a clear long-term vision, the Zymtrace team is addressing one of the most significant challenges in GenAI adoption.”

“The future of AI won’t only be defined by who can acquire the most GPUs, but by who gets the most out of them. As compute becomes the dominant cost center, Zymtrace is solving a problem every AI-driven enterprise will face,” said Fredrik Bergenlid, partner, Fly Ventures.

“Most organizations are still flying blind inside their GPU clusters, unable to see why their most expensive resources are sitting idle,” said Robin Vasan, founder and managing partner, Mango Capital. “The teams that can squeeze the most FLOPs from their GPU will have a decisive competitive advantage. That’s exactly why we backed Zymtrace from day one.”