zymtrace

Zymtrace develops a self-hosted, cluster-wide continuous profiling and optimization platform for AI infrastructure. Using eBPF, its profiler attaches at the CPU level and follows execution down through runtime layers (CUDA, XLA, ROCm, OFI) to the accelerator and individual GPU instructions, building stack traces that span application code (Python, C++, JAX, PyTorch), native functions, CUDA kernels, the Linux kernel and GPU instructions. The platform correlates GPU traces with the exact CPU code paths that triggered them, exposing stall reasons, memory bottlenecks, scheduling inefficiencies and kernel-level hotspots without code changes, recompilation or restarts.

Beyond profiling, Zymtrace surfaces GPU hardware and topology health signals such as NVLink CRC/replay/recovery counts, ECC and XID errors, and non-affine NUMA placement, and turns findings into a ranked list of recommendations with suggested commands and estimated cost and CO2 impact. It also covers inference engines including vLLM, SGLang, llm-d, TensorRT-LLM and Dynamo across prefill and decode phases, as well as distributed training runs. A "Profile-Guided AI Optimization" capability exposes flamegraphs, top functions and GPU metrics as Model Context Protocol tools plus prepackaged agent skills, so coding agents can pull profiling context, diagnose bottlenecks, apply fixes and re-profile; the company states it is building toward a closed loop that detects a GPU bottleneck and opens a pull request with the fix.

Deployment is via Helm charts on Kubernetes (backend plus a profiler DaemonSet) or Docker Compose, with data stored in customer-controlled ClickHouse, PostgreSQL and S3/MinIO backends, and access through a UI, a REST API and MCP. The product is OpenTelemetry-compliant for resource attributes and metadata, and by default uses a custom wire protocol the company says is six times more compute-efficient than the OTel Profiles proto.

Founding story

Co-founders Israel Ogbole (CEO) and Joel Höner (CTO) were part of the team at Elastic that pioneered, open-sourced and donated the eBPF CPU continuous profiling agent to OpenTelemetry, technology now used in production by Cisco, Datadog, Grafana and IBM. They cite the July 2024 release of Meta's Llama 3.1 and the resulting enterprise interest in training, distilling and serving their own models as the trigger: GPUs were spreading beyond frontier labs while teams had no way to see why their workloads stalled. They left their jobs to start Zymtrace and apply the same low-level engineering approach to GPU-bound workloads.

Business model

Zymtrace sells software deployed into customer infrastructure on a self-hosted basis, with tiered plans: a free-forever Basic tier limited to CPU profiling on up to five concurrent machines, a Pro tier priced by machines and GPUs that adds continuous GPU profiling, MIG-aware GPU metrics, unlimited machines, profile-guided optimization, recommendations and alerts, air-gapped installation support and optional managed hosting, and a custom-priced Enterprise tier adding dedicated support, custom SLAs, capacity planning, fully managed deployment and a forward-deployed engineer. Licensed versions are activated with a license key; a free perpetual trial license is included.

Subscription licensing of the self-hosted platform, priced by number of machines and GPUs at the Pro tier and custom-quoted at the Enterprise tier, with professional services (capacity planning, air-gap installation support, training, consulting, forward-deployed engineering) offered to enterprise customers.

Traction

Public customer results include Anam, which reported 2.5x faster inference and 90% higher throughput on its Cara3 model, and Vanellus, whose team reported a 40x faster GPU CFD solver after removing GPU memory transfer bottlenecks identified with Zymtrace. Walmart engineering leadership publicly endorsed the technology and Zymtrace joined the Walmart Sparkcubate program. The company's GPU profiler became publicly available in May 2025, and it reports customers reducing inference latency, increasing throughput and avoiding overprovisioning; documentation claims throughput increases of up to 300% and up to 14% faster performance for unmodified FOSS workloads via profile-guided optimization.

Latest developments

On March 11, 2026 Zymtrace announced $12.2 million in total funding, including an $8.5 million seed round led by Venture Guides with 6 Degrees Capital and Concept Ventures joining and existing investors Mango Capital and Fly Ventures following on, alongside strategic angels including Thomas Wolf (Hugging Face co-founder), Christian Bach (Netlify founder), Christopher Fregly, Jessica (Bartos) Thomas of Notion Capital and Reece Chowdhry of Concept Ventures. Proceeds are earmarked for product development, expanded enterprise deployments and U.S. go-to-market growth, with the roadmap focused on fully autonomous, profile-guided optimization that detects a GPU bottleneck and opens a pull request with the fix. Recommendations and alerts are listed as a coming-soon capability, and support for AWS Neuron, AMD ROCm and TPUs is in progress, with Cerebras support in early access.