What tech stack does Skild AI use?

Skild AI's core model training infrastructure is anchored in NVIDIA GPU compute delivered via HPE Cray XD670 systems populated with NVIDIA HGX H200 accelerators — the highest-density GPU configuration available for large-scale AI training at the time of deployment. The March 2025 HPE partnership press release detailed this stack: HPE Cray XD670 systems handle large language model training, natural language processing, and multimodal training workloads for the Skild Brain, while HPE ProLiant DL380a Gen12 servers with eight NVIDIA L40S GPUs each support model visualization and evaluation workflows. NVIDIA Jetson hardware provides real-time, on-robot inference at the edge for deployed robot embodiments in the field.

For inference and burst GPU compute, Skild uses STN's GPU One platform, hosted at CoreSite's CH2 data center in Chicago. The GPU One cluster, launching in April 2025, comprises more than 1,500 liquid-cooled NVIDIA B200 GPUs in SuperMicro servers across 24 racks within a 2.4 MW data center infrastructure, with direct high-speed connectivity to AWS, Microsoft Azure, Google Cloud, and Oracle Cloud for cloud overflow capacity. CoreSite has specifically named Skild Brain as a workload the cluster is designed to support.

For simulation and synthetic data generation, Skild leverages NVIDIA Isaac Lab (the industry-standard framework for robot learning), NVIDIA Isaac Sim (photo-realistic robot simulation), NVIDIA's Newton physics engine for high-fidelity contact dynamics, and NVIDIA Cosmos world foundation models for generating synthetic training episodes at scale — enabling the training of trillions of simulated robot action sequences that would be impossible to collect in the physical world.

Skild's training data pipeline is a proprietary multi-source system combining three distinct input types: trillions of synthetic simulation episodes generated via the NVIDIA Isaac and Cosmos stack; millions of real-world human manipulation and locomotion video frames ingested from internet-scale video sources; and live teleoperation data streams collected from robots deployed in the field by enterprise customers. This pipeline is custom-engineered — no commercial off-the-shelf data pipeline product is identified in Skild's public disclosures, and the company regards the architecture and data curation methodology as a core competitive moat.

On the GTM and operations side, Skild uses Greenhouse as its applicant tracking system, confirmed by the public job board at job-boards.greenhouse.io/skildai-careers, which lists open roles across simulation engineering, robot learning, OEM integration, and enterprise deployment. LinkedIn Recruiter is almost certainly in use given Miriam Elmasry's (Head of Talent Acquisition) background at Tesla and Uber, where LinkedIn Recruiter is the standard recruiting tool. The company's public-facing website (skild.ai) is maintained with engineering blog posts, press releases, and social media presence across X/Twitter, LinkedIn, and Instagram.

No CRM, marketing automation, or sales engagement platform is publicly identified — consistent with a ~64–100-person company that remains primarily founder-led in enterprise sales, without a formal SDR or AE motion as of June 2026. This will almost certainly change as the Zebra acquisition brings a larger enterprise customer base and the need for more structured account management.

Skild AI is deeply and durably committed to the NVIDIA ecosystem — from training hardware (HGX H200) to inference chips (L40S, B200, Jetson) to simulation tooling (Isaac Lab, Isaac Sim, Cosmos) — making it an unlikely buyer of competing GPU or simulation platforms from AMD, Intel, or non-NVIDIA simulation vendors. NVIDIA's Series B and Series C investments in Skild reinforce this alignment structurally: Skild's success drives NVIDIA GPU demand, and NVIDIA has an incentive to help Skild win. Vendors who natively integrate with or extend the NVIDIA Isaac platform have the highest entry opportunity.

On cloud infrastructure, the STN/CoreSite partnership covers current GPU burst needs, but the scale of Skild's training ambitions (trillions of simulated episodes) and the post-Zebra integration workloads create plausible demand for hyperscaler GPU capacity (AWS, GCP, Azure) for redundancy and geographic overflow — making cloud GPU reservation, spot-instance management, and MLOps tooling relevant vendor categories. The April 2026 Zebra acquisition creates the most immediate near-term buying signal: integrating the Symmetry Fulfillment fleet-orchestration software with Skild's model-serving API layer will require enterprise software integration middleware, API gateway and management tooling, and data warehouse / analytics infrastructure to unify operational telemetry from both systems.

Vendors offering data infrastructure (Snowflake, dbt, or comparable), API integration middleware, enterprise data labeling and curation platforms, or professional services for robotics software integration have a credible pitch during the 2026–2027 integration window. Sellers should target Deepak Pathak (CEO) for infrastructure and platform decisions, Abhinav Gupta (President) for model-training and simulation tooling, and Miriam Elmasry (Head of Talent Acquisition) for recruiting and people-ops software.