Stack Analysis: Workday — Human Capital Intelligence Driven by Distributed AI

Company Overview

Workday is a leading provider of enterprise cloud applications for finance and human resources. They have established a strong market position by offering comprehensive solutions that help businesses manage their most important assets: people and money. In 2026, Workday's competitive advantage hinges on its ability to leverage AI to deliver increasingly intelligent and automated workflows across its HCM suite.

Core AI/ML Stack

Workday employs a multi-faceted approach to AI/ML, relying on a blend of open-source frameworks and internally developed tools. Their core stack consists of:

• Model Frameworks: Primarily PyTorch 3.x for deep learning tasks, supplemented by JAX for research and computationally intensive projects like large language model (LLM) fine-tuning. TensorFlow is used for some legacy systems but is gradually being phased out.
• LLMs: Workday leverages a combination of proprietary LLMs trained on their extensive HR data and fine-tuned versions of open-source models like Llama 4 and Falcon-2 180B. These models are used for tasks like sentiment analysis, skills extraction, and personalized learning recommendations.
• Training Infrastructure: Workday utilizes a hybrid cloud approach for model training. Large-scale training jobs are run on NVIDIA H200 Tensor Core GPUs in AWS and Google Cloud, utilizing Kubernetes for orchestration. They are also exploring AMD Instinct MI400 series GPUs. Federated Learning is crucial, using a custom framework built on top of TensorFlow Federated to train models on anonymized and aggregated data from their customer base.
• Feature Store: They use Feast for managing and serving features for both training and inference. A key feature is real-time feature engineering using Spark Streaming and Flink for streaming data ingestion.

Hardware & Compute Infrastructure

Workday's compute infrastructure is primarily cloud-based, leveraging AWS and Google Cloud regions globally. They operate a smaller on-premise data center for development and testing, equipped with a mix of NVIDIA and AMD GPUs. Key details include:

• GPU Infrastructure: Heavy reliance on NVIDIA H200 GPUs for high-performance training. Investigating the performance and cost-effectiveness of AMD Instinct MI400 GPUs, especially for inference workloads.
• Networking: Utilizing InfiniBand HDR (200 Gbps) for low-latency communication between GPU servers within their data centers. Cloud regions leverage cloud provider's internal networking fabrics.
• Data Storage: Primarily relies on cloud-based object storage services (AWS S3 and Google Cloud Storage) for large-scale data storage. They use distributed file systems like Ceph for intermediate data storage during training.

Software Platform & Developer Tools

Workday provides a robust suite of developer tools and APIs to facilitate AI development and deployment. Highlights include:

• Workday Extend: A platform-as-a-service (PaaS) that allows developers to build and deploy custom applications on top of the Workday platform. It includes AI/ML APIs for accessing pre-trained models and building custom AI-powered features.
• Workday AI Studio: An internal platform that provides a unified interface for data scientists to build, train, and deploy AI models. It integrates with their feature store, training infrastructure, and deployment pipelines.
• Open-Source Contributions: Workday actively contributes to open-source projects related to Federated Learning and data privacy. They maintain a public repository of Federated Learning algorithms optimized for HR data.

Data Pipeline & Storage

Workday handles a massive volume of HR and financial data, requiring a sophisticated data pipeline. Their data infrastructure consists of:

• Data Lake: A central data lake built on Apache Iceberg, storing petabytes of structured and unstructured data.
• Streaming Data Pipeline: Utilizes Apache Kafka and Apache Flink for real-time data ingestion and processing. This powers real-time analytics and triggers automated workflows.
• ETL Pipeline: Employs Apache Airflow for orchestrating batch ETL jobs, transforming and loading data into the data lake and data warehouse.
• Data Governance: Implementing strong data governance policies using Apache Atlas to ensure data quality, security, and compliance. They are experimenting with differential privacy techniques to further enhance data privacy during model training.

Key Products & How They're Built

Workday leverages its AI stack to power key features across its product suite:

• Talent Marketplace: This product uses LLMs to extract skills from employee profiles and job descriptions, matching employees with relevant opportunities within the organization. The matching algorithm is continuously refined using Federated Learning on anonymized data from multiple customers. It is built upon PyTorch models served using Triton Inference Server.
• Skills Cloud: A dynamic, AI-powered skills ontology that automatically identifies and tracks employee skills. It uses natural language processing (NLP) techniques to extract skills from text data (resumes, performance reviews, etc.) and machine learning to infer skills from employee activities. The underlying models are fine-tuned using JAX on TPUs for faster experimentation.

Competitive Moat

Workday's competitive moat is multifaceted:

• Proprietary Data: Their access to a vast dataset of HR and financial data, combined with sophisticated anonymization techniques, provides a unique advantage in training AI models specifically tailored for the enterprise.
• Federated Learning Expertise: Their expertise in Federated Learning allows them to train models on data from multiple customers without compromising privacy. This results in more accurate and generalizable models.
• Integrated Platform: The seamless integration of their AI capabilities within the Workday platform provides a superior user experience compared to point solutions.

Stack Scorecard