Scaling AI in the gaming sector demands a reimagining of infrastructure that removes friction, accelerates deployment, and puts powerful tools into the hands of diverse teams. At Electronic Arts, a new approach is unlocking faster iteration, lower costs, and higher reliability across AI-driven development.
Electronic Arts faces a distinctly modern challenge. As AI increasingly shapes player experiences, the company must support teams with vastly different levels of AI expertise, from novices seeking straightforward support to research groups pushing the frontier of machine learning. Scaling AI at this level, across a globally distributed organisation, is not simply a case of adding more models or buying more GPUs. It demands the orchestration of workflows, infrastructure, and compliance without suffocating the creativity that drives innovation.
Fragmentation was quickly identified as a core problem. Each team had evolved its own processes for model training, storage, and deployment, creating bottlenecks, duplicated effort, and spiralling maintenance overheads. “Scaling AI is not just about provisioning more hardware or encouraging more projects,” Wah Loon Keng, Senior AI Engineer at Electronic Arts, explains. “It is about making AI useful, efficient, and accessible, removing barriers that prevent teams from doing what they do best.”
Without intervention, the natural result of fragmentation would be slower project timelines, inconsistent standards, increased security risks, and a rising operational burden that few researchers or developers were equipped to manage. Electronic Arts recognised that a coherent and scalable approach would not just be a technical enabler, but a strategic asset critical to maintaining leadership in an increasingly AI-driven industry.
Streamlining compute with Dstack
The priority was streamlining compute provisioning, the critical foundation for development and training. Traditional enterprise processes, bogged down by security protocols and procurement red tape, often took weeks to allocate a single GPU box. This friction directly stifled experimentation and iteration.
To address this, EA deployed Dstack, an open-source tool designed to automate the allocation of computing resources across multiple environments, whether on AWS, Google Cloud Platform, or the company’s own on-premises infrastructure. Dstack’s philosophy is simplicity: users define their requirements in a configuration file written in YAML (Yet Another Markup Language), apply the file, and within minutes receive a fully provisioned virtual machine equipped with the necessary GPU power. “We wanted to replicate the graduate student experience where you SSH (Secure Shell) into a machine and start working immediately,” adds Xin Gao, Lead Product Manager at Electronic Arts.
What makes Dstack transformative is not just speed but seamlessness. The system automatically clones local code repositories, installs dependencies, and provisions the environment with baked-in security and compliance, allowing researchers to focus entirely on development. Distributed training, often an operational nightmare involving multi-node orchestration and environment variable configuration, becomes trivial. “Dstack hides the complexity of multi-node setups,” Keng explains. “Users simply specify the number of nodes and GPUs required, and the environment variables and inter-node communication are handled behind the scenes.”
By cutting provisioning times from weeks to minutes, Dstack enables faster iteration, greater flexibility in hardware choice, and significant cost savings. Instead of locking expensive resources into permanent reservations, researchers spin up, use, and shut down resources as needed, achieving utilisation efficiencies that were previously impossible.
The familiar environment also means that onboarding new team members becomes significantly easier. Researchers and engineers can step into EA’s AI development ecosystem with minimal learning curve, accelerating productivity and promoting a culture of experimentation.
Consolidating storage with an internal Artifactory
While compute provisioning is critical, it is only part of the pipeline. Once models are trained, they need to be stored, versioned, and shared across teams in a consistent and reliable way. Previously, EA’s storage landscape was chaotic, with teams using a mix of cloud buckets, Git repositories, and ad-hoc local storage solutions.
The solution was the deployment of a centralised, internal JFrog Artifactory instance. More than a simple file store, this system acts as a universal repository supporting Docker images, Kubernetes Helm charts, and, crucially, Hugging Face-compatible ML models. “We wanted to provide a familiar experience for researchers used to Hugging Face, but within a secure, controlled environment,” Gao explains.
Authentication is handled transparently by the platform, allowing users to push and pull artefacts without worrying about access credentials. The Artifactory integrates natively into CI/CD pipelines, enabling teams to quickly deploy models into production environments or share them across projects without duplicating effort.
The consolidation of artefact management not only eliminates fragmentation but drastically reduces maintenance overheads. Researchers no longer waste time setting up and securing their own storage solutions, and compliance teams benefit from a single point of control for auditability and governance.
The standardisation brought by the Artifactory also plays a critical role in reproducibility. With centralised artefacts, teams can consistently reproduce experiments, compare model versions, and collaborate across departments with confidence that they are working with the correct assets.
Accelerating production with Kubernetes and Triton
With training and storage addressed, the final hurdle lay in production deployment. EA recognised that research scientists and ML engineers should not be burdened with managing Kubernetes clusters, service meshes, or inference server scaling. These operational complexities could not be allowed to slow down innovation.
The internal solution, codenamed AXS, is a production-grade Kubernetes cluster enhanced with ML-specific capabilities. At its heart lies KServe, orchestrating NVIDIA Triton inference servers that expose model APIs automatically. “If a researcher can export an ONNX file, they can deploy a production-ready model in minutes,” Keng says.
For teams simply looking to serve a model, KServe abstracts away almost every operational concern. A YAML file specifying the model location is all that is needed. The platform pulls the model from Artifactory, spins up the necessary Kubernetes pods, and exposes a high-availability HTTP and gRPC API endpoint, all secured and namespace-isolated to prevent cross-team conflicts.
Power users retain the flexibility to deploy full custom applications on Kubernetes if needed, but for the majority, the experience is reduced to a few simple commands integrated into familiar CI/CD workflows. This not only accelerates time to production from months to days but also improves cost efficiency.
Running models on NVIDIA Triton ensures that inference workloads are executed in the most hardware-optimised way possible, while Kubernetes’ autoscaling capabilities match resource allocation to real-time demand. “By centralising production ops, we ensure reliability, scalability, and cost-efficiency across every team’s deployment,” continues Gao. “It frees researchers to focus on model quality rather than infrastructure babysitting.”
The benefit of AXS extends further than efficiency alone. With centralised governance and robust service level agreements, the system enhances security, guarantees uptime, and mitigates the risk of errors or failures that could impact live games and player satisfaction.
The business impact of scaling AI intelligently
Beyond the technical achievements, EA’s approach demonstrates how intelligent infrastructure design amplifies organisational impact. Dstack accelerates experimentation by reducing time-to-GPU from weeks to minutes. The Artifactory eliminates storage fragmentation, ensuring models are versioned, discoverable, and reusable. AXS enables production deployment with minimal friction, ensuring that successful experiments transition to player-facing experiences quickly and reliably.
Cost optimisation is embedded at every stage. On-demand provisioning prevents idle resources from draining budgets. Centralised Kubernetes operations avoid duplication of DevOps effort. The use of standardised inference servers maximises hardware efficiency, ensuring that every dollar spent on cloud compute is delivering player value.
Most importantly, reliability and scalability are no longer afterthoughts. By offering teams SLAs derived directly from Kubernetes reliability standards, EA ensures that AI-driven features can be trusted in live production environments, protecting both player experience and business reputation.
Scaling AI at EA is not a theoretical exercise. It is a tangible operational advantage, allowing the company to move faster, spend smarter, and deliver richer, more intelligent gaming experiences. The infrastructure itself has become a platform for creativity, experimentation, and ultimately, competitive differentiation.
Unlocking the future of AI-driven gaming
Electronic Arts’ infrastructure journey is a case study in the realities of scaling AI across a complex enterprise. Fragmentation, operational bottlenecks, and resource inefficiencies are not unique to gaming companies, and the solutions EA has developed are highly transferable to other sectors facing similar challenges.
In a landscape where AI is moving from experimental curiosity to operational necessity, organisations must think beyond isolated projects. They must build scalable foundations that empower every team to innovate without friction, deploy without delay, and operate without fear. “Ultimately, it is about democratising AI,” Keng concludes. “Not every team needs to become infrastructure experts. They just need the tools to move fast, stay compliant, and deliver impact.”
For any enterprise seeking to scale AI sustainably, EA’s experience offers a clear lesson. Success lies not in throwing more resources at the problem, but in designing systems that allow people to focus on what they do best: building, experimenting, and innovating without limits.
