Most enterprise AI conversations still start with a cloud assumption: the model lives somewhere else, the data goes to the model, and the work happens through an API. That default has pushed many organizations into tradeoffs they should not have had to make: choosing between capability and control, speed and security, experimentation and governance - while creating a cost structure where AI economics scale linearly with token consumption
Poolside running on Dell gives customers another path: frontier AI that runs inside their own boundary, where enterprises can securely generate tokens closer to their data, with better performance, stronger governance, and more predictable economics. Built on Dell enterprise infrastructure with NVIDIA’s accelerated computing platform, networking, and AI software underneath and Poolside’s secure agent platform on top.
For highly regulated teams, that matters because the requirements are often non-negotiable. Defense, intelligence, financial services, healthcare, and critical infrastructure organizations operate in environments where sensitive code and data cannot leave the customer’s control. Air-gapped networks, STIG-hardened images, approved hardware lists, existing ATO boundaries, and established procurement paths are not edge cases. They are the operating model.
But the case for on-premises deployment is broader than regulatory mandate. Agentic AI can consume orders of magnitude more tokens and is not a neat, predictable API workload. As adoption grows, agents run longer, take more steps, call more tools, and sometimes do significant work before producing a useful result. In a metered model, that token uncertainty turns directly into cost uncertainty.
With on-premises deployment, organizations can plan around known infrastructure costs, improve token economics through higher utilization, and govern availability inside their own operating model rather than depending on shared capacity, upstream rate limits, or a vendor’s release schedule. Instead of treating every increase in usage as a new budget risk, teams can focus on putting agents to work where they create value.
Access and audit follow the same logic. Identity flows through the existing identity provider. Trajectory logs land in existing security tooling, including SIEMs. Retention follows the organization’s existing policy. Nothing about the deployment sits outside the security team’s model of the world, because nothing has been asked to.The common thread is control. For organizations running critical workloads, AI has to run inside the efficient infrastructure, security boundaries, and operating model they already trust. That is not a configuration preference. It is a precondition for using AI at all.
Most AI tooling still assumes the opposite: that teams can send work to a cloud service, adapt their controls around a vendor, or accept limited visibility into how agents operate.
That gap is why we built Poolside the way we did, and why our work with Dell matters.
What the hardest deployments taught us
Nearly two years ago, we were asked whether Poolside could bring frontier AI into environments most vendors never reach: air-gapped, STIG-hardened, classified, and sovereign. In under nine months, Poolside went from prototype to ATO-approved production deployment inside one of the most secure environments in the world. We learned that there is no single template for this kind of deployment. But we also proved there is a path.
That work shaped how we think about every layer of the stack. Ubuntu and RHEL hardened to DISA STIG. RKE2 on classified Kubernetes baselines. Full air-gap operation with no outbound calls. Identity, encrypted secrets, object storage, and a local container registry all bundled inside the boundary. Credentials injected at runtime and automatically redacted from outputs, logs, and agent trajectories. Every agent action is recorded as a searchable trajectory, exportable to the organization’s SIEM. None of this was retrofitted. It was the foundation we started from, because the teams we built for treat security as a precondition, not a feature.
We have kept pushing on the harder problems since. Our inference layer is an in-house variant of vLLM, substantially faster than upstream on the workloads our customers actually run, and the gap continues to widen. The result is more developers served per node within SLA, and a more efficient compute footprint with each release we ship. On the smallest end of that footprint, we are running our Laguna XS model on edge-class hardware like the Dell Pro Max GB10 or NVIDIA DGX Spark, with strong early results for single-developer and small-team use cases that previously had no good, affordable on-prem option.
The arc is consistent: more of the AI stack is moving out of remote cloud services and into infrastructure customers control, from secure data centers to smaller on-premises footprints.
Poolside on Dell
On Dell hardware, Poolside ships as a turnkey system in one of two configurations. The right path depends on what infrastructure the customer already has, and how much of it they want us to bring.
Single-node RKE2, fully self-contained.
Everything required to run Poolside ships in the bundle: an RKE2 Kubernetes cluster on STIG-hardened Ubuntu, PostgreSQL, SeaweedFS for S3-compatible object storage, Keycloak for identity, cert-manager, the NVIDIA GPU operator, a local container registry, the inference stack, and the Poolside Console. The customer provides the hardware. No external dependencies, no outbound network requirement. This is the path most defense, intelligence, and classified customers take because it fits cleanly inside an existing ATO boundary.
Helm on customer Kubernetes, flexible.
For enterprise customers running their own Kubernetes platform, whether upstream, Red Hat OpenShift, or a managed equivalent, the customer provides the peripheral infrastructure: container registry, identity provider, PostgreSQL, certificate management, ingress, storage class, and GPU operator.
Poolside installs on top via Helm. Same platform, same Console, same audit model. What changes is the seam between what we bring and what the customer already has in place.
For organizations serving Poolside to thousands of developers, we support multi-node clustering on Dell hardware as well. Multiple enterprise-rack nodes operate as a single Poolside cluster, with inference and platform services distributed across them while presenting the same Console and audit model to every user. This is the path large enterprises and federal program offices take when a single-node footprint cannot serve the whole user base, but the deployment still has to live inside the customer’s boundary.
Introducing the Poolside Blueprint for the Dell AI Factory
Dell customers now have a more repeatable path from purchase order to running deployment.
The Poolside Blueprint for the Dell AI Factory with NVIDIA is a Dell validated deployment automation that lives in the Dell Automation Platform catalog. It wraps the Poolside install inside Dell’s standard orchestration, lifecycle, and validation workflows. A pre-flight validator confirms the customer’s hardware and peripheral infrastructure match the Platform’s requirements before the install proceeds. The Blueprint then deploys Poolside on top under the same automation patterns Dell customers already use for other Dell AI Factory solutions.
For customers buying through the Dell AI Factory with NVIDIA, this creates a more repeatable path from hardware delivery to a running, governed Poolside environment. Dell brings the enterprise infrastructure and deployment automation. NVIDIA brings accelerated computing, networking, and AI software . Poolside brings the secure agent platform that runs on top. Together, that gives customers a clearer route from purchase order to production deployment, with fewer unknowns before install and less bespoke integration work along the way.
Meeting customers where they are
Efficient deployment is only half the path to production. The other half is procurement.
Public sector and regulated enterprise buyers often adopt new technology through established contract vehicles, OEM agreements, and reseller relationships their procurement and security teams already understand. For customers already buying Dell infrastructure at scale, Poolside on Dell AI Factory with NVIDIA gives them a more familiar route to adoption: one that can align with existing purchasing motions instead of forcing every deployment through a completely new vendor path. That matters for customer teams, because the barrier to adopting secure AI is rarely only technical. It is also operational, commercial, and procedural. The easier the path from approved infrastructure to governed deployment, the faster customers can move from evaluation to real usage.
One path among several
Dell hardware is one of four deployment paths for the Poolside Platform.Customers can also deploy Poolside on their own bare metal, inside their existing VPC on AWS, Azure, or Google Cloud, or via Helm on customer-provided NVIDIA AI infrastructure.The same foundation models, the same Console, the same agent orchestration, and the same audit trails run across all of them.
That matters because the journey rarely starts and ends in one place. An organization might begin with a tower in a secure facility, scale to an enterprise rack as the user base grows, and connect a regional VPC for adjacent lower-sensitivity work, all without changing the platform underneath.The Poolside Platform was designed for exactly that from the start.
What this enables
For customers whose data, workflows, and procurement paths require AI to run inside their boundary, Poolside on Dell AI Factory with NVIDIA respects every part of that requirement.The hardware is qualified. The deployment path is validated. The platform is built for environments where cloud-first AI tools are often difficult or impossible to adopt. And the customer keeps the model weights, the audit trail, the operational control, and more efficient token economics from day one.
For these organizations, that is what frontier AI on-premises has to look like.
→ Learn more about the Poolside Platform
→ Talk to us about deploying Poolside in your environment
And if you’re thinking through how this could work inside your own infrastructure, email me directly at colin@poolside.ai