Where should enterprises host their AI agents?

AI agents can be hosted in many places, from hyperscaler infrastructure and application platforms to specialized AI compute and enterprise-owned environments. Choosing where an agent runs, however, is about more than finding the right compute or model access.

The right platform depends on what the agent actually needs to run reliably in production. An agent that calls external models and a few internal APIs has very different requirements from one that executes code, runs local models, or handles sensitive enterprise data. This guide compares the leading platforms for hosting enterprise AI agents and what to consider when choosing between them.

An AI agent is usually an application made up of several components rather than a single process. The agent runtime might run as a container, serverless workload, Kubernetes service, virtual machine, or isolated sandbox, while databases, queues, object storage, model APIs, and other services run alongside it.

The execution environment becomes particularly important when an agent can execute code, access files, call internal APIs, or use credentials to take actions. An agent that only calls an external model API may run comfortably as a standard application, while an agent executing untrusted code needs a stronger isolation boundary.

Enterprises should therefore evaluate both where the agent runs and what infrastructure it needs around it. A production agent may need databases, background workers, scheduled jobs, GPUs, private networking, secrets management, and isolated execution in addition to its core runtime.

The right hosting platform depends on the agent's requirements and the organization's existing infrastructure. Evaluate these factors before choosing a platform:

• Data and compliance: What data can the agent access, and where must that data remain? Agents handling sensitive or regulated data may require specific regions, private infrastructure, or customer-controlled cloud environments.
• Runtime and isolation: Does the agent only call APIs, or does it execute generated code, shell commands, access files, or interact with untrusted packages? Higher-risk execution requires stronger isolation.
• Compute and GPUs: Agents using hosted models may primarily need CPU infrastructure, while agents running local models or inference workloads may require GPUs.
• Application infrastructure: Does the agent need databases, queues, object storage, background workers, scheduled jobs, or persistent volumes?
• Networking: Does the agent need access to private databases, internal APIs, or external services? Look for private networking and controls over ingress and egress.
• Enterprise security: Look for SSO, RBAC, secrets management, scoped access, audit logs, and controls that separate developers, automation, and production workloads.
• Deployment and operations: Production agents need repeatable deployments, health checks, observability, scaling, rollback, and backup capabilities.
• Infrastructure flexibility: Enterprises may need managed infrastructure, BYOC, multiple cloud providers, or self-hosted environments depending on their requirements.
• Operational capacity: A managed platform reduces the infrastructure your team needs to operate, while Kubernetes and bare-metal deployments provide more control at the cost of additional operational work.

The right platform is the one that meets the agent's actual requirements without forcing the enterprise to operate infrastructure it does not need.

There is no single platform that is best for every enterprise AI agent. The platforms below take different approaches, from full-stack application platforms to hyperscaler infrastructure and specialized AI compute.

Northflank is a full-stack application platform for running AI agents and the infrastructure around them. It combines secure agent sandboxes, GPU workloads, managed databases, CI/CD, preview environments, networking, and enterprise governance in a single control plane. Teams can run workloads on Northflank Managed Cloud or deploy into their own infrastructure through self-serve BYOC.

Key features:

• Sandbox isolation: Securely execute untrusted or AI-generated code in isolated environments using Kata Containers with Cloud Hypervisor, Firecracker, or gVisor, depending on the workload and infrastructure.
• GPU workloads: Run AI training, inference, batch jobs, and agent workloads on H100, H200, A100, L4, L40S, and B200 GPUs alongside CPU services, sandboxes, and managed databases.
• Managed databases: Provision PostgreSQL, MySQL, MongoDB, Redis, MinIO, and RabbitMQ with automated backups, point-in-time recovery, and database instances for preview environments.
• CI/CD and preview environments: Deploy from Git repositories with automated builds, deployment workflows, and isolated preview environments for testing agent applications and their dependencies.
• Enterprise controls: Secure workloads with RBAC, SAML and OIDC SSO, secrets management, and audit logs that integrate with SIEM platforms.
• BYOC and forward-deployed deployments: Run workloads in customer-controlled cloud, private, or on-premises infrastructure, with forward-deployed options for organizations that require the platform to run entirely within their own environment.
• Northflank Skills: Deploy, manage, troubleshoot, and scale workloads directly from AI coding agents such as Claude Code, Codex, Gemini CLI, and Cursor.

Northflank also supports the wider application stack around an agent, including networking, secrets, observability, deployment workflows, and supporting infrastructure. This means teams can run the agent and the services it depends on through the same platform instead of assembling separate infrastructure for each part of the workload.

Best for: Enterprise AI agent hosting, regulated workloads, agents that execute code, and teams that need GPUs, secure sandboxes, managed infrastructure, and flexible deployment options in one platform.

Amazon Web Services provides several ways to run AI agents, including managed AI services, containers, serverless compute, Kubernetes, and GPU infrastructure. Teams can build agents with services such as Amazon Bedrock and run supporting workloads using ECS, EKS, databases, queues, and object storage.

The main advantage is breadth and integration. Enterprises already invested in AWS can build agents directly into their existing identity, networking, security, observability, and infrastructure architecture. The tradeoff is that assembling a complete agent platform from individual AWS services can require significant expertise across multiple infrastructure layers.

Best for: Enterprises already standardized on AWS that want to build and operate AI agents directly within their existing cloud architecture.

Google Cloud combines managed AI and agent capabilities with its broader cloud infrastructure, including compute, GPUs, databases, networking, IAM, and data services. This gives enterprises multiple options for building and running agents while keeping them close to their existing cloud and data infrastructure.

Google Cloud is particularly suited to organizations already using Google Cloud for data, analytics, and machine learning workloads. Its broader AI platform also provides managed capabilities for teams that do not want to assemble the entire agent stack themselves.

Best for: Google Cloud enterprises that want AI agents closely integrated with their existing AI, data, and cloud infrastructure.

Microsoft Azure provides managed AI and agent capabilities alongside containers, Kubernetes, GPUs, databases, networking, and Microsoft identity infrastructure. Enterprises can combine these services to build agents that integrate with their existing Azure environment.

Azure is particularly attractive to organizations already invested in Microsoft 365, Entra ID, Azure infrastructure, and the wider Microsoft enterprise ecosystem. Its broad compliance portfolio can also make it suitable for organizations with strict regulatory and governance requirements.

Best for: Microsoft-centric enterprises that want AI agents integrated with their existing Azure and identity infrastructure.

Modal is a serverless compute platform focused on AI and machine learning workloads. It provides on-demand CPU and GPU compute with a Python-first development model, allowing teams to run workloads without managing traditional servers or Kubernetes infrastructure.

For AI agents that need GPU acceleration, model inference, or bursty compute, Modal provides a specialized infrastructure layer with a relatively simple developer experience.

Best for: AI teams building Python-based agents that need scalable serverless CPU and GPU infrastructure.

Enterprises can run AI agents directly on Kubernetes, private Kubernetes distributions, or bare-metal infrastructure. This provides extensive control over the execution environment, networking, data location, security policies, and underlying infrastructure.

The tradeoff is operational complexity. Kubernetes provides orchestration, but enterprises still need to provide or integrate CI/CD, secrets management, observability, databases, networking, access controls, sandboxing, and other platform capabilities required to operate agents reliably.

Best for: Enterprises with existing platform engineering teams, strict infrastructure requirements, or workloads that cannot run on managed platforms.

The best platform depends on which requirements matter most for your workloads.

For enterprises running a small number of conventional agents, a hyperscaler or application platform may be enough. For specialized GPU workloads, a platform such as Modal can make more sense. Organizations with significant existing platform engineering capacity may prefer Kubernetes when maximum infrastructure control is more important than operational simplicity.

Northflank is the strongest overall option when an enterprise needs more than an agent runtime. It combines application infrastructure, databases, GPUs, secure sandbox execution, networking, CI/CD, and enterprise governance in one self-serve platform. It can also run on Northflank's managed cloud or in customer-controlled infrastructure, giving enterprises flexibility as requirements change.

This is particularly useful for enterprises running different types of agents, where some workloads need standard application infrastructure while others require GPUs, databases, private networking, or isolated code execution.

Yes. Platforms that support BYOC can run agent workloads inside an enterprise's AWS, Azure, GCP, or other supported cloud environment. This can provide greater control over networking, data residency, and infrastructure while retaining a managed application platform.

Not necessarily. Agents that primarily call external model APIs can often run on standard application infrastructure. Dedicated GPUs or specialized execution environments become more important when agents run local models, execute code, or require stronger isolation.

Not every agent needs a dedicated sandbox. Agents that execute generated code, user-submitted code, shell commands, or untrusted packages need an appropriate isolation boundary. Conventional agents that only call APIs and access trusted services can often run using standard application isolation.

Managed platforms generally reduce operational overhead and provide a faster path to production, while self-hosted infrastructure provides more control. The right choice depends on the organization's security requirements, existing infrastructure, platform engineering capacity, and number of agent workloads.

Yes. Enterprises may use different platforms for different workloads. The important consideration is maintaining consistent identity, security, networking, deployment, and audit controls across the environments where agents run.

There is no single best place to host an enterprise AI agent. The right platform depends on the agent's runtime behavior, data requirements, compute needs, existing cloud infrastructure, security requirements, and the operational capacity of the team running it.

For enterprises running a growing portfolio of agents, the platform matters as much as the underlying infrastructure. A platform that combines application hosting, secure execution, GPUs, databases, networking, deployment workflows, and enterprise governance gives teams a consistent way to move agents from development into production.

• Best platforms to host AI agents
• What is BYOC in cloud computing?
• How to secure AI agents in the cloud
• How to manage enterprise AI infrastructure