Codex On-Premises: The Dell AI Factory Partnership and What It Means for Regulated Enterprise Teams

Updated

codexenterpriseon-premisesdellhybrid-deploymentregulated-industriesrequirements-tomlmanaged-configuration

Codex On-Premises: The Dell AI Factory Partnership and What It Means for Regulated Enterprise Teams

On 18 May 2026, OpenAI and Dell Technologies announced a partnership at Dell Technologies World to bring Codex into hybrid and on-premises enterprise environments 1. For the 67 per cent of enterprise AI workloads that already operate outside the public cloud 2, and for sectors where sending proprietary code to any hyperscaler is a non-starter — defence, central banking, healthcare — this is the first time OpenAI has offered a viable on-premises path for Codex specifically.

This article examines the architecture, the practical deployment stack, what Codex CLI configuration looks like in these environments, and the gaps that remain.

Why On-Premises Matters Now

Codex surpassed four million weekly active developers in May 2026 1. Yet adoption in regulated industries has been constrained by a fundamental tension: the agent needs access to internal codebases, documentation, and business systems, but those assets cannot leave the physical perimeter.

Previous workarounds — Azure OpenAI Service with private endpoints, or local model providers via Ollama — each had trade-offs. Azure still routes through a third-party datacentre. Local models lack the reasoning capability of GPT-5.5 3. The Dell partnership attempts to resolve this by keeping the full Codex stack within customer-owned infrastructure.

Architecture: Dell AI Factory + Codex

The deployment sits on Dell AI Factory, the branded hardware-and-software stack that has shipped to over 5,000 enterprise customers since 2024 2. The integration connects Codex to the Dell AI Data Platform for data preparation, governance, and federated SQL via Starburst 4.

graph TB
    subgraph "Customer Datacentre"
        subgraph "Dell AI Factory"
            PE["PowerEdge XE9680 / XE9812<br/>NVIDIA H200 / Vera Rubin NVL72"]
            ADP["Dell AI Data Platform<br/>PowerScale + Starburst"]
            OS["NVIDIA OpenShell<br/>Consistency Layer"]
        end
        subgraph "Codex Stack"
            CE["Codex Enterprise Runtime"]
            CLI["Codex CLI Instances"]
            APP["Codex App / IDE Extension"]
            CAdmin["ChatGPT Enterprise Admin Console"]
        end
        subgraph "Enterprise Systems"
            GH["GitHub Enterprise Server"]
            DOCS["Internal Documentation"]
            CB["Internal Codebases"]
        end
    end

    PE --> CE
    ADP --> CE
    OS --> CE
    CE --> CLI
    CE --> APP
    CAdmin --> CE
    CE --> GH
    CE --> DOCS
    CE --> CB

Hardware Options

Three PowerEdge configurations anchor the compute layer:

Server Form Factor GPUs Key Metric
PowerEdge XE9680 8U air-cooled 8× NVIDIA H200 141 GB Baseline enterprise inference
PowerEdge XE9680L 4U liquid-cooled 8× NVIDIA H200 141 GB 2.5× energy efficiency 4
PowerEdge XE9812 Announced DTW 2026 NVIDIA Vera Rubin NVL72 10× lower cost-per-token vs Blackwell 5

AMD MI300X and Intel Gaudi3 OAM are listed as alternative GPU options 4, though OpenAI model compatibility with non-NVIDIA silicon is unconfirmed at time of writing. ⚠️

NVIDIA OpenShell

OpenShell provides a consistency layer that lets teams build and test agents locally, then promote them to datacentre-scale infrastructure 2. This is significant for Codex workflows: a developer can prototype a skill or automation on a workstation with a smaller model, then deploy the same configuration against on-premises GPT-5.5 without rewriting the harness.

Codex CLI Configuration for On-Premises

Whether your organisation deploys on Dell infrastructure or any other self-hosted environment, Codex CLI’s enterprise configuration stack provides the policy enforcement layer. Configuration follows a strict precedence hierarchy 6:

  1. Cloud-managed requirements — pushed from the ChatGPT Enterprise admin console
  2. macOS MDM — via com.openai.codex:requirements_toml_base64
  3. System requirements.toml — /etc/codex/requirements.toml (Unix) or %ProgramData%\OpenAI\Codex\requirements.toml (Windows)

Earlier layers override later ones, and users cannot weaken admin-set constraints 6.

Sample On-Premises requirements.toml

# /etc/codex/requirements.toml — deployed via MDM or system image

[approval]
# Require explicit approval for all tool calls
policy = "on-request"

[sandbox]
# Restrict to workspace-write; no full access on regulated machines
mode = "workspace-write"

[web_search]
# Disable web search entirely on air-gapped nodes
mode = "disabled"

[network]
# Lock network to internal endpoints only
access = true
allowed_domains = [
  "github.internal.corp.example",
  "artifactory.internal.corp.example",
  "codex-runtime.internal.corp.example",
]

[features]
# Disable browser and computer use in regulated environments
browser_use = false
computer_use = false

[filesystem]
# Prevent the agent from reading secrets directories
deny_read = [
  "/etc/ssl/private",
  "/var/secrets",
  "~/.ssh",
]

[mcp]
# Allowlist only approved MCP servers
allowed_servers = [
  { command = "ghes-mcp", args = ["--host", "github.internal.corp.example"] },
]

Managed Defaults vs Requirements

A subtlety that catches teams out: requirements.toml locks values that users cannot change. For settings you want as defaults but allow developers to override during a session, use managed_config.toml instead. Codex reapplies managed defaults on restart 6.

# /etc/codex/managed_config.toml — soft defaults, user can override

[model]
default = "gpt-5.5"

[sandbox]
mode = "workspace-write"

Enterprise Admin Workflow

The ChatGPT Enterprise admin console provides the control plane 7:

sequenceDiagram
    participant Admin as Workspace Admin
    participant Console as ChatGPT Enterprise Console
    participant Cloud as Cloud-Managed Requirements
    participant CLI as Codex CLI Instance

    Admin->>Console: Configure workspace toggles
    Admin->>Console: Create RBAC roles (Codex Admin, Developer)
    Admin->>Console: Deploy cloud-managed requirements.toml
    Console->>Cloud: Push policies
    CLI->>Cloud: Fetch requirements on startup
    Cloud-->>CLI: Return merged policy
    CLI->>CLI: Enforce sandbox, network, approval rules
    Admin->>Console: Query Compliance API for audit logs

Key admin steps for an on-premises rollout 7:

  1. Enable surfaces — toggle “Allow members to use Codex Local” in workspace settings
  2. Set up RBAC — create a dedicated “Codex Admin” role with limited membership; assign developer groups separately
  3. Deploy managed policies — push requirements.toml controlling sandbox behaviour, web search, approval workflows, and feature flags
  4. Connect internal repositories — link GitHub Enterprise Server, configure IP allowlists
  5. Enable audit logging — create API keys for the Compliance API; query CODEX_LOG and CODEX_SECURITY_LOG event types

Compliance API Integration

For regulated environments, audit visibility is non-negotiable. The Compliance API exposes endpoints for usage tracking and code review monitoring 7:

# Query Codex usage for the workspace
curl -s -H "Authorization: Bearer $ADMIN_API_KEY" \
  "https://api.openai.com/v1/workspaces/${WORKSPACE_ID}/usage?surface=codex" \
  | jq '.data[] | {user: .user_email, tokens: .total_tokens, date: .date}'

# Retrieve code review audit logs
curl -s -H "Authorization: Bearer $ADMIN_API_KEY" \
  "https://api.openai.com/v1/workspaces/${WORKSPACE_ID}/code_reviews?status=completed" \
  | jq '.reviews[] | {repo: .repository, reviewer: .codex_model, findings: .finding_count}'

What This Changes for Development Teams

Before: The Workaround Stack

Teams in regulated industries previously cobbled together partial solutions:

  • Azure OpenAI Service with private endpoints — works, but data still traverses Microsoft infrastructure
  • Ollama + local models — fully air-gapped, but reasoning quality drops sharply below GPT-5.5 3
  • Codex CLI with network_access = false — prevents data exfiltration but cripples web search and MCP server connectivity

After: Full Codex Within the Perimeter

The Dell partnership promises the full Codex experience — including Goal Mode, MCP servers, code review automation, and Codex Security scanning — running entirely within the customer’s physical infrastructure 1. Task durations range from 1 to 30 minutes depending on complexity 2.

This unlocks use cases that were previously impossible in these environments:

  • Automated code review across internal repositories with access to internal style guides and architecture documentation
  • Test coverage analysis that can reason about proprietary frameworks and internal libraries
  • Incident response workflows with access to internal runbooks and production telemetry

Critical Gaps and Open Questions

The partnership announcement left significant implementation details unspecified. Teams evaluating this path should be aware of what remains unclear:

General availability date — not disclosed. Dell AI Data Platform upgrades are scheduled for Q2 2026, but a firm ship date for the Codex integration has not been published. ⚠️

Compliance certifications — FedRAMP, HIPAA, IL5, and SOC 2 status for the on-premises Codex deployment are pending 4. Regulated buyers will need these before procurement can proceed. ⚠️

Pricing — on-premises pricing has not been disclosed. Cloud Codex runs approximately $100–$200 per developer per month 8. On-premises will likely combine hardware capex with per-seat or per-token Codex licensing. ⚠️

Air-gapped support — the announcement describes “hybrid and on-premises environments” but does not address fully disconnected scenarios 4. Current documentation suggests external connectivity remains necessary for baseline model operations. ⚠️

Infrastructure prerequisite — this is not a “deploy Codex on your existing servers” announcement. Organisations must already operate Dell AI Factory systems or be prepared to procure them 2.

Competitive Context

The Dell partnership positions OpenAI against an increasingly crowded on-premises coding agent market:

Vendor On-Premises Option Model Data Residency
OpenAI + Dell Dell AI Factory (Q2 2026) GPT-5.5 Customer datacentre
GitHub Copilot GHES + Azure Private GPT-5.5 / Claude Azure tenant
Amazon CodeWhisperer AWS Outposts Amazon Nova Customer rack (AWS hardware)
Tabnine Self-hosted (Docker) Tabnine models Customer infrastructure
Sourcegraph Cody Self-hosted Multiple (BYO key) Customer infrastructure

The differentiator for the Dell path is access to the full Codex runtime — Goal Mode, the app-server protocol, Codex Security — rather than just a code completion endpoint. Whether that advantage holds depends on the unresolved gaps above.

Practical Recommendations

For teams evaluating on-premises Codex today:

  1. Start with the configuration stack now — requirements.toml and managed configuration work identically whether you’re on cloud Codex or preparing for on-premises. Deploy your policies, test enforcement, and iterate before the hardware arrives.

  2. Audit your MCP server dependencies — catalogue which MCP servers your teams use and whether they can run within the perimeter. Streamable HTTP servers that point to external SaaS endpoints will need on-premises alternatives or allowlisting.

  3. Map your compliance requirements — document exactly which certifications you need. When OpenAI publishes compliance attestations for the Dell deployment, you will need to match them against your requirements immediately.

  4. Budget for infrastructure — the PowerEdge XE9680 with H200 GPUs is not a trivial procurement. Work with Dell to size the cluster for your developer count and expected concurrent Codex sessions.

  5. Watch the changelog — Codex CLI v0.136.0 already introduced Windows sandbox provisioning and remote execution setup with CODEX_API_KEY registration for approved hosts 9. These enterprise primitives are shipping faster than the marketing announcements.

Citations

  1. OpenAI. “OpenAI and Dell Technologies partner to bring Codex to hybrid and on-premises enterprise environments.” 18 May 2026. https://openai.com/index/dell-codex-enterprise-partnership/ ↩ ↩2 ↩3

  2. ByteIota. “OpenAI Codex On-Premises: What Enterprise Devs Must Know.” May 2026. https://byteiota.com/openai-codex-dell-on-premises-enterprise/ ↩ ↩2 ↩3 ↩4 ↩5

  3. OpenAI. “Introducing GPT-5.5.” 2026. https://openai.com/index/introducing-gpt-5-5/ ↩ ↩2

  4. Digital Applied. “OpenAI + Dell Codex: On-Premises Enterprise Agents.” May 2026. https://www.digitalapplied.com/blog/openai-dell-codex-on-premises-enterprise ↩ ↩2 ↩3 ↩4 ↩5

  5. Dell Technologies. “Dell Technologies Delivers Production-Ready Agentic AI from Deskside to Data Center.” 18 May 2026. https://www.dell.com/en-us/dt/corporate/newsroom/announcements/detailpage.press-releases~usa~2026~05~dell-technologies-delivers-production-ready-agentic-ai-from-deskside-to-data-center.htm ↩

  6. OpenAI Developers. “Managed configuration — Codex.” https://developers.openai.com/codex/enterprise/managed-configuration ↩ ↩2 ↩3

  7. OpenAI Developers. “Admin Setup — Codex.” https://developers.openai.com/codex/enterprise/ ↩ ↩2 ↩3

  8. eesel AI. “A clear guide to OpenAI Codex pricing in 2026.” https://www.eesel.ai/blog/codex-pricing ↩

  9. OpenAI Developers. “Changelog — Codex.” https://developers.openai.com/codex/changelog ↩