OpenAI announced a partnership with Dell Technologies on May 18 that brings Codex on-prem — Codex’s first major hybrid and on-premises deployment path. The integration connects Codex with the Dell AI Data Platform and Dell AI Factory, the infrastructure stack already running at 5,000 Dell enterprise customers including major banks, hospitals, and government agencies. For OpenAI, this is the first explicit distribution channel for buyers who can’t send data to public cloud. For Dell, it’s a signal that the AI Factory pivot is producing real OpenAI-grade integrations. Codex on-prem changes the calculus for regulated industries that wanted Codex but were blocked by data-residency or air-gap requirements.
What’s actually new about Codex on-prem
Three concrete shifts from the existing Codex offering. First, Codex agents can now run against codebases, internal documentation, business systems, and team workflows that live entirely inside the customer’s perimeter. Until this announcement, Codex was a cloud-only product — every interaction sent code and context to OpenAI’s infrastructure. The on-prem option keeps the model weights and serving infrastructure inside the customer’s environment via Dell’s AI Factory hardware, with control plane connectivity to OpenAI for model updates and licensing.
Second, the integration is native to Dell’s existing data platform. Customers already running the Dell AI Data Platform — a unified system for storing, organizing, and governing enterprise data on Dell servers — can connect Codex to that platform directly. No new data pipelines, no separate vector store buildout, no migration of corporate knowledge to a third-party cloud. The model reads enterprise data where it lives, with the customer’s existing access controls and audit logs applying to Codex’s reads.
Third, the deployment posture is genuinely hybrid. Inference can run on-prem (for regulated workloads), in OpenAI’s cloud (for non-sensitive workloads), or split between them with workload-level routing. Customers don’t have to choose all-cloud or all-on-prem; they configure routing policies that send the sensitive workloads to on-prem and the rest to the public cloud, with consistent management across both.
Why it matters for Codex on-prem and enterprise AI
- This unlocks the regulated-industry buyer. UK and EU banks, healthcare systems, government agencies, and defense contractors typically have data-residency requirements that prevent cloud-only AI use. Codex on-prem is the first frontier-lab agent product that fits within those constraints.
- 4 million developers already use Codex weekly. OpenAI’s enterprise growth has been hampered by regulated buyers who couldn’t deploy; this partnership unblocks a large pipeline of those deals.
- Dell’s 5,000 AI Factory customers are an immediate distribution channel. Most are mid-to-large enterprises with existing Dell hardware footprints — the integration is incremental, not greenfield.
- The hybrid model becomes the default architecture. Going forward, expect enterprise AI deployments to default to hybrid: sensitive workloads on-prem, routine workloads in the cloud, with workload-level routing. This deal sets a template that other providers will copy.
- It’s a competitive response to Claude on AWS and Microsoft Foundry. Anthropic’s Claude is now on AWS with native IAM and billing; Microsoft Foundry includes Claude alongside OpenAI. The Codex-Dell deal gives OpenAI the equivalent enterprise-grade on-prem story.
- The Codex platform is expanding beyond code. OpenAI mentioned Codex-powered agents deploying for non-code workflows — report preparation, lead qualification, feedback routing — signaling that “Codex” is becoming the brand for OpenAI’s general agent platform, not just the coding-specific tool.
How to use Codex on-prem today
Codex on-prem is in early customer rollout as of May 18, 2026, with broader Dell AI Factory customer availability through Q2 2026. Here’s the activation path for organizations already on Dell hardware.
- Confirm your Dell footprint includes the AI Factory stack or compatible servers. The minimum supported configuration is Dell PowerEdge XE9680 or XE9685L with H200/B200 GPUs, plus the Dell AI Data Platform installed and operational.
- Work with your Dell account team to engage the joint OpenAI-Dell deployment service. The partnership ships as a managed onboarding (architecture review, capacity planning, model licensing) rather than self-service.
- Configure the OpenAI control plane endpoint. The control plane talks to OpenAI for model licensing, version management, and (optionally) telemetry — but inference traffic stays inside your perimeter.
# Sample on-prem Codex config (simplified)
# /etc/openai-codex/config.yaml
deployment:
mode: on-premises
control_plane: https://control.openai.com/v1
inference_endpoint: https://codex-internal.corp.example.com:8443
data_sources:
- type: dell-ai-data-platform
endpoint: https://aidata.corp.example.com
auth: oauth2
scopes: [code.read, docs.read]
- type: github-enterprise
endpoint: https://github.corp.example.com
auth: app-installation
policies:
routing:
sensitive_workloads: on-prem-only
routine_workloads: cloud-with-fallback
data_residency:
region: eu-west
log_retention_days: 365- Connect Codex to your code repositories and documentation systems. Codex supports GitHub Enterprise, GitLab Self-Managed, Bitbucket Data Center, and the Dell AI Data Platform for unified document access. Each connector uses your existing identity provider (Okta, Azure AD, Ping) for authentication.
- Pilot with a single team. The recommended rollout pattern is: pilot with a 10-20 engineer team for 4-6 weeks; gather usage and quality data; expand to a department; expand to the engineering organization. Skip the pilot phase only if you’re confident in the integration; most Dell-OpenAI customer rollouts have caught configuration issues in the pilot that would have been painful at full scale.
# Codex CLI usage in an on-prem deployment (same commands as cloud)
codex login --endpoint https://codex-internal.corp.example.com:8443
# Interactive coding session
codex chat "Refactor the auth module to use scoped tokens"
# Run a one-off task with access to specific repo paths
codex run \
--repo /workspaces/my-app \
--task "Add input validation to /api/users/* endpoints" \
--reviewer team-security
# Open a PR with the changes
codex submit --as-pr --base main --reviewers @lead-eng,@security- Configure audit logging to your SIEM. Codex on-prem produces structured audit logs for every model call, tool invocation, and data access. Ship these to Splunk, Datadog, Elastic, or your existing log aggregator for compliance evidence and incident response.
- For organizations with multi-region requirements (EU + US, for example), Dell and OpenAI support multi-region deployments with regional model serving and workload routing based on data origin. Configure via the deployment management console; ask your account team for the latest multi-region setup guide.
How it compares to Claude on AWS and other enterprise agent paths
| Capability | OpenAI Codex on-prem (Dell) | Claude on AWS Bedrock | Microsoft Copilot + Foundry | Local LLMs (Llama, etc.) |
|---|---|---|---|---|
| True on-prem inference | Yes (Dell AI Factory) | VPC endpoints + Outposts | Azure Stack / Foundry on-prem | Yes |
| Frontier-tier model quality | Yes (Codex / GPT-5.5) | Yes (Opus 4.7) | Yes (Opus 4.7 + GPT-5.5) | No (best open-weight gap is significant) |
| Native enterprise data integration | Dell AI Data Platform | Bedrock Knowledge Bases | Microsoft 365 graph | Self-built |
| Hybrid cloud + on-prem routing | Yes (workload-level) | Yes (regional + Outposts) | Yes (Foundry) | Limited |
| Identity / SSO | Customer IdP (Okta, AAD, Ping) | AWS IAM + customer IdP | Azure AD native | Self-built |
| Audit and compliance | SIEM-ready logs; SOC 2, ISO 27001 | AWS-native audit; HIPAA, FedRAMP | Microsoft 365 compliance suite | Self-built |
| Hardware footprint | Dell PowerEdge XE9680/85L | AWS-managed | Azure-managed or Azure Stack | Customer-managed (commodity GPUs) |
| License model | Per-developer seats + usage | Token-metered | Per-user seats | Open weights, self-managed |
The Codex-Dell offering’s clearest differentiator is the combination of frontier-tier model quality with true on-prem inference on Dell hardware most enterprises already own. Claude on AWS comes closest but requires AWS — buyers committed to non-AWS infrastructure have no comparable Anthropic option. Microsoft Foundry on Azure Stack is the parallel Microsoft offering. Local open-weight LLMs preserve full sovereignty but with a meaningful quality gap from frontier models.
The trade-offs. OpenAI Codex on-prem requires Dell hardware — customers on HPE, Lenovo, or Supermicro infrastructure don’t have a parallel path today (HPE and Lenovo will likely announce equivalent partnerships later in 2026). Pricing has not been published; the OpenAI-Dell deal structure is enterprise contract-based, so expect six-figure annual commitments minimum. The control plane connectivity to OpenAI means a full air-gap is not possible — the deployment is on-prem inference with cloud-based licensing and version management.
What’s next for Codex on-prem and enterprise AI distribution
Three threads to watch through Q3 and Q4 2026. First, HPE and Lenovo parallel partnerships. The competitive logic is obvious — both companies have AI server portfolios competing with Dell, and OpenAI will want broad hardware coverage. Expect at least one of HPE or Lenovo to announce a similar Codex partnership by end of 2026, with the other following in Q1 2027. Anthropic’s parallel motion will likely target the same hardware partners.
Second, the air-gap variant. Today’s Codex on-prem still requires control plane connectivity to OpenAI for licensing and updates. Highly-regulated defense and intelligence customers need a complete air-gap variant where the model weights ship as a deliverable and updates happen via physical media. OpenAI has not committed to this variant publicly but several Dell customers have indicated demand; expect a dedicated SKU within 12-18 months if the regulated buyer pipeline warrants it.
Third, the agent-platform expansion. OpenAI explicitly mentioned at the announcement that Codex on-prem is the foundation for non-code workflows too — report preparation, lead qualification, feedback routing, work coordination. Watch for Dell-OpenAI joint solutions in specific verticals (banking compliance, healthcare clinical documentation, government records management) emerging through 2026 and 2027. The Codex brand is broadening from “AI for engineers” to “AI for knowledge workers”.
Frequently Asked Questions
Does Codex on-prem require sending any data to OpenAI?
Inference data — your code, your prompts, your retrieval context — stays inside your perimeter. The control plane talks to OpenAI for model licensing, version management, and (if you opt in) telemetry about model usage. Customers can configure the control plane to send only aggregate license-verification messages with no payload data. A complete air-gap variant (no OpenAI connectivity at all) is not currently available but is on the roadmap for specific regulated customer segments.
What hardware do I need?
The minimum supported configuration is the Dell PowerEdge XE9680 or XE9685L with current-generation NVIDIA H200 or B200 GPUs, plus the Dell AI Data Platform stack installed and operational. Most existing Dell AI Factory customers already meet the requirements. For new deployments, Dell sizes the infrastructure based on expected concurrent developer count and workload mix — typical mid-size deployments need 1-4 PowerEdge XE-series servers. Dell’s sales engineering team handles capacity planning as part of the joint deployment service.
How does pricing work?
Neither OpenAI nor Dell published list pricing. The deal structure is enterprise contract — per-developer seats for Codex plus the Dell hardware and AI Data Platform license. Industry estimates put typical Codex Enterprise pricing at $60-120 per developer per month, and the on-prem variant is reportedly priced at a small premium reflecting the additional integration complexity. Total cost-of-ownership over three years for a 500-developer deployment is in the multi-million-dollar range, dominated by hardware and integration services in year one, by per-developer seat fees in years two and three.
Can I use Codex on-prem without Dell hardware?
Not currently. This partnership is specifically tied to the Dell AI Factory and Dell AI Data Platform. Customers on HPE, Lenovo, Supermicro, or other vendors will need to wait for parallel partnerships or use the standard OpenAI cloud offering. Both OpenAI and the hardware vendors have signaled that more partnerships are coming, but no formal timing has been published.
How does this compare to running open-weight models like Llama on Dell hardware?
Different value propositions. Open-weight models on Dell give you full sovereignty (no external dependencies) and lower marginal cost (no per-developer fees), but with a meaningful quality gap from frontier models — for code generation specifically, the gap between Llama 3.3-70B and GPT-5.5 / Codex is significant enough that most enterprise pilots end up choosing Codex despite the cost. The choice depends on stakes: for general copiloting and routine development tasks, open-weight may be sufficient; for high-stakes engineering where the quality difference matters, Codex earns its premium.
Is the on-prem deployment certified for compliance frameworks?
The Dell AI Factory infrastructure is certified for SOC 2, ISO 27001, and (for relevant deployments) HIPAA and FedRAMP Moderate. The Codex software layer adds its own SOC 2 and ISO certifications. For UK financial services, the FCA’s operational resilience expectations are met by the standard deployment pattern. For EU AI Act compliance, the on-prem deployment is well-positioned but specific high-risk system requirements (depending on use case) require additional customer-side controls. Work with your compliance team and Dell’s solution architects on the specifics for your sector.
How fast can a typical organization deploy Codex on-prem?
Dell estimates 6-12 weeks from contract signature to pilot-ready deployment for organizations already running the Dell AI Factory stack. Customers without an existing Dell AI footprint face longer timelines because the hardware procurement and AI Data Platform deployment have to precede the Codex install. Pilot phase typically runs 4-6 weeks; broader rollout follows pilot completion. The full path from “interested in Codex on-prem” to “1000+ engineers using it daily” is realistically 6-12 months for a mid-to-large enterprise.