OpenAI and Dell push on-premise Codex: the end of cloud-only for enterprise AI

OpenAI and Dell push on-premise Codex: the end of cloud-only for enterprise AI

🔎 Why OpenAI is abandoning its all-cloud dogma

On May 18, 2026, OpenAI signed an agreement with Dell Technologies that would have been unthinkable two years ago. The AI giant, built on a 100% cloud infrastructure, agreed to deploy its Codex agent on the customers' own servers. This is not a technical adjustment. It is a strategic U-turn dictated by a reality that Sam Altman could no longer ignore: companies are not going to send their most sensitive data to OpenAI's cloud.

The context adds a layer of tension. Recently, Anthropic denied China access to the Mythos model, illustrating the Sino-Western AI cold war. In this climate of geopolitical distrust, Western and Asian executive boards are demanding digital sovereignty guarantees that the cloud alone can no longer provide.

And then there are the numbers. Codex is used by more than 4 million developers per week, according to StartupHub (May 2026). But this success remains confined to non-sensitive use cases. The Dell partnership specifically aims to unlock blocked sectors: banks, hospitals, defense, government agencies. Enterprise AI will not happen exclusively in browser tabs.

The essentials

• OpenAI and Dell announce the Dell AI Factory with OpenAI Codex on May 18, 2026, to deploy Codex on-premise and in hybrid environments.
• Dell becomes the physical distribution channel for OpenAI's frontier models to enterprises, according to Forbes.
• The goal: enable companies to run AI where their data resides, without exposing it to the public cloud.
• The targeted sectors are finance, healthcare, defense, and any European company subject to the GDPR.
• This partnership is part of an AI infrastructure war between Google (Antigravity + Vertex AI on-prem), Microsoft (Azure AI), and now the OpenAI-Dell duo.

Recommended tools

What the Dell AI Factory with OpenAI Codex changes

Specifically, Dell integrates Codex into its AI Factory, a hardware-software infrastructure designed for local AI deployment. The Codex agent, which ranks 13th on the agentic benchmark with a score of 80 (GPT-5.3 Codex, OpenAI), is no longer confined to the cloud API. It runs in the customer's datacenter.

Ihab Tarazi, SVP and CTO of Dell Infrastructure Solutions Group, puts it clearly: the Dell AI Factory enables enterprises to deploy AI where their data resides. This marks the end of the "send us everything, we'll send you the result" model. The reverse paradigm is taking hold: AI goes to the data, not the other way around.

This hybrid architecture solves a fundamental problem. Companies possess proprietary codebases, patient records, and financial transactions that cannot cross the boundaries of the public cloud. The OpenAI-Dell partnership creates a bridge between the power of frontier models and local regulatory requirements.

The Dell AI Data Platform serves as the orchestration layer. It connects internal data silos to the Codex agent while maintaining a strict security perimeter. This is integration engineering, not magic. But it is precisely what CIOs have been waiting for.

Dell's positioning: much more than an integrator

Dell is not simply playing the role of a server reseller here. According to Forbes (May 2026), the company positions itself as the "distribution channel for OpenAI's frontier models to C-suites." The nuance is important. Dell sells physical access to OpenAI's models, not just hardware.

This is a major strategic move against hyperscalers. Google has its Antigravity and Vertex AI on-prem. Microsoft has Azure AI and Copilot, with a natural advantage through its historic partnership with OpenAI. By aligning directly with OpenAI, Dell bypasses the dependency on Azure for on-prem deployments.

The Dell Enterprise Hub on Hugging Face completes this strategy. According to Morningstar (May 2026), this hub provides on-prem access to recent open-weight models: MiniMax-M2.7, DeepSeek Pro, DeepSeek-V4, GLM 5.1 and Kimi K2.6, all optimized for the Dell AI Factory. The company therefore offers a complete catalog: proprietary OpenAI models on one side, open-weight models on the other. An ecosystem closed off by hardware but opened up by the diversity of models.

This multi-catalog approach is smart. It allows Dell to avoid being locked into a single model provider while remaining the sole physical point of entry. For the customer, it means a single point of contact for the entire AI stack, from server to model.

Sectors waiting for this offering

Banks are the most obvious use case. An agent like Codex can analyze codebases of trading systems, audit smart contracts, or automate legacy migrations. But no bank is going to send its source code to OpenAI's cloud. On-prem deployment via Dell removes this regulatory and cultural barrier.

Hospitals represent another critical segment. Medical records, subject to GDPR in Europe and HIPAA in the United States, require strict compartmentalization. A locally deployed Codex agent can assist with entering clinical notes, analyzing protocols, or developing internal tools without a single byte of patient data leaving the facility.

Defense is the most sensitive and potentially most lucrative sector. National security constraints rule out public cloud outright for critical applications. The Dell-OpenAI partnership opens the door to use cases that were previously inaccessible for OpenAI's models: cybersecurity log analysis, code generation for embedded systems, processing classified data within an air-gapped perimeter.

European companies, subject to GDPR, also benefit directly from this offering. The transfer of data to the United States remains a tense legal issue, despite the Data Privacy Framework framework. Being able to deploy Codex on a server based in Frankfurt or Paris eliminates the legal risk. This is a selling point that Dell's sales teams will exploit without moderation.

The AI infrastructure war: Google, Microsoft, OpenAI-Dell

The enterprise AI landscape in 2026 is a three-front war. Each camp offers a different vision of where AI should live.

Google has bet on Antigravity, its specialized compute infrastructure, coupled with Vertex AI on-prem. Gemini 3 Pro Deep Think, the world's second agentic model with 95.4 points, can be deployed in the Google Cloud Anthos infrastructure that extends to the customer's datacenter. Google's advantage: complete vertical integration, from TPU to model.

Microsoft remains OpenAI's historic cloud partner via Azure AI and Copilot. But this position becomes ambiguous with the Dell partnership. OpenAI is partly bypassing Azure for on-prem deployments, which creates tension within the Microsoft ecosystem. Azure remains essential for the cloud, but Dell is taking over the on-prem territory.

The OpenAI-Dell duo offers a third way: separating the model from the cloud infrastructure. OpenAI provides the intelligence, Dell provides the metal. This "decoupled" approach appeals to enterprises that want to avoid cloud vendor lock-in while accessing frontier models. It's pragmatic, not ideological.

The table below summarizes the positions:

The impact on self-hosting and sovereign AI

The Dell-OpenAI partnership accelerates an already ongoing movement: the rise of self-hosting AI models. The Dell Enterprise Hub on Hugging Face, with its models optimized for the Dell AI Factory, is the physical proof of this.

Moonshot AI's Kimi K2.6, ranked 7th with 88.1 points in self-host mode, and Z.AI's GLM-5 with 82 points, are available via this hub. These open-weight models offer a credible alternative to proprietary models for enterprises that want full self-hosting, without depending on OpenAI.

This is where Dell's strategy reveals its full subtlety. By offering both Codex (proprietary) and open-weight models on the same infrastructure, Dell covers both scenarios. The company that wants the best agentic model takes Codex. The one that wants total control and zero vendor dependency takes Kimi K2.6 or DeepSeek-V4. The hardware is the same. The lock-in occurs at the hardware level, not the model.

Sovereign AI, particularly in Europe, emerges strengthened from this dynamic. Governments can require that models run in a given territory, within a certified datacenter. Dell, with its physical presence in most European countries, is positioned to meet this demand in a concrete way. OpenAI, via this partnership, accesses these markets without having to build its own European infrastructure.

Codex vs. agentic competition in the enterprise environment

The on-prem deployment of Codex raises a valid question: is it worth deploying a model that ranks 13th on the agentic benchmark? GPT-5.3 Codex scores 80 points, far behind GPT-5.5 (98.2), Gemini 3 Pro Deep Think (95.4), or Claude Opus 4.7 Adaptive (94.3).

The answer depends on the use case. Codex is not designed for general reasoning. It is a code-specialized agent: generation, review, refactoring, migration. In its domain, its actual performance exceeds what a generic agentic benchmark suggests. A score of 80 in agentic does not mean 80 in code generation.

Furthermore, on-prem deployment changes the game. When the choice comes down to "Codex on-prem or nothing", the exact score matters less than availability. No bank is going to choose GPT-5.5 in the cloud over Codex on its own servers to analyze its source code. Security trumps raw performance.

Nevertheless, the competition will react. Anthropic could offer Claude Opus 4.7 (94.3) or Claude Sonnet 4.6 (81.4) on-prem via a hardware partner. Google could strengthen Vertex AI on-prem with Gemini 3 Pro Deep Think. The war is only just beginning on this specific front.

What this means for enterprise AI architecture

For technical architects, the Dell-OpenAI partnership adds a new option in an already complex landscape. The decision is no longer simply "cloud or on-prem". It is now "which model, on which infrastructure, with what level of control".

The hybrid approach recommended by Dell via the Dell AI Data Platform suggests a layered architecture. Sensitive data remains on-prem, processed by Codex deployed locally. Less sensitive tasks can continue to use cloud APIs for more powerful models like GPT-5.5 or Gemini 3 Pro Deep Think.

This multi-tier architecture demands new skills. Teams must master model orchestration between cloud and on-prem, manage response consistency, and monitor performance and latency across heterogeneous infrastructures. This is not trivial, and it is exactly where Dell hopes to sell its integration services.

Cost is also a determining factor. The on-prem deployment of frontier models requires premium GPUs, infrastructure management, and support. The TCO (Total Cost of Ownership) of a Dell AI Factory with Codex will be significantly higher than an API subscription. But for regulated industries, this cost is a compliance investment, not an optional expense.

❌ Common mistakes

Mistake 1: Confusing on-prem and air-gapped

Deploying Codex on a Dell server in the company's datacenter does not mean the model is disconnected from the Internet. The on-prem architecture described by OpenAI and Dell is hybrid: data remains local, but the system may need a connection for model updates or telemetry. If you need a true air-gap (defense, intelligence), require a specific architecture and validate every network flow.

Mistake 2: Choosing Codex on-prem for non-sensitive use cases

If your data is not subject to strict regulatory constraints, the on-prem deployment of Codex is probably an unnecessary extra cost. Cloud APIs give access to GPT-5.5 (98.2) or Gemini 3 Pro Deep Think (95.4), which are much more performant than Codex (80) in agentic. Reserve on-prem for data that genuinely needs it.

Mistake 3: Ignoring open-weight alternatives via the Dell Enterprise Hub

Dell doesn't just offer Codex. The Dell Enterprise Hub on Hugging Face gives access to Kimi K2.6 (88.1 in self-host), GLM-5 (82), DeepSeek-V4 and other models. Before signing for proprietary Codex, evaluate whether an open-weight model meets your needs. You will gain flexibility and avoid dependency on OpenAI.

Mistake 4: Underestimating the integration complexity of the Dell AI Data Platform

Marketing suggests a "turnkey" deployment. The reality is more nuanced. Connecting Codex to your internal data silos via the Dell AI Data Platform requires integration work, security configuration, and permissions management. Plan for a multi-month integration project, not a weekend deployment.

❓ Frequently Asked Questions

Is Codex on-prem identical to the Codex cloud?

No. The on-prem version via Dell AI Factory is optimized for local deployment, with possible trade-offs regarding model update latency and certain features requiring a cloud connection. The core of the model remains the same, but the execution environment differs.

What is the cost of a Dell AI Factory with Codex deployment?

Prices are not public and are based on enterprise quotes (May 2026, check on openai.com and dell.com). Expect a significant initial investment in Dell hardware, plus recurring license fees for Codex. The TCO is measured in the hundreds of thousands of euros per year for a deployment at scale.

Does this partnership threaten the OpenAI-Microsoft relationship?

Yes, partially. Microsoft remains OpenAI's main cloud partner for deployment in Azure. But for on-prem, Dell bypasses Azure Stack. This is a classic tension in tech ecosystems: the cloud partner wants to control everything, while the model provider wants to maximize its distribution channels.

Are the open-weight models from the Dell Enterprise Hub as performant as Codex?

In pure agentic tasks, Kimi K2.6 (88.1) outperforms Codex (80). But in specialized code generation, Codex remains competitive. The choice depends on your specific use case, not just benchmark scores.

Does this partnership concern models other than OpenAI Codex?

The Dell AI Factory is a generic platform. The partnership announced in May 2026 focuses on Codex, but the infrastructure can host other models. Open-weight models via the Dell Enterprise Hub (MiniMax-M2.7, DeepSeek Pro, DeepSeek-V4, GLM 5.1, Kimi K2.6) are already available on the same infrastructure.

✅ Conclusion

The OpenAI-Dell partnership of May 18, 2026, marks the end of the illusion that the cloud could absorb everything in enterprise AI. By enabling the on-prem deployment of Codex via the Dell AI Factory, OpenAI acknowledges that sensitive data will not move. For architects and decision-makers facing regulatory constraints, it is a new and serious option to evaluate — provided it is not confused with a universal solution.