The cloud is no longer the default answer for AI workloads.
At Dell Technologies World 2026, a partner list made that clear: frontier models once exclusive to the cloud are now being deployed inside enterprise data centers.
The list includes:
- OpenAI: Codex integrated with Dell AI Data Platform
- Palantir: Foundry and AIP deployed on-premises
- Google: Gemini 3 Flash on Dell Distributed Cloud
- Reflection AI: Full open-source frontier model stack on Dell AI Factory
- Mistral AI: Complete validated solution
- SpaceXAI: Grok as an enterprise assistant
- ServiceNow: Workflow automation
- Hugging Face: Enterprise Hub integrating open-weight models
One name is notably absent: Anthropic.
From Cloud Sales to On-Premises Rush
Reading the list backwards reveals the trend. A year ago, OpenAI was locked into an exclusive Azure deal. This May, it signed with Dell for Codex on-premises, a second path around Azure. Palantir's Foundry was built for clients who cannot use the cloud, but now Dell makes it a standard procurement item. Mistral is the European anti-US-cloud champion. Reflection AI, a US startup, handed over its entire model stack to Dell. Grok via SpaceXAI gives enterprises an entry point without buying xAI cloud.
Different motives, same action: pulling inference away from their own clouds.
Why? Jensen Huang provided data:
- 67% of AI workloads now run outside the cloud
- 88% of enterprises have at least one on-premises AI workload
The cloud-first narrative is crumbling under the agent economy. Agent workloads are long-tail, high-frequency, and persistent—SaaS billing by token is unsustainable. Uber's CTO admitted that its 2026 AI budget was exhausted in four months, mostly on agent tools. At that growth rate, cloud API bills break any budget.
Where Is Anthropic?
Many industry media asked that question at the conference. ServeTheHome bluntly stated:
"Where is Anthropic in this?"
Dell and Anthropic have not cut ties—Anthropic is prominent in SAP, Snowflake, and Dell's customer base. But the absence of Claude on-premises inference from Dell's list is a signal.
Possible explanations:
One: Anthropic's training-inference integration is too tight. Claude's inference performance relies heavily on Anthropic's proprietary optimization stack; handing over weights for standalone deployment degrades performance. OpenAI's Codex on-premises path is cleaner because Microsoft already built enterprise on-premises infrastructure.
Two: Anthropic's strategic choice. Dario Amodei has repeatedly said Anthropic aims to be a "deeply optimized managed service," not open-source or localized. Anthropic's recent Colossus 220,000 GPU contract with SpaceX goes the opposite direction—centralizing inference into a massive pool for deeper supply-side bargaining.
Three: Business logic differences. OpenAI, Mistral, and Reflection have partially open-sourced or semi-open models (Mistral under Apache, Reflection fully open, some older OpenAI weights on Hugging Face), making them technically suitable for on-premises. Anthropic's Opus 4.7 and Sonnet 4.6 are fully closed-source, monetized via cloud inference.
In the short term, Anthropic missing this on-premises frontier model wave is not a major issue. It already leads OpenAI in enterprise AI market share (Ramp April data: Claude 34.4% vs ChatGPT 32.3%), primarily targeting high-value cloud customers.
But long-term, this affects Anthropic's addressable market. When Dell+NVIDIA offers enterprises a complete on-premises package—hardware, runtime, agent sandbox, compliance, plus a choice of OpenAI/Mistral/Reflection—Anthropic will need another route into financial, healthcare, and government clients that require data to stay on-premises.
Sovereign AI Is No Longer Just a Political Term
"Sovereign AI" was a buzzword for European and Middle Eastern buyers meaning data cannot go to US clouds. Dell's list shows the boundary is now finer: every enterprise draws its own sovereign boundary.
Eli Lilly will not upload drug molecule structures to OpenAI's cloud. Hudson River Trading will not let Anthropic see its trading signals. Honeywell's industrial digital twin data involves supply chain secrets. These are business, not political, issues.
Dell packages hardware, models, runtime, and compliance, giving enterprise IT buyers a menu: choose hardware, choose model vendor, choose deployment form, and issue a purchase order.
Once this menu is standardized, the moat of "exclusive access to the most advanced models via the cloud" is partially filled. The second half of the year will show whether Anthropic joins the menu or continues its own path.
Sources: CocoLoop, Dell Tech World 2026: It's All About Sovereign and On-Premises AI (ServeTheHome); Dell builds a sovereign AI stack for the on-prem era (Fierce Network); Dell Technologies Closes the Gap Between AI Ambition and AI Outcomes (Dell official press release); NVIDIA CEO Jensen Huang at Dell Technologies World (NVIDIA Blog)