Four takeaways from Dell's AI summit: Cost concerns drive enterprises back on-premises

Moving artificial intelligence from experimental phases into working systems has become the critical challenge facing large organizations. As companies move past initial pilots, they are grappling with more fundamental questions around expenses, data ownership and operational control. This evolution is reshaping how enterprises think about AI infrastructure, moving the focus away from model selection and hardware specifications toward the underlying systems, data pipelines and governance frameworks that support production deployments.

The emerging battleground centers on who will manage AI systems once they enter production. Rising costs per token are driving workloads back to on-premises environments, while the rise of autonomous agents is creating new demands for oversight of runtime behavior. Companies including Sycamore Labs Inc. are securing funding to develop governance systems specifically designed for AI agents.

No one is asking whether to adopt AI; that's pretty much a done deal. They're asking who controls it, where it runs and whose law it operates.

John Furrier, executive analyst at theCUBE Research

Furrier shared these observations after interviewing technology leaders at the Dell Technologies AI Leadership Symposium held in San Jose. Discussions at the event examined the path to production deployment, the vendor partnerships supporting large-scale AI infrastructure, edge-based voice models, agent governance mechanisms and the resurgence of CPU-based approaches for agentic AI systems.

1. Rack-scale deployments strengthen Dell's competitive position

Dell Technologies Inc. is concentrating its efforts on organizations ready to transition from experimental AI projects to operational systems, supported by a partner network that now includes venture capital firms. The company's established relationships with chip makers enable rapid deployment—Sean O'Connor, regional manager of ENT West, large enterprise and acquisition at Dell, noted the company can deploy 350 racks in just two weeks.

2. Voice models move from centralized data centers to disconnected devices

Deepgram Inc. develops speech recognition, text generation and voice agent models using Dell's infrastructure, with many of its clients running these models independently. According to Nick Mann, staff technical program manager at Deepgram, the next evolution involves deploying these models directly on AI-capable laptops in air-gapped configurations, bringing AI capabilities closer to end users.

3. Governance at runtime becomes essential for autonomous agents

Organizations are demanding that AI agents operate within isolated network environments under their direct control, with oversight shifting from predetermined rule sets to continuous monitoring of agent actions during execution. This shift matters because always-active autonomous agents require real-time supervision, according to Victor Jakubiuk, co-founder of MisaLabs Inc.; Shiv Agarwal, co-founder and chief executive officer of Singulr AI Inc.; and Sri Viswanath, founder and CEO of Sycamore Labs Inc.

4. Escalating token costs drive workloads back to local infrastructure

Organizations exhausting their monthly token allocations in a matter of weeks are experimenting with on-premises alternatives and reassessing CPU-based solutions for agent workloads. Reducing the distance between data and AI processing—whether through on-premises systems or AI-equipped personal computers—delivers meaningful cost savings, according to Helen O'Sullivan, AI business development manager and solutions specialist at Dell, and Tim Wood, sales director for the Northwest region at Intel Corp.