Hybrid Cloud Hybrid Cloud Computing Platforms

Dell says AI costs and data control are reshaping enterprise model choices

Enterprises evaluating artificial intelligence are weighing more than access to GPUs as they move pilots toward production. At the SC26 supercomputing conference, Dell Technologies described token costs, data sovereignty and model choice as factors pushing customers to consider open-weight models and complete, locally operated AI systems.

Brad Maltz, Dell’s senior director of AI solutions, said customers increasingly want an end-to-end outcome that brings together hardware, software, open-source tools, models and operations. Dell has spent the past year expanding its AI Factory offering beyond servers and storage, he said. The company’s view is that infrastructure matters, but enterprises also need to plan for the systems and processes that use it.

Token spending changes the economics of pilots

Coding assistants are a common starting point for enterprise AI experiments, Maltz said, in part because employees have already tried products from Anthropic and OpenAI. But costs can become more significant when a pilot grows. Maltz said spending on tokens through frontier-model providers and hyperscalers was higher than some organizations had expected, citing potential monthly costs of $10,000 to $20,000 per person.

That figure was Maltz’s example of the spending enterprises may encounter; it was not presented as a typical bill for every customer or deployment. The broader issue, as he described it, is that a successful pilot can change the scale of inference demand. Organizations therefore have to assess operating costs alongside the performance and usefulness of a model before extending a trial to more users or business processes.

GPU procurement remains part of the equation, but hardware alone does not determine whether an AI project is practical to run. Dell’s full-stack approach is intended to account for the layers above the accelerators, including software and model operations. For enterprise teams, that means considering how a model will be deployed and managed, as well as the cost of generating its outputs.

Data sovereignty supports on-premises deployments

Cost is not the only reason companies may reconsider their initial model choices. Maltz also pointed to data sovereignty, intellectual-property exposure, security and governance as issues that need to be addressed before pilots become production systems. These concerns can lead organizations to favor open-weight models that they can fine-tune and operate on premises.

Keeping models near an organization’s data and GPUs can support deployments where customers want direct control over the environment, Maltz said. He described sovereign installations that may be disconnected from external networks. The data used to train models and the business information models operate on are central considerations, he added, because enterprise data comes from business applications and is closely tied to how organizations use AI.

Maltz cited Nvidia’s Nemotron as an example of a U.S. open-weight model option. He did not suggest that open-weight models will replace frontier labs. Rather, he described the market as likely to include both, with the enterprise question being whether the range of available open models is diverse enough to address specific business requirements.

Dell expands its AI Factory partner ecosystem

Dell has also introduced an AI Ecosystem Program to validate software partners for its AI Factory. Maltz said about 60 additional partners had applied or were under evaluation. He said each addition could enable Dell to address 10 to 30 more customer problems, linking partner breadth to the range of use cases the company can support.

The program is part of Dell’s effort to offer more than a hardware stack as customers develop AI systems. Software and model choices are particularly relevant for organizations seeking to run workloads in controlled, on-premises environments, where they may need to coordinate infrastructure with the data and applications they already manage.

The discussion took place in an interview at SC26 as Dell announced AI Factory updates at the conference. Dell sponsored the interview segment; the sponsorship disclosure stated that Dell and other sponsors did not have editorial control over the content. The remarks illustrate the practical trade-offs enterprises face as they move from experimentation toward deployment: model capability must be considered alongside token expenditure, data handling and the operational environment.