Cost and Control: Enterprises Abandon Frontier AI Labs for Open-Weight Models
As token expenses mount into the tens of thousands per employee monthly, companies are shifting away from proprietary AI services toward open-weight models they can run on-premises, a trend reshaping enterprise AI strategy.

While the artificial intelligence boom has driven enterprises to pursue GPU capacity, the real obstacles to moving AI projects into production now lie elsewhere. Token pricing, data sovereignty concerns and decisions about which models to deploy are increasingly the factors determining whether experimental AI initiatives become operational systems, according to discussions at the SC26 supercomputing conference.
Dell Technologies Inc. has expanded its AI Factory strategy over the past year to encompass more than just computing infrastructure and storage hardware. According to Brad Maltz, senior director of AI solutions at Dell, what customers are seeking is an integrated solution spanning software, open-source frameworks, model selection and operational management.
The hardware's key, because that drives the tokens. You really, truly need a full-stack AI factory view above and beyond just the hardware stack.
Brad Maltz, Dell Technologies
Maltz discussed these shifts during an exclusive interview at SC26 with theCUBE Research's John Furrier and Gemma Allen, where Dell also announced updates to its AI Factory platform. The conversation centered on how token expenses and data sovereignty requirements are pushing enterprises toward open-weight models and comprehensive AI infrastructure solutions.
Why GPUs alone won't carry AI pilots into production
Numerous enterprises remain trapped in the pilot phase of AI deployment, with coding assistants representing one of the most accessible entry points. Since employees have already experimented with offerings from Anthropic PBC and OpenAI Group PBC, organizations naturally want to extend that familiarity to their own business problems. However, the financial reality emerges once these experiments scale up, Maltz noted.
The token spend right now in the frontier labs and the hyperscalers. That's a little more than we expected. We didn't expect to spend $10,000, $20,000 per person per month.
Brad Maltz, Dell Technologies
Cost represents just one consideration. Before moving a pilot into production, enterprises must address data sovereignty, protection of intellectual property, security protocols and governance frameworks. These requirements are directing many organizations toward open-weight models that they can customize through fine-tuning and operate within their own data centers, including domestically developed options like Nvidia Corp.'s Nemotron, Maltz explained.
There will always be frontier labs. There will always be open-weight models. The question is going to be, do we have enough diversity with the open weights to meet the enterprise use cases?
Brad Maltz, Dell Technologies
To support this shift, Dell established its AI Ecosystem Program to certify software vendors for compatibility with its AI Factory platform. Approximately 60 additional partners have submitted applications or are undergoing evaluation, and each new partner enables Dell to address between 10 and 30 additional customer requirements, according to Maltz. For organizations requiring sovereign deployments, this typically translates to isolated, on-premises infrastructure that keeps models and data within the organization's own hardware and control.
https://www.youtube.com/embed/JRm3qGGzpVE?feature=oembed
People understand data is the lifeblood of AI. It's the thing that trains the models. It's the thing the models operate against. It comes from your enterprise business applications.
Brad Maltz, Dell Technologies