A soaring bill for rented coding tools forced engineers to rethink how they pay for intelligence. They came up with a better idea: build their own token machine.
Nutanix invested $20 million to build its own AI factory in 2026, betting that owning graphics processing unit (GPU) infrastructure to run open-source models will cost less over time than renting coding tools by the token. The move reflects a broader shift as enterprises seek control over artificial intelligence (AI) costs, data sovereignty and performance.
An AI factory is an integrated on-premises infrastructure stack that combines compute, storage, networking and software to run AI inference workloads and produce tokens inside a company's own data center. For Nutanix, sticker shock from soaring AI coding costs set them on this journey.
AI factories are emerging as an effective option for enterprises to reduce cost, accelerate token throughput and provide greater control as AI scales, according to Deloitte's AI factory industry trends report. "With an AI factory, organizations can build governed capabilities that improve innovation, risk management, token optimization, resilience and measurable value," the report said.
Nutanix took action in the summer of 2026. Ganesh Ramesh, technical director and AI agent architect at Nutanix, said his engineering teams adopted an AI-coding tool last year and saw near-full adoption. Then the coding tool shifted from API pricing to per-token pricing.
"We found the cost untenable," Ramesh said. “If we are going to spend $20 million on some solution that is going to only last us a year, we might as well take that and invest it in something that will meet our token needs for three or four years," Ramesh said.
New iterations of frontier AI models had shrunk from months to weeks. Ramesh said things kept evolving faster, forcing his team to stay nimble to keep up with new coding capabilities. Then pre-trained, publicly available open-weight AI models arrived and the path forward clicked into focus.
"My gut feeling was that these open-weight models are only going to get better and better," Ramesh said. "If we hosted them on our hardware, these models could free us from third-party providers. Then we can actually focus on making it better and better, and the model is no longer a bottleneck."
Chief Executive Officer Rajiv Ramaswami first addressed the $20 million investment during a September 2026 earnings call.
"We will, of course, still continue to go to the frontier models to get the best of the best, but we expect that to be a smaller portion," he said.
"For all the stuff that runs on our clusters with open-weight models, we no longer have to pay on a per-token basis. That's a one-time investment that we make, and then we can use it to the maximum extent over many years."
Nutanix aims to allocate 80% of tokens from the AI factory to on-premises AI models and the remaining 20% to frontier models.
Shifting to an owned asset posed many challenges. The team had to secure substantial funding, source hardware from a constrained supply chain and demonstrate a return on investment (ROI). Bringing teams together took work, and the early days were chaotic, according to Ramesh.
"It's almost like I'm not in an enterprise but in a startup," Ramesh said. "There is a lot of frenzy, a lot of things to get done."
The IT team, engineering, product and other teams together explored what it would look like to run the Nutanix software stack on the new hardware. This would give engineers a unified platform for accessing frontier and open-weight models, Ramesh said. It would enable them to select the right capabilities for each task in the software development process.
Inbar Stolberg, infrastructure software engineering lead at Nutanix, said the AI factory needs to provide a lower cost per inference than external services. The team would have to maximize the use of the AI infrastructure and maintain a rapid schedule for managing code changes to stay competitive. And the new clusters would require strict governance, resiliency and reliability.
"And we would have to move very fast or not do this at all," Stolberg said, describing the pressure IT professionals feel when they get the green light to embrace new AI capabilities.
The AI factory could provide a living use case for other IT teams to validate the cost, control and performance of using Nutanix software to manage infrastructure. It could become a model that helps enterprise IT teams across sectors develop strategies for managing AI costs and ensuring data sovereignty, explained Ramesh.
"We want to own our own intelligence, and that means managing AI on premises," Ramesh said.
Stolberg said the exploding demand for AI resources is driving the need for an AI factory.
"Any company that has thousands of software engineers needs this like oxygen," Stolberg said.
He is keenly watching how engineers are evolving in their role.
"The real transformation is how every single engineer at the end of the day is going to use this," Stolberg said. "There's lots of work ahead."
Editor's note: Read the Nutanix AI Factory solutions document to learn more about how Nutanix implemented an AI token factory using the Nutanix Agentic AI suite. The initial implementation focuses on developer workflows, providing a shared platform for model access, inference, policy, observability, and enterprise context.
Related stories:
Ken Kaplan is Editor in Chief for The Forecast by Nutanix. Find him on X @kenekaplan and LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.