Open models gain traction in enterprises, tech CEOs say

Technology leaders from IBM, SAP and Meta all said that usage of open models internally and among customers is gaining momentum. The comments are just the latest indicator that frontier models are sucking wind in the enterprise due to costs.

As previously noted, frontier large language models don't have much of a moat. Anthropic and OpenAI are pushing multiple models to cover use cases and price points, but enterprises are wary of locking into one model provider. The future will look a lot like what a CEO showed me: A Google Sheet of dozens of models doing various tasks and optimized for cost.

There was a mix of models with a few pricey LLMs and a lot of cheap ones. Loyalty? Absolutely none. He was managing to optimize price and performance. If one LLM fades, it will simply be replaced by another.

Enterprise vendor CEOs are seeing more model portfolio management with open models taking on the bulk of tasks. The upshot is that enterprises are taking aim at their frontier model costs.

Speaking on IBM’s earnings call, CEO Arvind Krishna said enterprises in 2024 had to go with a frontier model because of capability. No enterprise was going to accept an LLM that was cheap but 70% or more less capable.

The gap between frontier and open models closed in 2025, said Krishna. "In late, 2025 we began to see the gap shrink," said Krishna.

In 2026, open models can do about 90% of what the most advanced model can with lower infrastructure costs. Krishna said few enterprises are going to pay 4x more for what amounts to an easy answer for an LLM.

"Our computation is that for an average large enterprise, you can get going with maybe 2 or 3 racks, which is not a huge amount of infrastructure, but it gives you the ability to run those open-weight models currently anywhere from 25% to 10% of the token cost of the frontier models, that is very attractive given that they are now 90% as effective," said Krishna. "Would you like something that's 10% of the cost if it's 90% as good as the best? Or do you want to pay that 10x higher cost?"

• Microsoft's releases MAI-Base-2 aimed at open models with a twist
• Meta launches Muse Spark Open, may appeal to enterprises
• Nvidia’s Nemotron hit enterprise AI inflection point
• Mistral makes its open model case with Mistral Large 4
• Evidence Shows Enterprises Use of Open Weight Models is Mainstream
• Nvidia’s $12.93 billion Hugging Face acquisition validates open weight models

Krishna said the open model push rides along with sovereign AI and on-premises deployments.

SAP's Klein had a similar riff on open models:

"We are now checking every agent use case, we are shipping also on what is the best model and not only from a performance, from an outcome perspective, but also from a price perspective. And we are just at the beginning. So we are switching the models now more often. And we see also that we have an increasing share of open-weight models. By the way, not only for what we ship to our customers, but also what we do internally."

• Agentic AI deployments: What you should, and shouldn't do
• Enterprise AI projects mature: What you need to know now

Meta is creating its own models and using them internally. CFO Susan Li said it's an ongoing game of optimization and price and performance as the company rolls out a mix of open models with frontier efforts like Muse Wonder. "We are focused on compute efficiency and there are a couple dimensions to achieving that efficiency. Some of the near-term pieces include leveraging off-peak or spare computing capacity to run jobs using small models," said Li. "Over the longer term as LLMs scale we will get more tasks done with fewer tokens."

Mark Zuckerberg, CEO of Meta, added on the company’s earnings call: "We will be able to deliver the same intelligence that we have now in Muse Wonder or any of the current frontier models at a fraction of the cost in a smaller model a year from today."