Where did you get your AI model? In 2024, 53% of enterprise AI solutions were purchased rather than built internally, according to Menlo Ventures. A year later, that share rose to 76%.
Nick Patience, vice president and practice lead for AI at The Futurum Group, attributes that uptick in the “buy” strategy to two factors. A lot of 2024-era build efforts didn’t go so well (often due to a lack of data readiness), while the frontier API market matured, offering lower upfront cost and faster time to value.
But two things changed in 2026. As AI providers move to usage-based pricing models, customers are now grappling with unpredictable and often exorbitant inference bills. Secondly, access to the models themselves turned out to be conditional. When Anthropic temporarily suspended access to its Fable 5 and Mythos 5 models to comply with U.S. export controls, it made the risks of relying on models you don’t control very real.
Nick Patience. Credits: The Futurum Group
Some companies that have been building their own models argue it’s allowed them to save costs, achieve better quality, and maintain control. But training a model from the ground up is still a major endeavor. For companies that want all of the benefits without committing to a full build, a third option is emerging.
“Training a purpose-specific model on open weights that the company fully owns and can run anywhere, that's a distinct third category from either buying an API or building from scratch — and it's the one gaining the most real traction this year,” said Patience. “It gives portability without the enormous cost of pretraining, which is why I'd expect it to keep growing faster than either pole.”
Building from scratch
DeepL is one rare example of a company that chose to train its own translation-specific model.
Launched in 2017 and long a leader in translation, the company’s product predates the LLM boom and was built on earlier neural-network technology. But when DeepL decided to redesign its product with a new LLM-based architecture in 2024, the decision to build was an easy one for co-founder and CEO Jarek Kutylowski.
“[We’re using] LLMs that are really specialized for the kind of translation task that we're doing, and it's extremely important when you want to maintain certain quality levels and create models that are actually excelling at this particular task. You don’t want the model to be doing everything,” he said.
Loora team. Credits: Loora
Roy Mor, co-founder and CEO of English-learning company Loora, said that building a language model from scratch these days “is a crazy thing to do,” arguing there’s no reason to go that route because you don’t gain anything. But when it came to integrating a new video model to power the mobile app’s AI avatar, developing a custom model in-house won out on both cost and quality.
“The cost [of the off-the-shelf solution] just doesn't work out from a unit economics perspective. It's like 10x [the cost] for ElevenLabs, and also they're really not built for on-device and to be quick,” he said.
Because the scope was narrow, Mor said they were able to train a really small model using a custom architecture that runs on-device, eliminating the network round trip and creating a better user experience at a much lower cost.
This is where Patience says building makes more sense: when an AI capability is part of the actual product, pricing or competitive differentiation, or when the workload has ownership economics beating ongoing usage-based fees over a multi-year horizon. Another reason to build is when regulatory requirements make dependency on a single external provider a business risk rather than a convenience, he said.
A new way to own your model
Oumi, a Seattle-based startup founded by former Google and Microsoft AI engineers, is betting more and more companies will want ownership over their models, but they won’t want the burden of building them.
As an answer to this, Oumi created a platform to enable customers to automate the post-training of open-weight models to build AI models for specific use cases. Users start by describing the model they want to build in a prompt box using natural language. The platform then helps them compare open-weight models and curate their data before building it out. Models built with the platform are infrastructure-agnostic and can be exported and run wherever the user wants.
Manos Koukoumidis. Credits: Oumi
“You own everything. You own the model weights, the data, the recipes, which is how every step was executed,” said co-founder and CEO Manos Koukoumidis, adding that he believes companies need optionality and to not be beholden to anyone else for this vital technology.
Koukoumidis said their customers are seeking quality differentiation and control. When AI is a big part of their product, they feel they’re not differentiating if they’re using the same AI as everyone else, he said. Others want full control over privacy, security, where they deploy models and to make sure no one can take them away, he said. Because of the latter in particular, he’s found Oumi has especially resonated with companies in regulated industries. Oumi’s first enterprise customer was a top 10 global bank, which he said wanted higher quality, lower costs, and full control.
“I think the mindset is shifting,” said Koukoumidis.
