Custom model training is bringing enterprise AI from experimentation to production

As artificial intelligence moves from proof of concept into enterprise production, custom model training on governed data is emerging as the critical unlock for organizations that need domain-specific accuracy without sacrificing security or control.

The shift is pressing enterprise platforms to rethink how they deliver model training infrastructure. Rather than forcing customers to move sensitive data to external GPU clouds, the winning approach keeps training inside the governed environment where the data already lives, according to Dwarak Rajagopal (pictured, left), vice president of AI engineering and research at Snowflake Inc.

“We are extending the Cortex platform to actually train custom models in a safe, governed environment,” Rajagopal said. “What we hear from a lot of our customers is that training customized models for their unique use cases and their unique enterprise data is super critical. What we are making it easier for them to do is enable them to just focus on that problem and not worry about infrastructure, distributed systems [and] GPU sourcing. All of that, we handle it, and it makes it much easier.”

Rajagopal and Spiros Xanthos (right), founder and CEO of Resolve AI Inc.,  spoke with theCUBE’s Dave Vellante and Rebecca Knight at Snowflake Summit 2026, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how Cortex Training enables governed custom model training and why the combination of frontier and specialized models is becoming the standard for production AI. (* Disclosure below.)

Custom model training closes the gap between frontier AI and domain expertise

The agentic enterprise push at Snowflake Summit 2026 has made custom model training a centerpiece of the company’s strategy, with Cortex Training offering managed GPU infrastructure, enabling enterprises to fine-tune open-weight models such as Qwen and Mistral on proprietary data without leaving the Snowflake environment. A key engineering innovation driving efficiency is multi-tenant GPU utilization across workloads, which Snowflake said can deliver up to twice as many training runs for the same GPU budget, Rajagopal noted.

“Inference compute as well as training compute is a big currency right now,” he said. “One of the things our technology allows is to actually do multi-tenant [workloads] across different workloads within the same GPUs. We dramatically improve the utilization through technologies we’ve built recently; for example, ZoRRo, which is a zero redundancy rollout specifically for reinforcement learning training.”

For Resolve AI, which deploys site reliability engineering agents that autonomously debug and manage production software systems, the case for custom model training comes down to domain depth and performance economics. Production environments are noisy, tooling-heavy and data-infinite — characteristics that stretch general-purpose frontier models beyond their comfort zone, Xanthos noted.

“I see models as a way where intelligence is compressed within a model,” Xanthos said. “If you want to have a very specific business use case, you can do a better compression with a domain-specific model — not just the speed and latency, but also actual accuracy. For more generic use cases, frontier models are generally the best answer. But if you want to have … what Resolve is working on, that actually has a very good product market fit for a customized model to come and actually help.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of Snowflake Summit 2026:

(* Disclosure: TheCUBE is a paid media partner for Snowflake Summit event. Neither Snowflake, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE