AI infrastructure costs have been one of the biggest friction points for enterprise AI adoption. The more capable the model, the higher the inference bill. The more agents you run, the faster costs scale. Many businesses have been caught choosing between performance and price.

On June 4, Glean announced it added support for NVIDIA’s Nemotron 3 Ultra model to its platform — and the implications for enterprise AI economics are worth paying attention to.

What Glean Actually Did

Glean is an AI-powered knowledge platform used by enterprise teams to search and work across all their internal tools and data. Think of it as an intelligent layer on top of Slack, Drive, Confluence, Salesforce, and dozens of other systems that organisations accumulate over years.

Up until now, Glean customers have largely relied on frontier proprietary models from OpenAI, Anthropic, and Google to power the agents and search experiences inside the platform. These models are capable, but they come with frontier pricing.

NVIDIA’s Nemotron 3 Ultra changes that equation. It is an open model that, according to Glean’s announcement, delivers 91% of frontier LLM performance across key enterprise benchmarks — specifically completeness, the measure of how thoroughly an AI agent handles a complex task — while carrying the cost profile of an open model.

That gap between open model and frontier model used to be a meaningful one. For many practical enterprise tasks, it no longer is.

Why This Matters More Than It Looks

Glean’s platform now supports more than 30 models, including the newly added Nemotron 3 Ultra. This is not just a feature addition. It reflects a broader shift in how enterprise AI platforms are approaching the infrastructure cost question.

The key concept here is token economics. Every AI interaction involves processing tokens, and tokens cost money at scale. The more agents you run, the more queries your teams make, the more data you are processing — the more those token costs compound.

Glean’s own agentic search model, called Waldo, is post-trained on NVIDIA’s Nemotron 3 Nano and delivers 50% lower latency and 25% fewer tokens compared to previous approaches. That is not a minor operational improvement. At scale, it is a significant reduction in compute cost without a meaningful drop in output quality.

For enterprise teams, this means they can run more agents, more frequently, across more workflows — without hitting the cost ceiling that has kept many deployments conservative.

The Vendor Lock-In Question

Glean specifically highlights that Nemotron 3 Ultra gives customers flexibility to avoid provider lock-in. With 30-plus models available in the platform, customers can route different workloads to different models based on cost, capability, and risk tolerance.

This is becoming a real strategic concern for enterprise IT and data teams. Relying on a single AI provider creates dependence on their pricing decisions, uptime, and policy changes. Multi-model platforms that route intelligently across open and proprietary models give businesses more control over their AI stack.

The shift is subtle but meaningful: enterprise AI is moving from a “pick a provider” decision to an “orchestrate across models” operation. Glean’s move is one more step in that direction.

What This Means for Business

For business leaders thinking about AI costs, the pattern emerging here is worth tracking closely. The frontier performance ceiling is dropping in price. Open models are catching up in capability. In practical terms, the cost argument against deploying AI agents at scale is getting harder to justify with each passing quarter.

The constraint is no longer whether capable, affordable AI exists. The constraint is whether your business has the data foundations, the processes, and the people to actually make use of it.

Teams building data literacy and AI fluency now are the ones who will be positioned to take advantage as the cost floor drops. That is the window that is open right now — and it closes as competitors catch up.

If your business is exploring how to build those capabilities, the EDNA Learn platform has helped more than 220,000 data professionals develop the kind of practical, applied skills needed to get real value from enterprise AI investments.