Who Gets the Compound Interest on Your AI?

This week is Open Source AI Week, a community-led celebration of open-source AI.

To mark the week, we're sharing a conversation about one of the biggest questions facing organizations adopting AI:

Who owns the intelligence your business depends on?

In this episode of AI Masterminds, Raffi Krikorian, CTO of Mozilla, joins Manos Koukoumidis, co-founder and CEO of Oumi, to discuss the difference between renting intelligence and building systems that compound over time.

Meet the guests

Raffi Krikorian is Mozilla's CTO, where he leads efforts to build technology that serves the public interest and strengthens human agency.

Manos Koukoumidis is the co-founder and CEO of Oumi. He previously led generative AI work at Google Cloud, including Cloud PaLM, which evolved into what became Gemini. He focuses on helping organizations build, own, and compound their own intelligence.

Renting intelligence versus building it

The conversation starts with a simple premise: using a frontier model isn't the same as owning the intelligence behind your workflows.

Organizations that rely entirely on external model providers can face questions around control, economics, continuity, and differentiation. The answer isn't to abandon frontier models. It's to understand where ownership and specialization matter most.

The discussion covers five provocative questions:

• What happens when a provider changes access, pricing, or product direction?
• If every company uses the same frontier model, where does differentiation come from?
• Should organizations measure cost per token, or cost per completed task?
• Is the model really the product, or is the real value in the harness built around it?
• What's the first critical workload your team should move from dependency to optionality?

Raffi and Manos return to the same idea throughout the conversation: the goal isn't to pick one model forever. The goal is to preserve choice.

Start with one critical workload. Evaluate it carefully. Test alternatives. Build a runbook. Then decide where a specialized or open model can give your team better control, economics, or performance.

Oumi Fall Launch: AI to Build Agents and the Models Behind Them

Frontier models made it easy to start building agents. On November 4, Oumi will show how to define real tasks, build and evaluate agents, train the models behind them, and deploy the result. Your team keeps control of the models, data, evaluations, methods, and deployments that make the agent useful.

Register for Oumi Fall Launch →