Jeffrey Morgan on Open Models, Ollama and the Economics of AI

Ollama co-founder and CEO Jeffrey Morgan joins YC’s Lightcone to discuss open models, developer demand and the company’s path from an uncertain idea to a widely used tool. The discussion reflects his company’s perspective in September 2026.

Morgan reports growing enterprise interest in open models, particularly for coding agents and assistants. His observations come from Ollama’s users and should not be treated as a representative measurement of the entire market.

He says cost is an immediate reason customers consider open models, while control and customization can be longer-term goals. Actual savings depend on hardware, utilization, support and the workload.

Coding agents and other assistants can issue many model calls while working across tools and information sources. Morgan connects that behavior to the rise in token consumption he observes.

He describes Ollama Cloud usage rising sharply and distinguishes per-user growth from aggregate growth. The figures are company-reported usage for its own service, not an industry-wide count.

The hosts ask whether fine-tuning has returned after earlier enthusiasm faded. Morgan says frequent model releases can make a custom model obsolete quickly, while improved tooling may make updating it easier.

He discusses security and governance as obstacles to enterprise adoption. Open weights can offer control and flexibility, but a business must still evaluate data handling, model behavior, permissions and deployment risks.

The discussion includes security testing as a potential use case for specialized models. Morgan’s comparison of model behavior is his experience; suitability depends on authorization, safeguards and the actual system.

Launching a new model involves more than publishing weights. Ollama must support the architecture in inference software, check accuracy and speed, and make the model usable with relevant tools.

Morgan describes coordinating model developers, inference providers and hardware partners. A model that cannot run efficiently or cannot access the right capabilities may disappoint users despite strong benchmark results.

He likens the surrounding software to an operating-system layer. It must connect models, hardware, inference engines and applications across many combinations, which creates integration work.

The conversation turns to layers hidden between a model and an application, including serving, routing, security and state. The model is one component of a complete product.

Morgan expects open models to encourage specialized providers at different layers, as open-source software has in other parts of computing. That is a prediction about market structure, not an established outcome.

He distinguishes tasks that can be performed by a model from stateful systems that must store and manage information. Applications still need durable data and reliable operational components.

Morgan predicts that many business tokens will eventually run through open models, with frontier models used for some harder decisions or orchestration. The final mix will depend on capability, cost and customer requirements.

He imagines routing between models and providers rather than requiring a single model for every task. Customers would have to evaluate quality, latency, security and the complexity of operating the router.

Local inference and cloud inference serve different needs. Running locally can offer control and use existing hardware, while cloud infrastructure can provide larger models and shared capacity.

Morgan says smaller models have improved enough to be useful for some coding tasks. Specific model comparisons are time- and benchmark-dependent and should not be generalized to all workflows.

He reports different model preferences in Ollama’s local and cloud usage. Those observations reflect its user base and deployment options rather than a complete picture of global adoption.

The speakers discuss NVIDIA’s role in the open-model ecosystem, from hardware and software to models. Morgan sees a wider ecosystem forming around the ability to run models on different machines.

They return to the idea of running more capable models locally as hardware improves. Local deployment still has hardware purchase, power, maintenance and opportunity costs even when there is no per-call API charge.

Securing enough GPUs is a practical challenge for cloud services. Morgan says Ollama works with multiple providers and must match capacity to uncertain demand and model releases.

He highlights smaller, faster models as a possible way to reduce cost per completed task. Speed and token price matter, but the right comparison also includes how reliably a task is finished.

The hosts ask whether one exceptionally powerful model or many coordinated models will dominate. Morgan favors a mix of models and orchestration for different requirements; that remains a hypothesis about future systems.

The discussion also covers geopolitics and model origin. Morgan says customers differ in what they prioritize, including security, data handling, language behavior and where a model comes from.

Morgan and his co-founder previously worked on Docker Desktop. Their experience making software run across machines later influenced Ollama’s focus on developer experience and model execution.

The founders entered YC with an earlier idea and spent a long time searching for the right problem. Morgan says being experienced founders did not eliminate uncertainty or guarantee a quick product fit.

They explored security-related ideas and repeatedly spoke with potential customers. When conversations did not reveal a compelling need, the team revised its direction rather than treating the initial idea as fixed.

The eventual pivot toward running open models drew on the team’s systems background. Morgan wishes they had reflected earlier on what their particular experience made them well suited to build.

Ollama attracted substantial developer interest before it had a mature revenue model. Morgan says seeing users adopt the product was encouraging, while monetization still required deliberate work.

Ease of installation and local use can help a developer try a tool without a lengthy procurement process. Enterprise-wide use, however, can still require security review, governance and operational support.

The company considered how to monetize open-source adoption and eventually launched cloud services. Morgan frames the decision around solving a customer problem rather than assuming GitHub attention alone produces revenue.

As a repeat founder, he values candid conversations with other founders about what worked and what failed. A network can provide examples, but it cannot replace direct learning from customers.

The discussion shifts to which lessons from earlier infrastructure businesses still apply. Morgan says teams should reexamine assumptions as the needs of AI applications and the surrounding technology change.

He closes by emphasizing curation and integration: with many models and providers, a useful developer experience comes from making the pieces work together and helping users evaluate them.