Ollama: How Agents Are Changing Open-Model Economics (#AI4SDLC)

This YC Lightcone episode with Ollama co-founder Jeff Morgan brings together several subjects that had previously appeared separately in the channel: token economics, local models, routing, and the rapidly changing agent harness. What makes the conversation valuable to me is that a familiar architectural pattern is now turning into a market structure.

Morgan's thesis is that agents sharply increase token consumption, making it uneconomical to use an expensive frontier model for every intermediate step. Low-cost open models will handle most of the work, while the strongest closed models remain available for difficult decisions and escalation.

The episode includes several estimates from Ollama 1️⃣ According to the company, total Ollama Cloud token volume has grown roughly 150-fold since the beginning of 2026. This is Ollama's internal usage metric, not an independent measure of effectiveness. 2️⃣ Morgan predicts that open models could process 80–90% of enterprise tokens while receiving only 10–20% of total AI spending. This remains a market participant's hypothesis, but the asymmetry is revealing: most of the work and most of the margin may settle in different layers. 3️⃣ The architecture will be hybrid: simple and sensitive tasks can run locally, heavier ones in the cloud, behind a shared interface.

Much of this supports ideas I have covered before. Token price is a poor description of agent economics: the useful unit is a completed and accepted task, including verification, retries, and the consequences of failure. Token-volume growth should not be confused with value growth either. In my OpenCode data analysis, I showed how session length, caching, and a small group of heavy users can reshape the aggregate number.

Routing has also moved beyond an elegant slide diagram. AT&T describes a workload of 45 billion tokens per day, a cache-aware router, and cost reductions of up to 90% in its use cases. The OpenCode story showed the same product bet: aggregate demand, validate model-provider combinations, and let an application switch between models.

What is new to me here

First, agent economics is joining privacy and sovereignty as a reason to choose open models. When one request expands into hundreds of model calls, a low-cost executor becomes necessary simply to scale the loop.

Second, Ollama wants to monetize the most perishable part of the stack: keeping the model, inference engine, hardware, cloud capacity, and agent harness compatible on release day. I recently argued that a generic harness is sensible to rent, while retaining ownership of the task, context, and evals. Ollama is trying to supply a similar volatile layer one level lower.

Third, the roles of familiar tools have become clearer. LM Studio gravitates toward the personal workbench, vLLM toward the high-performance serving engine, while Ollama is aiming at the distribution and compatibility layer between local execution, cloud capacity, and agent applications. That makes the Docker comparison more useful than treating Ollama as merely another way to run an LLM.

Open weights do not automatically make a system local, cheap, or secure. In my agent-stack configuration analysis, I separated the model, harness, tools, identity, and execution boundaries. A local model also moves spending from APIs into GPUs, operations, and underutilization risk.

If Ollama's bet works, open models will not eliminate platforms. They will create a new class of platforms above interchangeable weights. Companies should therefore own the layers where their real value resides: the task contract, state, permissions, reproducible episodes, and outcome evals. The remaining layers can be rented, but they should be revalidated after every meaningful model or harness shift.

https://www.youtube.com/watch?v=rY0wnfFHYbs

#AI4SDLC #AI #Agents #Architecture #PlatformEngineering #FinOps