Open Models Change The Economics of AI

Open models are quietly taking over enterprise AI workloads

Open models have become good enough that enterprises are moving real workloads off frontier APIs, and Ollama sits in the middle of that shift. Morgan argues most tokens will run on open models while most spend stays with the frontier labs, which keep the hardest tasks and the stateful pieces open models still lack. Ollama's bet is that curation across a fragmented ecosystem is the valuable layer.

Takeaways

• Open models now carry enterprise token load. AT&T has shifted 40 percent of its token consumption to open models, according to The Information. Morgan says cost is the largest pain point open models solve, but control over AI and customization are the deeper draw for businesses.
• Coding agents and OpenClaw drove usage. Ollama Cloud token usage is up 150 times since January. Coding agents on Kimi and GLM arrived, then OpenClaw in April pushed context windows past a million tokens. Per-developer usage rose about five times, to roughly 15m tokens a week.
• Open models take tokens, labs take spend. Morgan predicts 80 to 90 percent of a business's tokens will run on open models while only 10 to 20 percent of spend goes to them. Frontier labs keep the hardest tasks, and a router can send orchestration there while line-item work runs open.
• Stateful layers stay outside the model. Morgan says the hard problems sit above the model, in coordination and execution. Memory and credentials stay outside it, and open models lack the security tooling closed providers ship, though security testing of software is one use case where they lead.
• Local hardware now runs serious models. Next-generation machines run 20 to 40bn parameter models, sometimes 120bn. NVIDIA's DGX Spark has 128GB of unified memory and stacks to run a 400B model. Tan says Qwen 3.8 38B now benchmarks with Opus 4.6 for coding.
• Customers ask where models run. Geopolitics starts with a model's origin, but Morgan says what matters to many customers is how and where it runs and whether that environment is secure. He cites a Finnish power plant running analytics on Ollama.

Ollama's day-zero model playbook

• Model developers contact Ollama before general release, so it gets previews and sees a usage spike when a new model drops.
• Ollama packages a harness or SDK for the model, the model itself with availability and capacity, plus the hardware and providers. Its Codex harness is open source.
• Morgan likens the stack to an operating system, where Ollama supplies the drivers and the inference layer and glues them together. That combination is combinatorially large.