Ollama Built a $65 Million Round on an Engine It Borrowed

Type one line - ollama run llama3 - and about ninety seconds later, a language model that would have needed a research team and a rack of GPUs three years ago is answering questions on a laptop. No cloud account, no API key, no bill. That trick just earned its maker a $65 million Series B on top of 8.9 million developers running it every month, a spot inside 85% of the Fortune 500, and 176,000 stars on GitHub.¹² The natural assumption is that Ollama built the trick. It didn't.

The engine that actually does the work - compressing a multi-billion-parameter model down to something a laptop's chip can run - is llama.cpp, a separate open-source project a Bulgarian engineer named Georgi Gerganov started in early 2023, weeks after Meta's LLaMA weights leaked online.⁹ Ollama is a wrapper: a one-line installer, a curated model registry, and a friendly command sitting on top of an inference engine it didn't invent. For the better part of a year after Ollama's own July 2023 launch, its documentation never said so. A developer had to open a GitHub issue in April 2024 just to ask for a mention, writing that the project was 'heavily dependent on llama.cpp' with 'no mention of that in the readme.'⁶ Ollama's co-founder eventually added one line, roughly halfway down the README, under a header called 'Supported backends': 'llama.cpp project founded by Georgi Gerganov.'⁵ A separate GitHub issue, open since March 2024, still tracks a related complaint - that Ollama's release binaries didn't include the copyright notice llama.cpp's MIT license requires.⁷

Ollama's moat was never the engine. It's the on-ramp - and the two people who built it are running the exact play that already made them money once, wrapping a different piece of hard infrastructure somebody else had built.

It didn't build the engine. It built the front door - and this isn't the first door these two have sold.

The founders have run this exact play before: the same interface trick, a different engine underneath

Jeff Morgan and Michael Chiang aren't new to this. Before Ollama, they built Kitematic, a graphical interface that made Docker - genuinely hard, unix-permissions-and-namespaces infrastructure - installable and runnable on a Mac in minutes. Docker Inc. announced its acquisition of Kitematic on March 12, 2015, on undisclosed terms, and folded it into what eventually became Docker Desktop.³⁴ Docker didn't buy Kitematic for a container engine - Kitematic didn't have one; it drove Docker's own engine underneath. It bought the front door. Ollama is the same acquisition thesis, run by the same two people, minus the acquirer: instead of selling the front door to the company that owns the engine, they built the front door for an engine nobody owns, and this time they're trying to keep the toll themselves.

What Ollama actually owns: distribution, not the difficult part

None of this makes Ollama a fraud. Distribution is real value - arguably the only kind of value that compounds in open source. The company built a genuinely good curated model registry, a one-line install that behaves the same on a Mac, a Windows box, or a Linux server, and - since 2024 - an OpenAI-compatible API that lets developers point existing code at it without a rewrite.¹¹ That's the product 8.9 million developers show up for every month, and it's what Ollama is now monetizing directly: Ollama Cloud, a subscription layered on the same free local tool, priced at $0, $20 a month, and $100 a month, billed by GPU-time rather than a token count, for models 'too big to run on your own computer.'⁸ It's the Kitematic trade again - get millions of people hooked on a free, frictionless interface, then sell the next layer up before anyone else builds the same front door.

The honest objection: interfaces are real moats too: and Ollama is quietly paying down the debt that started this

The fair objection is that 'just a wrapper' undersells what a good interface is worth. Docker's own engine has been open and forkable since 2013; that never stopped Docker Desktop from becoming the default way people actually use it, because habit and polish are their own kind of lock-in. GitHub proved the same thing against raw git. Ollama is also less of a pure llama.cpp skin than its 2024 critics suggest - the company has been building pieces of its own inference engine, particularly for multimodal models that llama.cpp's architecture doesn't handle cleanly, rather than riding someone else's code forever.⁹ A company that starts by borrowing infrastructure and ends up owning more of it isn't running a con. It's walking the standard path from wrapper to platform - the one Docker walked, the one AWS walked on top of Linux, and the one Ollama is a few years into.

What an on-ramp can't lock down: the interface Ollama is betting on isn't proprietary either

The crack is that the interface itself is now a commodity. The OpenAI-compatible API - the exact habit Ollama's 8.9 million developers have built - isn't something Ollama owns. LM Studio, vLLM, and even llama.cpp's own official server all expose the identical routes, so a developer can repoint existing code at any of them with a one-line change.¹⁰ Kitematic's moat, such as it was, ended the day Docker decided to buy it instead of competing with it. Ollama doesn't have an acquirer waiting this time. It has a cloud giant, or Hugging Face, or Meta, any of which could bundle the same one-line local install into a tool millions of developers already have open - and erase the on-ramp overnight. The $65 million isn't proof the front door is locked. It's the clock Ollama is racing before someone bigger notices it never was.

Ollama didn't build a better engine. It built a better front door, and sold admission to the same people who already had a key to the building next to it. The last time Jeff Morgan and Michael Chiang made that exact bet, Docker paid to find out they were right. This time there's no buyer lined up - just a very fast clock, and a habit they're hoping 8.9 million developers keep long enough to matter.