Enterprises will run 80% of tokens through open models, but spend only 20% of budget
Open models and closed frontier models are not substitutes: the vast majority of tokens will go through open models, the hardest work stays with the frontier labs, and routing and orchestration connect the two in between.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The driver behind open models is control, not saving money
Jeff says cost is the biggest pain point open models can solve immediately, but the enterprise north star is ‘better control over AI and customising it for yourself’. Cost is just the piece that can be solved in the short term; only once that is unlocked does customising the model to your own use case come into play. The order matters: if you treat open models purely as a cost-cutting lever, you will misread why they are spreading so fast inside the Fortune 500. AT&T has already moved 40% of its token consumption onto open models, currently mostly US and European models, while also evaluating Chinese ones.
— Jeffrey MorganTwo inflection points: coding agents, then OpenClaw
Ollama's cloud per-user token usage curve has two clear inflection points. The first came early in the year, driven by coding agents — Kimi, GLM and MiniMax launched in succession, and for the first time open models could carry a coding agent. The second came in April, brought by OpenClaw, as users expanded from developers to non-developers: finance, customer support, marketing and sales can all hand a hard problem to an open model and let it run to completion, picking tools and pulling data along the way, burning an enormous number of tokens. Context windows growing from 128K to over a million is the technical precondition for this explosion.
— Jeffrey MorganModel launch day is a 24-hour fire drill
Ollama is a coordination channel for model makers ahead of launch, but Jeff says getting a model a few weeks early counts as lucky; most of the work is crammed into the final 24 hours before release. Three things have to land at once: inference engine support (often weeks of work to make sure it is fast, accurate and consistent with the reference implementation), the accompanying harness (there are plenty of open-source harnesses now, like Codex and OpenCode), and capacity from hardware and inference providers — launch day is usually the single biggest day of growth. He compares it to an operating system: drivers, the inference layer and the application runtime all have to be glued together, and it is a combinatorial explosion problem.
— Jeffrey MorganThe new scarcity sits above the token
Jeff cites three directions from Anthropic's platform team: knowledge (wiring company data and context into the model), coordination (a single request spawns a pile of sub-agents, some in the cloud and some local, that need scheduling), and execution (sandboxes and compute). His judgement: tokens are already in oversupply, dozens of open model providers can supply them, and what is scarce is how to orchestrate an agent from A to B above the token. Following the path of cloud computing, developers will ultimately prefer the best standalone product at every layer rather than something packaged together like Heroku or Google App Engine.
— Jeffrey MorganStateful things will not sink into the model layer
Asked whether the harness will be absorbed by the model, Jeff's judgement is that the core loop and hooks will sink into the model, but a few categories of things cannot. One is state — models are trained on a monthly cadence and can never catch up with the latest data, so storage is a huge standalone problem space. The other is credential management and security. Open models do not come with a whole security toolkit the way closed providers do, and that is critical to enterprise adoption. This explains why the stronger the models get, the more middle layers there are.
— Jeffrey Morgan80% of tokens run on open models, but only 20% of budget
Jeff's steady-state judgement: the vast majority of tokens inside an enterprise — his figure is 80% to 90% — will run on open models, but that does not mean 80% to 90% of budget flows to open source. The open community is driving cost extremely low, so maybe only 10% to 20% of spend lands on open models. The hardest work goes to the frontier labs, because that is where the best researchers are; the big middle is open and closed combined. He uses a law firm analogy: the partner hands work down to a pile of associates.
— Jeffrey MorganLocal and cloud model mixes are not the same thing at all
On local models, US and Chinese models are roughly neck and neck; on cloud-hosted models, it is almost entirely Chinese models. Jeff says local is a three-way mix of US, European and Chinese, while cloud coding agents are dominated by Chinese models. The division of use cases is clear too: coding agents write code and tests, hit many hard problems, and mostly rely on large cloud models; end-to-end low-difficulty workflows like document processing run very well locally. So it is a hybrid execution model — simple tasks run locally, plus a router deciding when to go up to a large model.
— Jeffrey MorganThe Flash tier is the precondition for unlimited tokens
Jeff says that early in the year open models had just caught up to the intelligence frontier, less than 3 months behind the closed frontier, and the next problem to solve is extreme efficiency. The key metric for models like DeepSeek Flash is not just low cost per token but low cost per task. This tier is good enough for 80% of tasks, fast and extremely cheap, and will become the workhorse doing the grunt work. He recalls that in the ChatGPT era nobody thought about token usage, and we will eventually get back to that state, but it requires model architectures trained specifically for high token volumes. The fastest-growing segment on Ollama's cloud is precisely the DeepSeek line, driven mainly by Flash adoption.
— Jeffrey MorganTwo years without finding the problem is scarier than no hype
Jeff says that when he entered YC in 2021 he wrote down a direction like ‘Kitematic for Kubernetes’ or ‘Docker Desktop for Kubernetes’, building security products such as SSO for Kubernetes. The company entered YC in 2021, and Ollama did not launch until July 2023 — more than two years in between. He says the scariest thing is not whether there is hype, but that ‘you cannot even see the north star’ — customers are usually a good north star, and not being able to see it is scarier still. The team was already more than 10 people at the time, and he is especially grateful to that group for sticking through several pivots.
— Jeffrey MorganFirst version in two weeks, and Llama 2 launched at exactly the right moment
The trigger for the pivot was a company-wide discussion in Toronto: what would you build if you started from zero today. They saw two problems — one, build a gateway to access any model (framed at the time as Segment for LLMs, the router idea that came later), and two, they were ex-VMware, ex-Docker people who were good at making things run, so make open models run. They gave themselves two weeks to build the first version, and Llama 2 launched exactly at the end of those two weeks, so they shipped straight away. Jeff calls this bias to action: from idea to launch in two weeks, with more users than all their previous products combined.
— Jeffrey MorganFree and runnable anywhere took a hobbyist tool straight into the Fortune 500
Jeff says what surprised them most was the speed from hobbyists to the Fortune 500. Two reasons: open models are free to start with, and they can run anywhere — which is critical for Fortune 500 IT developer teams, because they do not need to request permission to use them. What is good for hobbyists translated directly into something good for developers inside enterprises. He draws a database analogy: MongoDB also went from developers to the enterprise, but LLMs are stateless, which makes that transition even easier than databases.
— Jeffrey MorganThe rules of the infrastructure era do not hold in AI
Jeff gives two pieces of old muscle memory that have to be broken. One, ‘building a layer on top of someone else is dangerous’ — true in the infrastructure era, not true in the AI era, where moving up the stack may actually be better because you are closer to the customer. Two, the systems world demands that everything run exactly as designed, tested and verified, whereas an LLM by definition does not work that way — that is a feature, not a bug. He also mentions team size: some problems do not need that many people now, such as customer support pipelines, the way cloud services are delivered, and the very fact that ‘no engineer knows how all the code works’.
— Jeffrey MorganIn their own words · checked verbatim
I think the biggest thing we're seeing is a shift to open models, especially in enterprise and that's from a mix of US and and Chinese origin models and it's predominantly driven by coding agents and also AI assistants more co-work cases like openclaw and Hermes
Jeffrey Morgan1:03
Cost is by far the largest pain point that open models can jump in and solve but you know every business has a vision of getting better control over AI and customizing it for their business and that's really their north star.
Jeffrey Morgan2:04
The new scarcity the problems now are what's above the tokens right? You know how do you orchestrate an agent from you know A to B.
Jeffrey Morgan15:12
The super majority of tokens and this is our take it will be open models within a business. Call it 80 90%.
Jeffrey Morgan18:15
It's also low cost per task which is a really important metric and that class of models in my mind will be the first ones that come down to this idea of like unlimited tokens.
Jeffrey Morgan29:25
being lost in the wilderness like what's your north star that customers are generally a great north star, but not seeing the north star is even scarier, right?
Jeffrey Morgan41:30
there's this concept that if you're a layer on top of something else that you're in kind of a vulnerable position as a startup which is absolutely not true in the eye world and in fact going up the stack can sometimes be even better because you're closer to the customer.
Jeffrey Morgan53:38
Figures
| Ollama developers | 9 million | 1:03 |
| Ollama GitHub stars | 178,000 | 1:03 |
| Share of the Fortune 500 using Ollama | 85% | 1:03 |
| Share of AT&T token consumption already migrated to open models | 40% | 2:04 |
| Gap between open models and closed frontier models | less than 3 months | 29:25 |
| Growth in Ollama cloud token usage since the start of the year | 150x | 4:05 |
| DGX Spark unified memory | 128 GB | 25:22 |
| Parameter range runnable locally | 20 billion to 40 billion, some up to 128 billion | 21:18 |
| Expected share of enterprise tokens running on open models | 80% to 90% | 18:15 |
Glossary
- harness
- The outer code that connects model capability to a concrete task, handling tool calls and the loop.
- open weights
- Model parameters that can be downloaded and self-deployed, as opposed to closed models that only offer an API.
- router
- A middle layer that dispatches requests across multiple models by task difficulty and cost.
- MLX
- Apple's local inference stack for running large models on a Mac.
- DGX Spark
- A device that sits on a desk and can run models from 20 billion to 120 billion parameters.
How to listen
Founders building AI applications and agents, engineers who own enterprise AI procurement and architecture, and investors trying to judge the path to commercialisation for open models.
The one-minute show intro at the start and the YC recruiting spot at the end are skippable.