Everyone knows the frontier labs. Media coverage centers on their fight for the best frontier models. It is true that the hardest problems will require frontier intelligence. However, I'd like to shed light on the emerging open-source ecosystem taking shape.
Enterprises will use a portfolio of models. Frontier models for critical jobs, smaller models for easier tasks, and customized models where proprietary knowledge matters most.
I saw this first-hand on a personal project. A proprietary frontier model cost roughly $4,000 for a given workload. An open-weight model did the job at negligible cost. This was an email parsing and classification workload that did not require a frontier model. I would not have approval to run it on a recurring basis for $4,000.
There are useful workloads that become viable only when inference is dramatically cheaper.
Open weights make this possible, but downloading a model is not the same as running it in production. Requests arrive in bursts and use different amounts of memory. Providers have to batch traffic, manage memory, and spread work across GPUs without slowing the application or hurting quality.
The setup changes by model and workload. Each new release requires more engineering. Companies that serve many production workloads build experience balancing speed and cost while keeping expensive hardware busy. This is why the same model can have very different economics depending on where it runs.
When I think about the open-source ecosystem, a few players come to mind: NVIDIA, Nebius, Fireworks, and Palantir. NVIDIA is developing Nemotron as a family of open models, datasets, and training tools.
Most open models today focus on language, but this market should reach the physical world as well. NVIDIA Cosmos and Alpamayo are early examples for robotics and autonomous vehicles.
Nebius is building Token Factory on top of its AI cloud to serve open models at scale. Fireworks helps developers train, customize, and run those models. Palantir connects models to company data, permissions, and workflows.
A company may not use all of them. They show how much infrastructure is starting to form around open models.
Open models do not need to replace OpenAI or Anthropic to become a large market. Companies will use frontier models where maximum intelligence matters and open models where cost, control, or customization matter more.
As open models spread from text into the physical world, the infrastructure around them should become more valuable.
Sources