Who builds custom AI models? Applied Compute vs Prime Intellect vs Mercor and the rest
A year ago, "build our own model" meant hiring a research team. Now there is a small market of companies that will do it for you, and it moved fast this month: Palo Alto Networks is buying Applied Compute for $4.5 billion, and Prime Intellect launched a self-serve Model Factory. Here is who does what.
Updated 20 November 2026. Prices and features in this space change monthly; we re-check this page every month.
The main players at a glance
| Company | What they do | Best for | Watch out for |
|---|---|---|---|
| Applied Compute | Post-trains open-weight models on a company's own tasks and data, then runs them. Founded 2025 by former OpenAI researchers. | Large companies that want a finished model without building a team | Being acquired by Palo Alto Networks; expect a security-first focus and enterprise-size deals after the deal closes |
| Prime Intellect | Model Factory: train, deploy and keep improving your own model on one stack, on their GPUs or yours. Also open-source training tools and a hub of 2,500+ reinforcement learning environments. | Companies with ML engineers who want control and a fast start | You still need to define the task and the test; the platform doesn't do that for you |
| Mercor | A large network of paid domain experts (lawyers, doctors, bankers, engineers) who create training data and grade models. Now sells custom evaluations and post-training data to companies, not only to AI labs. | Building the benchmark that proves a model works, and the expert examples to train it | Not a model trainer on its own; pair it with a lab or platform |
| OpenAI and Anthropic | Fine-tuning and custom-model programmes on their own models, run on their platforms | Teams already on their APIs with a narrow task | You don't get the weights; the model lives on their service |
| Together, Fireworks and similar | Self-serve fine-tuning and hosting of open-weight models | Engineering teams doing a smaller, well-understood fine-tune | Simple fine-tuning, not reinforcement learning on complex tasks |
Sources: Palo Alto Networks press release, 18 November 2026; Prime Intellect's Lab and Model Factory announcements and documentation, November 2026; Mercor's enterprise evaluations pages, 2026. Customer results reported by vendors are their own claims and are not repeated here.
Applied Compute: the done-for-you lab, now inside a security company
Applied Compute was founded in 2025 by Yash Patil, Rhythm Garg and Linden Li, all former OpenAI researchers. It takes an open-weight model, trains it with reinforcement learning on a customer's real tasks, and runs it for them. It raised $350 million at a $3.25 billion valuation earlier this month. On 18 November Palo Alto Networks announced it will buy the company for about $4.5 billion in stock, with the deal expected to close in Palo Alto's fiscal third quarter of 2027. Patil becomes a co-chief AI officer there.
What it means for buyers: the clearest proof yet that large companies will pay for models they control. It also means one of the few done-for-you labs will now be part of a security vendor. If you are a Palo Alto customer, that may be convenient. If you aren't, ask any lab you talk to who might own it in a year.
Prime Intellect: the platform, under the line "own your intelligence"
Prime Intellect started with decentralised training and open-source tools (prime-rl for training, verifiers for scoring) and released its own models, including INTELLECT-3. Its new Model Factory packages the same stack for companies: train a model, deploy it, and keep improving it as new data comes in, on their GPUs, on a dedicated cluster, or on your own. Its Environments Hub, a public library of reinforcement learning tasks and graders, passed 2,500 community environments in November. It is SOC 2 Type II certified.
What it means for buyers: if you have even two good ML engineers, you no longer need to build training infrastructure. What you still need is the hard part: a clear task, a test, and examples.
Mercor: the experts and the test
Mercor built a marketplace of paid experts that the big AI labs use to create training data and grade models. In 2026 it began selling the same thing to companies: custom evaluation suites for a company's own agents and tasks, built in about six weeks according to Mercor, and post-training data from experts in the field. Its pitch for post-training is that you should know whether a custom model will beat the general one before you pay to train it.
What it means for buyers: the benchmark is the contract. Whoever writes the test decides what "good" means, so it should be you, with help, before you pick a lab.
How to choose
| If you… | Start with |
|---|---|
| Have no ML team and a high-value task at scale | A done-for-you lab, after you have a benchmark |
| Have ML engineers and want to keep improving the model yourself | A platform such as Prime Intellect |
| Don't yet know if a custom model would beat ChatGPT or Claude | An evaluation first: 200 to 500 real examples with correct answers |
| Just need consistent formatting or tone | Neither: good prompts and a well-built project usually get you there |
How we run it
We don't train models ourselves. We help you write the benchmark, keep the test set private, and let specialist labs bid with a price, a timeline and the score they commit to. You pay each milestone only when an independent run shows the model hitting the agreed score, and you own the weights. The full process.
Questions
Who is buying Applied Compute?
Palo Alto Networks announced on 18 November 2026 that it will acquire Applied Compute for about $4.5 billion in stock. The deal is expected to close in Palo Alto's fiscal third quarter of 2027.
What is Prime Intellect's Model Factory?
A platform launched in November 2026 for companies to train, deploy and continuously improve their own models on one stack, on Prime Intellect's GPUs, a dedicated cluster or their own hardware.
Does Mercor build custom models?
Mercor supplies expert-made training data and custom evaluations, and positions them for post-training. It is usually paired with a lab or platform that does the training itself.
How much does a custom model cost?
It depends on the task. A small fine-tune can cost thousands of dollars; reinforcement learning on a complex task with expert data usually costs from the low hundreds of thousands. Get a benchmark first so you can compare bids on results.
Do we own the model?
Only if the contract says so. With open-weight models trained by a lab or on a platform you can own the weights. With fine-tuning on OpenAI or Anthropic, the model stays on their service.