---
title: Time to Stop Renting Your Intelligence
description: Companies are shifting from renting AI models to building proprietary ones, unlocking cost efficiency and improved performance in the evolving landscape of agentic intelligence.
---

[HubSpot Ventures Blog ](https://hubspotventures.com/hubspot-ventures-blog)

# [Time to Stop Renting Your Intelligence](https://hubspotventures.com/hubspot-ventures-blog/time-to-stop-renting-your-intelligence)

 Written by [Adam Coccari](https://hubspotventures.com/hubspot-ventures-blog/author/adam-coccari) | Jul 14, 2026 4:14:34 PM

We are entering an era of token scarcity, cost discipline, and shrinking gains from each new frontier model. For the last few years the default for agentic app was simple: call the best general-purpose model through an API, steer it with detailed prompts, give it tools, and improve behavior through evals. Things are changing quickly. Now, companies running AI products at scale will increasingly build and host their own models to power their flagship agents, rather than renting general-purpose intelligence from the frontier labs.

This is already happening, and the trend is accelerating. Cursor's in-house coding model, Composer, is the canonical example. It was built on an open source base and post-trained with reinforcement learning on real agentic coding tasks. Last week, Harvey announced that they post-trained an open-source Kimi 2.6 model for long-horizon legal work. Using this model as the main worker with Opus as an advisor, **they were able to** **match or exceed Opus’ quality at** **11x cheaper cost. **

These are some of the first clear examples of a trend we will see accelerate in the months ahead, as the benefits on cost for equal or better performance can be massive.

### How Companies Get There

These gains are available to any company with a large volume of real product usage data from customers and detailed agent eval traces. This data is the raw material for post-training an open-source foundation model with reinforcement learning and related techniques. A company like HubSpot, with millions of real customer interactions across its products and detailed evaluation traces from our production agent, is in a great position to do this.

The payoff is a proprietary model that deeply understands your product to drive your flagship agent experience. The companies doing this well are reporting higher accuracy on long-horizon tasks and better token efficiency, along with much lower long-term costs when hosting and optimizing the model for yourself.

Post-training is only one part of a larger puzzle though, as doing this effectively also means creating private eval rubrics for your models and (ideally) optimizing it for a specific agent harness. After deployment, you need to continually update and manage this model over time as the product changes and newer open-source base models become available. It’s a complex process, but we strongly believe that this the path to building differentiated app-layer businesses and capturing value through greatly improved margins. Companies that train on their proprietary data, evaluate against their own private benchmarks, and tune to their own agent harness will build the most durable moats.

### Startups Supporting Custom Model Development

A growing set of startups exists to help. Companies like [Trajectory](https://trajectory.ai/) and [Applied Compute](https://www.appliedcompute.com/) bring deep reinforcement learning expertise from frontier labs to help enterprises with the entire process. They embed engineers directly with their customers to turn product usage data into continually improving models. For the post-training with RL, startups like [Fleet](https://www.fleetai.com/) and [Deeptune](https://deeptune.com/) are building reinforcement learning environments, or “AI training gyms”, that simulate messy enterprise environments. To host and continually update the models, full-stack inference and fine-tuning providers like [Fireworks](https://fireworks.ai/), [Baseten](https://www.baseten.co/), and [Modal](https://modal.com/) supply the infrastructure. It shouldn't be a surprise that the usage and valuations of these companies are exploding as more companies decide to train and serve their own models.

The pieces are now in place: capable open-source foundation models, partners with real post-training expertise, environments to train agents safely, and infrastructure to serve the result at a competitive cost. We expect most scaled players building agentic AI to start down this path, and we’re excited to work closely with the startups helping them get there. The era of renting the best frontier intelligence for all tasks is over, and we are now moving to a world of building our own.

[View full post](https://hubspotventures.com/hubspot-ventures-blog/time-to-stop-renting-your-intelligence)

```json
{
  "@context" : "http://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Adam Coccari"
  },
  "dateModified" : "2026-07-14T16:14:34.480Z",
  "datePublished" : "2026-07-14T16:14:34Z",
  "headline" : "Time to Stop Renting Your Intelligence",
  "image" : {
    "@type" : "ImageObject",
    "height" : 719,
    "url" : "https://50339659.fs1.hubspotusercontent-na1.net/hubfs/50339659/1781109297518.png",
    "width" : 1280
  },
  "mainEntityOfPage" : "https://hubspotventures.com/hubspot-ventures-blog/time-to-stop-renting-your-intelligence",
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "height" : 60,
      "url" : "/hs/hsstatic/content_shared_assets/static-1.4092/img/default-amp-logo.png",
      "width" : 60
    },
    "name" : "HubSpot Ventures Blog"
  }
}
```