Free Scope
✕
Want an estimate for something like this?
Tell us what you're building. We'll send a rough scope and timeline within two business days. No call required.
Start a project

Why pay for frontier models when open ones exist?

If open models are free to download, why do so many companies keep paying for frontier access? It’s one of the most common questions we hear from executives, and the honest answer depends on the task in front of you.

Achal Aggarwal, Rishav Dixit, Sid Verma
November 5, 2026

NVIDIA wants more than the chips

NVIDIA has been one of the biggest beneficiaries of the AI boom, supplying the hardware behind nearly every frontier lab. Achal’s read is that it has every reason to move into software as well. It already writes software for its robotics work, and once agent frameworks like OpenClaw took off, a share of that market became too attractive to ignore. In March it launched the Nemotron Coalition, a group of AI labs building open frontier models on NVIDIA’s cloud.

Achal can picture where that leads. A company spending heavily on tokens from a frontier lab every year could instead be offered a server rack it hosts itself, running open models, for a flat fee. Its usage would no longer be metered, its CFO would have a fixed number to plan around, and its data would stay in-house. The frontier labs, which currently sit between NVIDIA’s chips and the customer, would be left out of that arrangement entirely.

It’s the 1990s again, with better chips

Before AWS made cloud hosting easy, companies rented bare metal servers. Someone had to physically plug in the hardware and set it up, and you got direct access to the actual machine. Self-hosted AI racks would bring a version of that model back.

“It's actually going back to the '90s and early 2000s. It's such a weird cycle.”

— Achal Aggarwal

The tradeoffs haven’t changed much since then. You get control and a predictable bill, and you also take on the work of running everything yourself.

Why open models aren’t the default yet

Achal is direct about the downsides. Open models still trail the best frontier models on the hardest tasks. Most teams don’t have local hardware that can run them at the speeds OpenAI and Anthropic deliver, and the tooling for running open models well on your own servers is still difficult to use. Amazon, Microsoft, and Google have all tried to make it easier, and none of them has solved it for most teams.

That’s why we use both. Our PR review experiment runs Qwen 3 on our own servers, because that task is small and repetitive and the model handles it well. Work that needs the strongest reasoning still goes to frontier models, because nothing open comes close yet.

Is open safer or riskier?

The public debate tends to split in two. One side argues that a highly capable open model can’t be recalled or patched once it’s released. The other argues that open models make everyone safer, because more people can look for problems.

Our CTO, Siddhaarth Verma, sees it differently.

“Open source is not inherently less secure or more secure. If a vendor says your data is secure, you have to take their word for it. With open source, if you're good enough, you can go look.”

—Siddhaarth Verma, CTO

An open project can have weak security or excellent security. What it gives you is the ability to check, and anyone who finds a flaw can send in a fix. Open also doesn’t mean that anyone can change anything. Sid points to Linux as the example, where you can submit a contribution, and whether it gets accepted is still up to the foundation that maintains it.

There’s a real discomfort here too, and Maury Boswell names it plainly. Relying on a community feels reasonable for an ordinary app and much less comfortable for a technology with capabilities nobody has had before. Achal’s response is that most critical systems already run on software built by communities, and the people who build in the open are just as capable as the ones paid by a company.

Both kinds of models ask you to trust someone. With open models, you also get to check their work. Most open models are open weights, though, which isn’t the same thing as open source.

How we decide

We pay for frontier models when a task needs the best reasoning available and nothing open is close. We use open models when a task is narrow, repetitive, sensitive, or high volume, and a flat cost makes more sense than paying per token. We revisit that split often, because open models keep getting better.

Lessons you can steal

  1. Decide between open and frontier models one task at a time.
  2. Price self-hosting as a fixed cost plus the people who run it.
  3. Treat open models as something you can inspect, and check them before assuming they’re safe.
  4. Find out who decides which contributions get accepted in any open project you depend on.
  5. Keep an eye on NVIDIA’s coalition, since the math changes quickly once open models reach frontier quality on your tasks.

Free weights still cost something. Someone has to run them.

We work with open source and frontier models every week. Let’s talk about how it fits for you. Book a 15-minute intro.

FAQ

What’s the difference between open and frontier models?
Frontier models are the most capable models from labs like OpenAI and Anthropic, and you access them through those labs’ services. Open models publish their weights, so you can download them and run them on your own servers.
If open models are free, why isn’t everyone using them?
They still trail the best frontier models on the hardest tasks, most teams don’t have hardware that runs them fast enough, and the tooling for running them well is still difficult to use.
What is the Nemotron Coalition?
The Nemotron Coalition is an NVIDIA-led group of AI labs, announced in March 2026, that is developing open frontier models on NVIDIA’s cloud.
Why compare self-hosted AI to 1990s hosting?
Before AWS, companies rented bare metal servers that someone had to set up by hand. Self-hosted AI racks bring back the same tradeoff: more control and a predictable bill, along with the work of running everything yourself.
Are open models more secure than proprietary ones?
Neither is more secure by default. Open models can be inspected, so anyone with the right skills can check how they work and send in a fix if they find a problem.
Does open source mean anyone can change the code?
No. Major open projects have maintainers who decide which contributions get accepted. Linux is a good example, where the foundation still controls what goes in.
Interested in working together? Schedule a call.

More articles

Tech
Business

How to save $$$ on PR code reviews: What $25 a review does to good habits.

Business
AI
Tech

Does it matter where an open-weight model comes from?

Business
AI

How we use OpenAI Dots for business development: 5 agents, a single Trello board, and none of them can hit send.

‍