Editor's Note: The shutdown of Mythos this week has made many AI entrepreneurs re-acknowledge a problem obscured by cost discussions: When a product's core capabilities are built upon external models and platforms, what does a company truly own?
Over the past few years, open-source models have often been discussed within the framework of being 'cheaper alternatives to frontier models.' However, this article argues that cost is not the most critical variable; control is. For an AI company, calling frontier model APIs can quickly launch a product and lower technical barriers, but it also means core capabilities may be subject to the rules, pricing, strategy adjustments, or even takedown decisions of the model provider.
The article further proposes that 'owning intelligence' does not mean abandoning frontier models. Instead, enterprises should embed their own data, workflows, domain knowledge, evaluation standards, and edge cases into a controllable model system. Future AI competition may not be dominated by a single largest model but will likely feature multiple 'frontiers': general frontier models, enterprise-specific post-trained models, vertically specialized models, and routing systems composed of multiple models working in concert.
Therefore, the shutdown of Mythos serves as a reminder: the true moat in the AI era is not merely about accessing powerful models, but about whether intelligence can be turned into a company's own asset.
The following is the original text:
Mythos was shut down this week. Whether you agree with this decision is no longer the main point.
The real sting for many is this: a company built on intelligence it cannot control was suddenly exposed to a set of decisions it cannot influence. After witnessing this scene, many founders are likely asking themselves the same question: Which parts of my business are essentially just 'rented'?
In recent years, discussions about open-source models have mostly revolved around cost: Can they actually do the job? If so, how much cheaper are they compared to calling frontier model APIs?
Now, we have fairly clear answers. We've worked with companies like @RampLabs, @cursor_ai, and @harvey, following similar paths: starting with a powerful open-source model, post-training it with work content that genuinely matters to the company, and continuously benchmarking it rigorously against frontier models.
The results have been consistently surprising. For tasks that matter most to the business, an optimized open-source model can often approach or even reach the quality of frontier models at a fraction of the cost.
But what this week truly clarified is that cost was never the most important issue.
The deeper question is control. Who ultimately owns the intelligence your product relies on?
Many recent discussions have been summarized as the difference between 'renting' and 'owning.' This analogy isn't perfect, but it's useful.
Renting Intelligence
Renting works wonderfully—until it doesn't. An apartment is move-in ready; the lights work, the plumbing works, and repairs are someone else's responsibility. That's why most companies start down this path.
Frontier model APIs are phenomenal products. They enable startups to build things that seemed impossible just a few years ago.
But renting also means constraints. A landlord can raise the rent, decide what renovations you can make, change the rules. And occasionally, for reasons that have nothing to do with you, they can tell you it's time to move out.
You haven't done anything wrong. You've just been operating on someone else's turf.
That's also why the Mythos story resonates with so many people. When your core competency is entirely dependent on someone else's platform, you are exposed to a set of decisions outside your control.
Most of the time, it doesn't matter. But sometimes, it matters immensely in an instant.
Owning Intelligence
The lesson here isn't that companies should stop using frontier models. Far from it. The frontier model labs have produced extraordinary technology. Most products should use them. We use them ourselves.
In many senses, frontier models are becoming infrastructure. But infrastructure and ownership are two different things.
You can use public infrastructure while still owning what truly creates value for your business. In the AI world, 'owning' means starting with a state-of-the-art open-source model and shaping it around what is most unique to your company.
Your data.
Your workflows.
Your domain knowledge.
Your edge cases.
Your evaluation criteria.
Your definition of 'good.'
Over time, this model becomes less general and more reflective of the actual work your company handles every day. Value is created right here.
Think of it like a house. Moving furniture is easy; painting a wall is easy. But if your future depends on the layout of the house itself, then sooner or later, you'll want the ability to move walls. The same goes for intelligence.
When intelligence truly belongs to you, no one can silently pull the floor out from under your product.
This is also why we build Fireworks this way.
We integrate training and inference within the same system, enabling companies to adopt the best open-source models, shape them around the most critical aspects of their business, and deploy them stably into production.
Not just consuming intelligence. Owning intelligence.
There Is No Single Frontier
This week also held an optimistic revelation: the future of AI is not determined by one model winning everything.
There is no single frontier. There are many frontiers.
Frontier models are one frontier.
A model post-trained on years of company-specific knowledge is another frontier.
A specialized model that solves a narrow problem better than any other is another frontier.
A system that routes requests to multiple models and orchestrates them to work together, outperforming any single model on many tasks, is also a frontier.
The most interesting shift in AI is not that any single model is getting smarter, but that intelligence is becoming increasingly customizable.
The companies that ultimately win may not be those with the largest models, but those that can turn intelligence into their own unique asset.
Looking Ahead
Much of this week was spent reacting to the news, while we chose to continue releasing products: @Kimi_Moonshot K2.7 Code, @MiniMax_AI M3, @Alibaba_Qwen 3.7 Plus.
The future I anticipate is not one where a single model quietly consumes everything it sees.
But rather, one where many teams can own their own piece of the frontier.
If the shutdown of Mythos has prompted you to reconsider these trade-offs, we'd love to talk.
