Are Foundation Models and Application Companies Friends or Foes?
In the early years of LLMs, there has been tight alignment between foundation model companies and the application builders on top of them.
The labs invested ungodly amounts of money to maintain a steady stream of model improvements. And the application companies took these models and built systems on top of them to productize AI for domains like coding, legal, finance, and security.
It didn’t make sense for either side to target the other. The labs were already underwater splitting their time across pushing the frontier and building consumer platforms. For application companies, trying to train models just wasn’t worth the opportunity cost given the speed of new model releases.
It was a good deal, powering the fastest-growing software companies in history and minting trillions in enterprise value. But this era of peace is coming to an end.
As OpenAI and Anthropic have focused their efforts on enterprise agents, they have started expanding up the stack, capturing more value and threatening to compete directly with their customers. This changes the equation for application companies, who now must own their technology to control their destiny.
Why Owning a Model Didn't Make Sense
Over the past few years, we’ve met dozens of AI application companies thinking about developing or fine-tuning their own model. For the vast majority, it just didn't make sense.
Frontier models were getting better and cheaper on a steep curve, and every dollar and every engineer-month you spent on your own model was a bet against that progress. Domain-specific data seemed promising, but in practice didn’t hold a candle to the rapid improvements coming from general pre- and post-training progress, and the development of reasoning capabilities.
The opportunity cost was killer. By the time you trained something good, the labs shipped something better and cheaper, and your investment was instantly obsolete. You would need to invest constantly just to try to keep up with the frontier, instead of spending your scarce engineering talent on the product your customers actually paid for.
What’s Changed?
That advice is starting to change. Three forces have shifted the calculus.
1: The Cost of Frontier Intelligence Isn't Falling
The labs keep shipping better models, but they're not cutting prices on the old ones; instead they deprecate them and push you up a tier. GPT-5.5 launched at roughly twice the API price of GPT-5.4. Opus 4.7 came in around 1.4x the price of Opus 4.6 due to its new tokenizer.
The flat-rate era is ending too. Anthropic shifted enterprise customers to per-token billing and cut off heavy third-party token consumers. Pricing is converging towards metered infrastructure, not all-you-can-eat.
This is the behavior of a supply-constrained market: as long as demand for intelligence outstrips the compute available to serve it, the labs have no reason to pass savings downstream. As they move towards IPOs, profitability pressure will only increase. For an application company, that means AI cost of goods sold is now a structural problem, not one you can wait out by riding the cost curve down.
2: The Labs Are Climbing the Stack
In addition to their pricing power over tokens, the labs are moving up the stack.
OpenAI and Anthropic both launched managed agent platforms this spring, and are pushing their customers in that direction. This does two things for the labs. It lets them capture even more of the value from their API customers as they move into more complex orchestration. And it gives them control and visibility over the agents. The labs frame this as a technical improvement: they can manage an agent's internal state and tools better than you can from the outside. That's partly true. But it also means API customers no longer see or control how the agent actually works, and that visibility is exactly what an application company needs to differentiate its product and own its relationship with the user.
Their aspirations go even further. Labs are starting to compete directly with their customers, both with out-of-the-box products (e.g. Claude Cowork) and bespoke implementations at large enterprises (e.g. OpenAI DeployCo). For any AI application company in general knowledge work, these developments are terrifying.
3: Open-Weight Models Are Good Enough for Many Tasks
While frontier intelligence holds its price, near-frontier open-weights intelligence keeps getting better and more abundant.
While it’s natural to want the shiniest new model, the truth is that we’ve already saturated performance for many economically valuable tasks: you don't need the absolute frontier to do most enterprise jobs.
A real menu of strong near-frontier bases now exists: Kimi, Qwen, MiniMax, DeepSeek. These models have the core capabilities required for enterprise work: reasoning and tool use. Continued training gets you to frontier or near-frontier performance on a narrow domain, at a fraction of the cost and time of building from nothing.
The Impact: A Renewed Desire to Own Your Models
Because of these forces, we're seeing renewed interest from application companies that want to own their own model stack.
The reasoning has evolved from the early days of LLMs. Back then, many companies had a thesis that proprietary data would let them train a better model than the labs. It was mostly wrong. The labs have developed new capabilities so quickly that a proprietary data advantage rarely closed the gap.
Today, the rationale isn't as much about beating the frontier on quality. It's about controlling your own destiny: predictable cost, strategic independence, and a hedge against suppliers who are now competing with you.
It doesn’t make sense for every company; owning your own model is still capital, talent, and infrastructure intensive. But for AI application companies or enterprises at large scale, the calculus may now favor a build approach. Cursor is the flagship example of this strategy with their Composer 2 model trained on a Kimi K2.5 base, but we have now talked to many companies that are considering or implementing their own in-house model strategies.
A Broader Infra Decoupling of Apps from the Labs
Owning the model is just the first sign of a broader shift. If your most important supplier is also your competitor, and it's bundling the agent harness as its own product, then standardizing your infrastructure on that lab's proprietary SDK is just deepening your lock-in to a potential rival.
We think this competitive evolution will drive broader demand for independent, model-agnostic AI infrastructure. This is the only way that application companies have a chance at capturing value over time, making the model a swappable component instead of a dependency.
Some examples of infrastructure that we think must remain independent from model providers are:
- Fine-tuning and post-training: most directly related to model ownership, independent serving and training stack to adapt and own your own model weights.
- Model gateway and routing: a layer that sends each request to the best or cheapest model for the job, across providers.
- Agent runtime: a provider-neutral environment to execute agent logic and tools, rather than a lab's proprietary SDK.
- Retrieval and knowledge: a multimodal, multi-workload data platform like LanceDB that gives the agent access to your corpus, independent of any provider's built-in RAG.
- State and memory: the agent's working set, including session state, checkpoints, and the filesystem it uses as a scratchpad and long-horizon memory.
- Version control: tracking changes to prompts, context, and agent logic the way you track code.
- Evals and observability: measuring quality and behavior consistently as you swap the model underneath.
Keeping Your Friends Close and Your Models Closer
We are still in the early innings of the co-evolution of foundation labs and application companies. But it’s clear that the relationship has changed; foundation models are now simultaneously the most important supplier and a credible competitor to the application layer.
We think this will drive a marked change in strategy for large application builders, and will be a huge tailwind for infrastructure platforms that enable companies to pursue a model-agnostic stack.
If you’re working in this space, I'd love to hear from you: at@theoryvc.com.
