# AI application startups can build their own models

Markdown copy of https://realalpha.app/s/ai-application-startups-can-build-their-own-models/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Marc Andreessen](https://realalpha.app/people/marc-andreessen/) · Marc Andreessen is cofounder of a16z. · Co-founded Netscape and co-created Mosaic, then co-founded a16z, a venture firm backing AI, crypto and frontier tech.
- Host: [Erik Torenberg](https://realalpha.app/people/erik-torenberg/) · General partner, Andreessen Horowitz · Helped build Product Hunt; founded On Deck, Village Global and Turpentine.
- Show: a16z Podcast
- Episode: [Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI](https://realalpha.app/v/xRh2sVcNXQ8/), Jan 7, 2026
- Moment: 50:35 to 54:49 of the episode (4 min), playing on https://realalpha.app/s/ai-application-startups-can-build-their-own-models/; the episode from there: https://realalpha.app/v/xRh2sVcNXQ8/?t=3035

## Quote

> “the leading edge application companies are actually backward integrating and actually building their own AI models because they have the **deepest understanding of their domain**.”
> — [Marc Andreessen](https://realalpha.app/people/marc-andreessen/) at [54:02](https://realalpha.app/v/xRh2sVcNXQ8/?t=3242) · a16z Podcast

## Takeaways

1. A mature AI product may use many models for different tasks.
2. Application companies may build models tailored to their domains.
3. Open-source models offer an alternative to buying cloud tokens.

## About this segment

Marc Andreessen says leading application companies often use multiple models and may build models of their own. Their domain knowledge can make them more than wrappers around a model provider.

From “Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI”, Jan 7, 2026. [Watch the full episode](https://realalpha.app/v/xRh2sVcNXQ8/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[50:39](https://realalpha.app/v/xRh2sVcNXQ8/?t=3039) Another question you had posed at that point in time was, will incumbents versus startups win? And at that point in time, I think there was a mixed bag of where the incumbents were approaching AI. I think that's radically changed in the last two years. And then on the counter-example, the blossoming of startups increasingly now may be migrating into the incumbent category, just since that time. You want to take that question and give your assessment of where the state of the world is?

[51:07](https://realalpha.app/v/xRh2sVcNXQ8/?t=3067) Yeah, so, I mean, look, you know, big companies that are definitely, you know, playing hard, you know, Google's playing hard, Meta's playing hard, Amazon, Microsoft, you know, there's a bunch of these companies that are, you know, that are kind of in there, you know, very aggressively. And then you've got these, you know, what we call the new incumbents like Anthropic and OpenAI. But you also have like, you know, even in the last two years, you've had this birth of all of a sudden like brand new companies that are almost instant incumbents.

[51:30](https://realalpha.app/v/xRh2sVcNXQ8/?t=3090) And you could say XAI is one of those. Mistral, by the way, Mistral is the great outlier to my Europe thing from earlier. Like Mistral is actually doing very well as sort of the European kind of, you know, French national, European continental, you know, kind of AI champion. Sort of, you know, the exception that proves the rule. But, you know, there's a bunch of these now that are like, you know, doing quite well and are kind of becoming new incumbents.

[51:52](https://realalpha.app/v/xRh2sVcNXQ8/?t=3112) And then, of course, there's tons of startups. By the way, there's, and then there's actual foundation model startups, right? And so, you know, we funded, you know, we funded Ilya Sutskever out of OpenAI to do a new foundation model company. We funded Mira Murati also out of OpenAI. We funded Faye Faye Lee out of Stanford to do a world model foundation model company. And so, you know, there are new swings all, you know, all early but very promising to kind of build, you know, new incumbents quickly.

[52:16](https://realalpha.app/v/xRh2sVcNXQ8/?t=3136) And so, you know, that's all happening. And then, you know, and then on top of that, there's just this giant explosion of AI application companies, right? And so there's basically companies that then, usually startups, that basically take the technology and then, you know, field it in a specific domain, whether that's law or medicine or education or, you know, creativity or whatever. But again, here, it's just like, it's amazing kind of how sophisticated things are getting very quickly.

[52:43](https://realalpha.app/v/xRh2sVcNXQ8/?t=3163) So, I'm going to talk about the application companies for a moment. So, like an application company, a classic example is like a cursor. It's like an application company. So, they take the core AI capability, which they purchase by the drink from, you know, Anthropic or OpenAI or Google, you know, tokens by the drink. And then they build a code, basically a code editor, what we used to call an IDE, integrated development environment, or basically like a software creation system.

[53:06](https://realalpha.app/v/xRh2sVcNXQ8/?t=3186) So they'd build like an AI coding system on top of the Anthropoc or OpenAI or whatever, you know, kind of big models. So until that, and the critique of those companies in the industry has been, oh, those are what are called GPT wrappers. It's kind of the pejorative. And the idea basically being is, well, they're not actually like, they're not actually doing anything that's going to preserve value because the actual, the whole point of what they're doing is they're surfacing AI, but it's not their AI.

[53:29](https://realalpha.app/v/xRh2sVcNXQ8/?t=3209) The AI that's being surfaced is from somebody else. And so these are kind of these pass-through shell things that ultimately won't have value. It actually turns out what's happening is kind of the opposite of that, which is the leading AI application companies like Cursor. I mean, first of all, what they're discovering is they're not just using a single AI model. They actually, as these products get more sophisticated, they actually end up using many different kinds of models that are kind of custom-tailored to the specific aspects of how these products work.

[53:54](https://realalpha.app/v/xRh2sVcNXQ8/?t=3234) And so they may start out using one model, but they end up using a dozen models. and then in the fullness of time, it might be 50 or 100 different models for different aspects of the product, A. And then B, they end up building a lot of their own models. And so a lot of these, the leading edge application companies are actually backward integrating and actually building their own AI models because they have the deepest understanding of their domain.

[54:13](https://realalpha.app/v/xRh2sVcNXQ8/?t=3253) They're able to build the model that's best suited to that. And then by the way, also AI open source, they're also able to pick up and run on open source models. And so if they don't like the economics of buying intelligence, you know, buy the drink from a cloud service provider, you know, they can pick up one of these open source models and implement it instead, which, you know, which these companies are also doing.

[54:33](https://realalpha.app/v/xRh2sVcNXQ8/?t=3273) And so the best of the AI application companies are actually, they are actually full-fledged deep technology companies actually building their own AI. And so that, you know, that's, I think. Small models, though, right, Mark? When you think about God models versus small models