# Applied AI will rely on a portfolio of specialized models

Markdown copy of https://realalpha.app/s/applied-ai-will-rely-on-a-portfolio-of-specialized-models/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Bret Taylor](https://realalpha.app/builders/bret-taylor/) · Bret Taylor is co-founder of Sierra. · Co-founded and leads Sierra, an enterprise AI agent company; chairs OpenAI's board and co-created Google Maps.
- Host: [Joubin Mirzadegan](https://realalpha.app/people/joubin-mirzadegan/) · Partner, Kleiner Perkins · Took Palo Alto Networks' cloud business from $2M to $50M+ ARR in a year.
- Show: Grit
- Episode: [How Sierra Is Pulling Ahead in the AI Race | Co-founder Bret Taylor](https://realalpha.app/v/6v9q6k8aF6E/), Mar 9, 2026
- Moment: 22:35 to 26:51 of the episode (4 min), playing on https://realalpha.app/s/applied-ai-will-rely-on-a-portfolio-of-specialized-models/; the episode from there: https://realalpha.app/v/6v9q6k8aF6E/?t=1355

## Quote

> “we use the term internally at Sierra, **a constellation of models** where you have different price performance, which is both latency and throughput.”
> — [Bret Taylor](https://realalpha.app/builders/bret-taylor/) at [25:19](https://realalpha.app/v/6v9q6k8aF6E/?t=1519) · Grit

## Takeaways

1. Venture capital expands the field of competitors, not token demand.
2. Model costs can fall even as some valuable tasks use more tokens.
3. Different AI applications will need different model trade-offs.

## About this segment

Bret Taylor argues that venture capital increases the number of competitors, but real business demand is driving token use. He expects AI products to combine models chosen for cost, speed and quality.

From “How Sierra Is Pulling Ahead in the AI Race | Co-founder Bret Taylor”, Mar 9, 2026. [Watch the full episode](https://realalpha.app/v/6v9q6k8aF6E/)

## Transcript

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

[22:39](https://realalpha.app/v/6v9q6k8aF6E/?t=1359) it's a good question i i don't think the venture economy is what's driving token demand i think it's actual like the revenue numbers are real in these companies so i think the venture this is my take and i'm not a sophisticated economist i'm just like a business person in this world but to me it's inflating valuations and it's like making it just really easy to start company it's like too easy to start companies like you're starting a company in a crowded market with no particular uniqueness you know in a in a market with less venture capital you wouldn't get financing for that now we're just in a different world we're like so that's sort of that's where i think the venture is playing in more in the quantity of competition as you said there's really smart people building these companies i don't mean it in a disparaging way at all it just uh you know

[23:27](https://realalpha.app/v/6v9q6k8aF6E/?t=1407) it's not it's there's not enough room in the market for all these players to to thrive as an independent company. But I think demand, like the revenue is real. And I think that's what's driving token demand. I mean, just look at software engineering. Like if you just look at all the other markets, you know, Quip and Sierra, Harvey, they go away and you just have software engineering that could probably saturate demand. That and chat GPT. And you've got enough demand for all these tokens. And I think the Moore's law characteristics of, you know, these models will drive token costs down.

[24:00](https://realalpha.app/v/6v9q6k8aF6E/?t=1440) because there's every incentive to do so and also scientific breakthroughs that facilitate it, just like the transistor in the early days of Moore's Law. The interesting thing is the way I see the models playing out is a little bit more of a heterogeneous set of models for applied AI applications. I think if you look at the pursuit of AGI, you end up using a lot of tokens because you do both that training but also inference time with these reasoning models, you're using a ton of tokens, but producing remarkable amounts of intelligence.

[24:37](https://realalpha.app/v/6v9q6k8aF6E/?t=1477) In the more applied AI world, you're seeing this with codecs and modern versions of Claude, where they're actually thinking for long periods of time, consuming a lot of tokens, but writing extremely high quality software. So, you know, it's not like the actual, you know, iterative, like each iteration of using codecs, like the cost hasn't gone down, it's probably gone up, but you're actually building software you couldn't do before.

[24:59](https://realalpha.app/v/6v9q6k8aF6E/?t=1499) So it's quite valuable. And then the other end of the market, you might be using, let's say you're making an LLM augmented thing to detect credit card fraud. Well, you can't, you don't want to use a reasoning model for that. It's too expensive and high frequency for that. So you might, that you probably want, you care about latency and other things. If you're doing voice AI, you care a ton about latency.

[25:20](https://realalpha.app/v/6v9q6k8aF6E/?t=1520) And so my sense is we're going to end up with a, we use the term internally at Sierra, a constellation of models where you have different price performance, which is both latency and throughput. And different applications are sensitive to both quality. And you're going to trade off all those different levers for your application. And if there's an analog in the pre-LM world, there's probably databases.

[25:43](https://realalpha.app/v/6v9q6k8aF6E/?t=1543) You know, like, I don't think there's a database you use for everything, right? If you want a trillion data points and you want to do analytics on it, you'll use one column or data store type thing. If you want a transactional database for credit card transactions, you'll use something different. And if you want to, you know, eventually consistent cash, you'll use a different thing. And engineers over the past three decades of building cloud applications have pretty much a shared set of rules of thumb to figure out which data storage technology you use for different applications.

[26:16](https://realalpha.app/v/6v9q6k8aF6E/?t=1576) uh i think that's the way models will play out over time and i think i i think we'll sort of grow up in you know five years and there'll be a lot of different um things that we call large language models today but sort of in that heritage of model um probably with new advancements just like the reasoning models and there'll be a lot used in concert for to power an agent and i think that'll be a really healthy um healthy progress in the ecosystem earlier you mentioned um hey like a market correction might be good here let's like play fake economists for a second like we don't