back to indexState of Startups and AI 2025 - Sarah Guo, Conviction

00:00:25.800 |
What is definitely happening by the end of 2026? 00:00:29.680 |
AI agents ship code directly to prod in your environment, right, not in some playground. 00:00:36.960 |
Voice AI replaces text for most business communication. 00:00:42.120 |
Inference cost drops below a cent per million tokens. 00:00:54.800 |
First one, ship code directly to prod, okay, this is a hopeful set of engineers. 00:00:59.360 |
All of you want to get rid of your own jobs, I love that. 00:01:06.640 |
The good thing is I also don't have Internet, so I can't look at my next question. 00:01:14.640 |
No, it's going to be good, it's going to be good. 00:01:25.200 |
Oh, no, I was going to go through poll questions while we're trying to do AV setup. 00:01:41.480 |
While this is happening, I'm actually just going to introduce myself so we're not wasting the time. 00:01:53.760 |
We got going about two and a half, almost three years ago now, just before the starting gun of Chad GPT. 00:02:01.040 |
As always, in technology, investing most of life, it's better to be lucky than right. 00:02:07.360 |
And the point of having a new venture firm, I worked at Greylock. 00:02:13.600 |
It's kind of a traditionalist venture firm, a great one. 00:02:16.080 |
My partner, Mike Vernal, used to work at Sequoia. 00:02:19.440 |
Was that we think, like, actually, you know, at risk of sounding like those people, this time it's 00:02:26.880 |
That this is the largest technology revolution that we get to be a part of. 00:02:30.640 |
And that there's so much change in the technology, the types of businesses you can build, the product 00:02:35.840 |
decisions you make, what challenges these startups and big companies face that, you know, maybe there's 00:02:45.040 |
And so, you know, I'm thrilled to be working with, like, really interesting people in the industry 00:02:50.960 |
Mike and I are investors in companies like Cursor, Cognition, Mistral, Thinking Machines, Harvey, 00:02:58.480 |
So a mix of base 10, like a mix of infrastructure, model, and application level companies. 00:03:05.520 |
And, you know, one more, are my kids coming up yet? 00:03:11.120 |
One more just observation from the last two and a half, three years of doing venture. 00:03:17.040 |
I was an investor for about 10 years before that, is I have never seen the, like, just the uptake 00:03:23.840 |
from users that has been possible in the last couple years. 00:03:30.880 |
You know, AI product and AI engineering, and this is kind of the theme of my talk, so I'm 00:03:37.040 |
sorry to give away the punchline, but it's quite a bit harder than people had hoped. 00:03:43.840 |
We see companies going from 0 to 10, 50, 100 million in run rate very, very quickly, 00:03:50.400 |
faster than we've ever seen in any technology revolution before. 00:03:53.200 |
And I get asked a lot, like, where are we in the AI hype cycle? 00:04:02.720 |
And I would say, having actually been an investor or an operator through a macro cycle at this point, 00:04:08.800 |
like, I try to pay very little attention to what the marketing world is saying, 00:04:15.280 |
Because, you know, if you're an operator or an investor, maybe you care about what the stock price 00:04:21.200 |
does every day, but really you want to figure out if the company you're working for or starting is 00:04:27.280 |
And if the products are going to work long-term. 00:04:29.040 |
And the things that I get most excited about are seeing, like, crazy usage numbers. 00:04:52.480 |
So, I want to talk really quickly about just a few things today. 00:04:57.360 |
I think we lost a little bit of time, but let's say let's talk about capabilities, 00:05:01.040 |
what we're seeing work in the market, and then maybe some advice on, like, 00:05:07.040 |
what to build, if those are, you know, a question you're considering. 00:05:10.320 |
I think the shorthand that we're going to use in this presentation is, like, cursor for X, right? 00:05:16.400 |
And I do think that's a really massive opportunity. 00:05:18.960 |
The first thing in capability for this past year is clearly reasoning. 00:05:23.520 |
Reasoning's a new vector for scaling intelligence with more compute. 00:05:27.280 |
The labs are really excited about this because they get to spend more money and get more output. 00:05:31.120 |
But we should also be really excited about this in terms of unlocking new capabilities, right? 00:05:37.600 |
If you just put aside how it works, it's a confidence-boosting implementation detail. 00:05:45.360 |
You're unlocking a new set of use cases, like transparent, high-stakes decisions 00:05:52.320 |
Sequential problems, problems where you need to do systematic search. 00:05:56.000 |
I think this looks like a lot of problems that we're excited about and face in knowledge work every day. 00:06:03.200 |
As you have just seen demos of and I'm sure are working on, given reasoning, people are really excited about agents. 00:06:09.680 |
To put a -- you know, I want to do, like, the Steve Ballmer impression that's, like, "Agents! Agents! Agents! Agents! Agents! Agents!" 00:06:17.760 |
But I -- you have to give me more than 12 minutes to, like, get that sweaty. 00:06:22.000 |
But, like, the non-marketing definition that I think of is it's software that -- 00:06:31.040 |
it takes some set of steps. It, like, plans. It includes AI. It takes ownership of a task. 00:06:37.680 |
And it can hold a goal in memory, you know, try different hypotheses, backtrack. It ranges from 00:06:42.800 |
super sophisticated to super simple. Some of the tools that you might use to accomplish a task include 00:06:48.320 |
other models or search. And largely, it's just, like, AI systems that do something. And that's not a chatbot 00:06:55.520 |
that looks more like a colleague. And, you know, one thing that I think we have a really unique vantage 00:07:00.880 |
point on is we back a small number of companies at Conviction, but we also run a grant program for AI 00:07:07.280 |
startups. It's called Embed. We get thousands of applications every year. And it includes, like, 00:07:11.600 |
user data and revenue data and, like, really amazing people. And the number of agent startups has gone 00:07:16.560 |
up 50 percent over the last year. And a lot of them are working. Like, we do see stuff that's working 00:07:21.520 |
in the real world. And that's super exciting. Other modalities are progressing, too. I'm sure a lot of 00:07:27.360 |
people are using voice, video, image generation, even beyond, you know, Studio Ghibli. But you have 00:07:34.480 |
companies like HeyGen and Eleven and Midjourney that are rocketing past 50 million of ARR. These are real 00:07:39.840 |
businesses now. I want to see if I can quickly play for you. 00:07:44.080 |
They told me to express myself. So I did. They told me to express myself. So I did. Now I'm banned 00:07:50.400 |
from three coffee shops. Hands can hurt or heal. That's the difference between chaos and creation. 00:07:56.000 |
So if you're wondering where Q3 is headed. So if you're wondering where Q3 is headed, 00:08:01.040 |
here's the thing. Consistency always beats urgency. We've got the projections ready. And let's just say, 00:08:07.680 |
it's looking solid. I would definitely recommend it to anyone. I would definitely recommend it to any. 00:08:12.720 |
So I think, like, if you just are looking for artifacts of improvement, this is from a company 00:08:18.160 |
called HeyGen. You can make clones of yourself, of fake people. And it's like you have gestures and 00:08:24.560 |
expressions that reflect emotion and content now, right? So these models work together. And like, 00:08:31.200 |
I don't know about you guys, but looking at that last gal, like, I feel influenced. I don't know what 00:08:34.800 |
the bunny is, but I would buy it. And so I think, like, huge swaths of the economy are going to be 00:08:40.160 |
affected by this sort of multimodality. Some investors or operators would say multimodality 00:08:46.000 |
would just be for niche verticals that enterprises don't have. You know, your average enterprise doesn't 00:08:50.960 |
have that much voice, video, image data today. But I think that changes, right? When you can do stuff with 00:08:56.400 |
this data when it is structured and understood, there's more reason to capture it. And I think of, like, 00:09:02.640 |
how much video do all of us watch every day? It's one of the highest bandwidth communication methods, 00:09:06.880 |
and we're just going to use more of it. We think voice is where we're going to see applications first 00:09:12.880 |
in business workflows, because it's already a very natural communication mode. So everything from 00:09:19.040 |
medical consults to lead generation, places you already had business voice, you just couldn't scale it 00:09:24.480 |
before. I think that's where we're going to see it first. But as these other modalities become 00:09:29.120 |
more controllable and also less costly, we should see all of them. I think it's safe to say you can 00:09:35.440 |
expect capability improvement in every part of the model layer, which is really exciting. A lot of people 00:09:40.800 |
are talking about the data wall or, like, the end of AI summer. But for anybody who's building applications, 00:09:47.360 |
I'm at least to tell you one person's opinion is not coming. And then, usefully for all of us, 00:09:55.760 |
that market for model capabilities is getting more competitive, not less. Sam Altman himself, I think, 00:10:03.920 |
said it best. Last year's model is a commodity, which is a scary thing for a model provider to say. 00:10:08.720 |
Because last year's model is now pretty damn good, right? The numbers tell the story. GPT-4 00:10:14.080 |
went from $30 per million tokens to $2 in about 18 months. The distilled versions of that are, like, 00:10:19.760 |
now 10 cents. So we can really use them very broadly. If you look at this chart, green is Google, 00:10:26.480 |
yellow is Anthropics. You see, you know, it's a real mix. This is data from Open Router. So thank you, 00:10:32.000 |
Open Router for that. But you really saw Claude cut into OpenAI's market share and Google come 00:10:38.880 |
roaring back with Gemini. This data is obviously a little biased because a lot of people just go 00:10:42.960 |
direct to OpenAI. But if you're into multi-model, there really is a mix. And you do have credible new 00:10:47.920 |
players like SSI and Thinking Machines, some of the best researchers in the business, with orthogonal 00:10:53.040 |
technical approaches entering the fray as well. And I'm sure many of you have experimented with DeepSeq, 00:10:59.760 |
coming out with releases of both base and reasoning models that are reasonably competitive with a 00:11:06.640 |
claimed fraction of the training cost. Like, we should just assume that open source will do as 00:11:11.280 |
open source does. And we can rely on the model market to compete for our business, which is really 00:11:15.760 |
exciting. And so the view is plan for a world that is multi-model. Tools like Open Router or inference 00:11:21.680 |
platforms like Base 10 help that, and I think, like, be comfortable with that. I am. Okay, so we have all 00:11:28.560 |
this capability. Let's shift quickly to the application layer. We have to start with Cursor. A million to 100 00:11:35.120 |
million of ARR in 12 months and half a million developers. I assume all of you. Zero sales people to 00:11:41.040 |
start. That's not growth. That is a killer application. Cognition, which started with more autonomy, is already the top 00:11:47.760 |
committer in many companies. Feeling a little threatened, but also excited because recruiting 00:11:52.000 |
is hard. And then Windsurf, who's on a tear itself and really beloved, is being acquired by OpenAI for 00:11:57.520 |
three billion dollars. So we know for sure that the labs don't think that they can just, you know, steamroll 00:12:03.600 |
everyone. Right? Lovable and Bolt hit 30 million of ARR each in a handful of weeks, helping non-engineers 00:12:12.480 |
vibe as well. So, you know, our ranks are expanding. And I think it's useful to just, like, analyze a little 00:12:19.440 |
bit why code was first. Fundamentally, it is text with its, like, logical language with structure. Right? So much 00:12:28.640 |
of coding is sophisticated boilerplate. Like, we all love engineering, but some of it is, like, craft work, 00:12:34.240 |
not new algorithm work. You don't need AGI to write, like, an API endpoint or a React component. 00:12:42.000 |
Second, you have deterministic validation. You can automatically check if code works. Run tests, 00:12:48.400 |
compile, execute, do things developers would do. And third, researchers believe code is crucial for AGI. 00:12:55.360 |
Right? So they poured resources into it. And code became a key benchmark and a training priority and 00:13:02.400 |
an area for data collection. But I think the last point is the money point to me. Engineers built tools 00:13:09.920 |
for engineers. They understood the workflow intimately, and that made all the difference. And that last part 00:13:14.800 |
is the playbook for every other industry. I'm sure people are building things that serve beyond engineers. 00:13:20.320 |
And I don't think the winners will just be AI experts learning those domains. They'll be customer-centric, 00:13:26.400 |
like, problem-centric builders who understand AI and then redesign workflows from first principles around 00:13:32.000 |
manipulating those models. And so I think that's really the opportunity to build Cursor for X. Let's think a 00:13:38.560 |
little bit about what that means. Cursor is not a single model. You know, one model's doing diffs, one's doing merge, one's embedding the files. 00:13:46.240 |
They manipulate and package up the context. They prompt the models very skillfully. They let engineers 00:13:53.440 |
avoid repetitive tasks and standardize with things like Cursor rules. And then if you're using Cursor 00:13:59.280 |
in a team or even yourself, regularly retrieval accuracy gets better the more you use it with coverage and 00:14:04.320 |
freshness. And so all of this happens in a UX that makes sense. Right? Like, I use VS Code. I'm familiar with it. My shortcuts work. 00:14:11.760 |
And they make it safe to say yes. Right? Like, green for add and red for subtract makes sense. I can 00:14:18.960 |
scroll through it. And it's fast enough that I don't get frustrated. So my view is Cursor, if it's a wrapper, 00:14:25.120 |
it's like a very nice, thick, perhaps 14 or 15 billion dollar wrapper. Right? It's like if your burrito was 80% 00:14:32.880 |
wrap and 20% fill, but you got to choose the fill and there's like an empty, like an open market for fill. 00:14:38.640 |
Right? And so where's the value now? It may not be in the protein. It's kind of in the company. 00:14:44.480 |
So, like, if we try to generalize that recipe a little bit, if you are building a generic text box, 00:14:53.120 |
like, unless you're just like learning to do this, please don't. Like, OpenAI already won that. 00:14:58.880 |
Or it's just not very valuable to do so. Your domain knowledge, your workflow knowledge can be the 00:15:03.360 |
bootstrap. If you already know what users in your industry need, don't make them explain it. Build 00:15:10.240 |
products that show up informed. They collect and package context automatically, including from other 00:15:14.960 |
sources, not just natural language. Present it to the models. Use the right models at the right time, 00:15:19.760 |
now known as orchestration, and present the outputs to the users thoughtfully. Right? So I do not think 00:15:25.760 |
this is the end of the GUI. I think you can capture and enable workflow with these models. And all this 00:15:30.880 |
requires taste and a ton of work. I'd argue that, like, some version of this recipe is much of the work 00:15:36.560 |
each of us is going to do. So don't listen to the labs from a user experience perspective. The prompt is a bug, 00:15:42.960 |
not a feature. I think it's like a stepping stone. Don't make me think as a user. The best AI products, 00:15:48.640 |
they feel like mind reading because they are. There's enormous headroom in building these products, 00:15:53.760 |
and I think that's really exciting because that's what most of us in this room have alpha on. 00:15:57.040 |
What is a software company if not a very thick workflow wrapper most of the time? That was true in 2015. 00:16:05.360 |
It's true in 2025. Besides code, where might you go apply this? We think the opportunities to build value 00:16:14.960 |
around the LLMs exist in every vertical and profession. But here's something counterintuitive. 00:16:21.040 |
Beyond coding, one of the things that I've been surprised by is that the most conservative, 00:16:25.600 |
low-tech industries seem to be adopting AI fastest. We call this the AI leapfrog effect internally. 00:16:31.040 |
These are three portfolio companies. They're working. Sierra resolves 70% of customer service queries 00:16:38.720 |
for their customers. They serve people that you guys use, like SiriusXM or ADT. Harvey is two years in, 00:16:46.800 |
well over 70 million of ARR. Its AI is essential now to being competitive in the legal industry. 00:16:52.960 |
There's a company called Open Evidence, which helps doctors stay up to date with medical research. 00:16:58.960 |
You have to be a clinician to use it, but you give it your medical ID number, and you can do intelligent 00:17:04.080 |
search against medical research at the point of clinical decision making. Today, it reaches a third 00:17:11.840 |
of doctors in the U.S. weekly, and the average user uses it daily, right? And so I think there's just 00:17:17.760 |
examples of huge value beyond ChatGPT. These are companies that know their customer and solving real 00:17:24.720 |
problems. As a piece of trivia that you may or may not know, Brett at Sierra is the chairman of the 00:17:30.800 |
board at OpenAI. OpenAI was Harvey's seed investor, and if these people are not fretting about thin 00:17:43.200 |
Okay. Finally, I'll make an observation. A lot of people are excited about full automation. Now I'm 00:17:48.400 |
sweaty enough, so agents, agents, agents, agents, agents, agents. But when we analyzed the applications 00:17:54.240 |
to embed, I said, you know, it's gone up to 50 percent, you know, doubling applications for agentic 00:18:01.040 |
startups in the last year. I think some people think copilots are yesterday's news. They want to get to 00:18:07.760 |
the endgame, right? Like, you know, your colleague and AGI. But in terms of what works, like the data 00:18:13.840 |
on what's driving revenue, I think copilots are still really underrated. We see a whole spectrum 00:18:19.120 |
of how much automation. And I think the Iron Man analogy is still really great here. Tony Stark's Iron 00:18:25.760 |
Man suit augments him, right? He can do all these amazing things, but could also fly around on command, 00:18:30.880 |
could do some basic tasks without Tony. And my experience with these companies has been that human 00:18:36.480 |
tolerance for failure or hallucinations or lack of reliability, it just reduces dramatically as latency 00:18:42.720 |
increases, right? So the path of least frustration today for many domains is to build great augmentation 00:18:48.960 |
and then just ride the wave of capability because we know it's coming. And so my advice for many domains 00:18:54.720 |
would think about, like, build the suit and you can extend out to the suit that flies on its own once 00:19:00.480 |
Tony or any of us is wearing it. I'm not going to go through each of these, mostly because I lost time, 00:19:08.000 |
but there are a ton of opportunities. We put requests for startups on our website. We're interested in a 00:19:13.840 |
couple different categories of things. They go from, like, just good fit for purpose, like the law is a 00:19:22.000 |
space of lots of text generation, right? To things that weren't possible before AI. My partner, Mike, 00:19:28.880 |
will say, like, this is a really interesting era of machines interrogating humans. What can you do if 00:19:34.080 |
you can go, like, collect data on demand from people? We could talk to every customer, not just the top 5% by 00:19:40.640 |
contract value. We could root cause every alert proactively, right? Versus, like, just firefight. 00:19:47.920 |
And the mental model is how can you build as if you had an army of compliant, infinitely patient 00:19:52.960 |
knowledge workers. You know, one aside here is I think there are many hard problems where, like, 00:20:01.840 |
the basic premise is the answer to them is not in common crawl, right? The reasoning around them is not in 00:20:06.720 |
common crawl. So this would be robotics, biology, material science, physics simulation. They require 00:20:14.160 |
clever data collection, probably interaction with atoms, not just bits. Super scary for a software 00:20:20.320 |
person, but I think the juice is worth the squeeze, right? The same reasoning that crushes math olympiads 00:20:25.520 |
can seemingly navigate molecular space, and I think there are some really fundamental questions for human 00:20:30.640 |
society that can be answered when people work on these problems. And it's really cool as a machine 00:20:35.520 |
learning person to meet people in their, at the top of their field at the intersection of machine 00:20:41.120 |
learning in all of these other areas because, like, you guys would also understand, the same architectures 00:20:45.840 |
apply, right? And that's just, that's really exciting. 00:20:53.040 |
How should we think about defensibility to this advance? 00:20:56.880 |
Okay. So one last point, and then I'll conclude here. Some would say stay out of the weight of the 00:21:05.520 |
labs, don't pick up pennies in front of the steamroller, right? But I would offer what I think is an 00:21:10.880 |
uncomfortable truth. Execution is the moat in AI, and that's available to all of us. Cursor arguably did not 00:21:18.080 |
invent code completion, they did not invent the model, they didn't invent their product surface 00:21:22.160 |
area, right? They just out-executed on every dimension of this. They shipped a great experience 00:21:27.360 |
faster than their competitors could copy, and they captured the hearts and minds of developers, at least 00:21:31.760 |
in this term. I don't mean this to be cruel, but I often get asked about, like, counter cases and the 00:21:38.560 |
importance of first-mover advantage. Let's be brutally honest. In contrast, like, Jasper had first-mover 00:21:43.920 |
advantage brand that raised $125 million, but its first product was a series of prompts in a text box 00:21:50.720 |
and, like, very good SEO, and, like, you have to keep running. Like, ChatGPT, you know, crushed the first 00:21:56.240 |
iteration pretty quickly. And so I don't think this is satisfying advice, but I think it is, like, real from 00:22:01.440 |
the trenches. Build something thick and stay ahead, and, like, no domains are out of question. Magical AI 00:22:07.520 |
experiences, they build customer trust and drive adoption. And a lot of the data we need to improve 00:22:13.520 |
these experiences and the context we need, it is not easily available today. And that advantage is, 00:22:20.000 |
you know, open for the taking and not for the labs. 00:22:23.520 |
So I guess in conclusion, I think the opportunity is early and really massive. Like, I've made a career bet 00:22:31.120 |
on it. I think many of you are. We're in the dial-up era of AI, and we're moving pretty quickly to 00:22:36.720 |
broadband. Instagram came four years after the iPhone. Like, I was there when Greylock made that 00:22:41.920 |
investment. Uber, five years. DoorDash, six, right? So the truly transformative companies, 00:22:47.280 |
they weren't necessarily the first people to recognize the changes or the opportunity. It was 00:22:53.200 |
those who reimagined the experiences. And the game board keeps getting shaken up. That's the thing that's 00:22:58.000 |
different this time, right? It's like getting a new iPhone. That's actually different every 12 months. 00:23:04.240 |
And so you have, like, new model release, new capability breakthrough, you know, one-tenth the cost. 00:23:09.840 |
And every time the game board turns, I think there are, like, there's an opportunity to to win again. 00:23:15.280 |
Okay, so I'll give you one last sentence and be chased off the stage. This was not my fault. 00:23:21.040 |
Here's what I really want you to remember. You, as the engineers, got the magic first. 00:23:26.240 |
The anthropic, like, economic index says that 40% of use was still coding. That's not, like, 00:23:32.000 |
40% of the economic opportunity in the world, right? And so it is the job of everyone in this room 00:23:37.520 |
and, you know, globally online to be the translators for the rest of the world. So I encourage you to