Hello, and welcome back to Equity TechCrunch's flagship podcast about the business of startups. I'm TechCrunch venture editor Julie Bort, in for Rebecca Balan while she's at Web Summit this week. TechCrunch readers have probably heard some form of agents are the future or agents are coming for our jobs. But today we're going to find out where one major investor in agents and infrastructure stands on the issue. Today I'm joined by Jennifer Lee, general partner, leading infrastructure investments at Andreessen Horowitz. Jennifer, welcome to the show. Hi, Julie. Thank you for having me on. Yeah, well, you've got some good news that just happened to your team. You guys recently raised 1.7 billion in new funds. So I guess the, I mean, the biggest thing I think people want to know is like, what are you going to spend it on? So, and I'm curious, what are you going to spend it on in 2026 this year that you probably wouldn't even have considered spending it on last year or the year before? It's truly a lucky time to be alive and we're literally in this super cycle that not just I've never seen that many of the industry veterans on our team have never seen and infrastructure is getting re-bued in real time. Every single layer, every step of the way. So as an infrared investor, it's certainly a very, very exciting time, which is the reason why we raised 1.7 billion fund to really back infrastructure founders to go from all the way on the bottom of chips, chips design, building the real hard infrastructure that supports sort of where models are going to run and the software layers, the communication layers, the developer tooling layer to all the way the model layer. And all of these are our mandate to really looking at looking to the future, looking at the workloads, looking at use cases to see where we need to retool. And the answer honestly is we need to retool pretty much everything because all the infrastructure that's AI running on today are not built for AI workloads. So give me a few samples of some of your portfolio companies and why their infrastructure. I work with a few companies such as Edelman Labs. Edulgram, these are on the foundation model side that they are developing their own model from pre-training to post-training to the product itself. And Edelman Labs is one of the voice players that is building both the creative platform for people to use synthetic voice or clone voice to using podcasts, YouTube, creative expressions, and also agent platform that are powering a lot of the voice agents today, be it for customer support use cases, for HR, for sales and marketing purposes. And Edelgram is an image generation company that's generating a graphic design, realistic photos, typography, also from pre-training to post-training to the consumer facing and professional facing products. These are on the foundation model layer and on the infrastructure layer, fall, FAL is a great example of being the inference cloud for all the diffusion models or I'd say multimedia creative models from video image to audio that is powering a ton of the current at scale, creative expressions, creative tools and marketing and advertising use cases. And there's a slew of DevTool companies like Snineless, Munlefy, Mox, Astro. I can talk about them probably for the next hour. I know all your babies. You don't want to leave anybody out. Exactly. I get it. But this is sort of below the application layer. Layer, right? So this is the stuff that's going to power the application level. Sounds like mostly. Some of them do have their own applications because of this working integrated notion of AI company. But yeah, largely these are infrastructure and APIs and capabilities that are powering application layer. So you're back in these companies that do voice, that do video, they do well. So what's your personal opinion on AI Slop? I mean, mine is it's actually some of the most entertaining stuff on the internet. I feel like. But also it's problematic. So I'm curious, like you've got companies in the trenches of that. So what what's your thoughts on it? And how do you think we're going to get the best out of it without having it become a real liability? Yeah. I'd say every technology go through its infancy phase, go through its adolescence phase and to maturity. And I am in all of the technology evolution in all of these models. Like three years ago, we're looking at a gender image. We can distinctly tell this is a gender image, whether it's the hands or the fingers or the eyes, the inconsistency, the lighting, the shadow. It's very clear that that's a fake or AI gender image. Only six months to a year later, you really cannot tell anymore. The lighting is perfect. It's hard. Yeah, the facial expression is perfect. And we really across the uncanny valley really quickly with image and same for audio. Like honestly, I was astonished when at first heard 11 laps clone of my own voice. And it's a little uncomfortable. I have to say that to hear your voice. Which do they have you say? I was just cloning in there reading a blog, a blog post. It was not the most like glamorous thing. But at the same time, I was like, oh my gosh, I can turn this into another language. I can really like hear my voice. And I recently did a clone of myself speaking Japanese. I showed it to my husband who speaks Japanese. And he's like, wow, it's quite surprising to hear my wife who doesn't speak Japanese, speak Japanese to him. And it's just like, you know, it definitely feels like living in the future. But again, these these technology just evolves into like crossing the uncanny valley phase really quickly. I don't think we're there yet for video. That's why we probably see a lot of sloths on on the social media today. But like you said, it's entertaining. And we already are seeing the progress really was the Grox recent video model that these sort of imperfection is getting close to high quality, high speed, really hard to tell that it's generated pretty quickly. And there's of course good and bad with that too. It's just, you know, the quality side and sort of the sloths that we're as users were complaining are quickly getting removed. I mean, we like knowing that it's AI generated, you know, like a cat jumping on a trampoline. We like knowing that that's AI generated, right? I still love to be able to tell us some of the breadcrumbs. I don't want to see in the video that like, you know, there are these little imperfections to know that it's generated. Yeah, I agree. But I'm sure for professionals, that's like a complete note now. So what are you seeing in 2026? That's sort of new already. I mean, 2025 was a crazy year. We went from, oh, the models are good. We could maybe use them to companies raising Buku amounts of money and attacking every layer of the infrastructure with AI. So I'm curious, like, as we start 2026, what do you think is new right now that we haven't ever seen before? I am sure you have seen the moat, moat bot now in that clop bot, proliferating through Twitter. I think that's a great sign of just how people start to realizing agent can be real and they can actually help you with personal productivity for like, you know, longer running processes. And this is probably all set of the moat sort of use cases. But that I'd say, it's not new as like a concept who I've been talking about agents for a couple of years. But it's finally at the stage where we are able to yield real ROI and productivity out of some of the long running agents and agent take workflows. That I think is a huge unlock. So do you think this year agents are going to come into the hands of the average knowledge worker? I definitely think so. With clop skills, with open source like moat bot, we'll certainly see a ton of personal adoption. Do you use any agents? What do you use them for? I definitely use research agents to help me prepare my day. Get to know the topics and also pull up, know what's that information from from my own drive and so on. I use a ton of voice agents just to try out and test out capabilities and sort of advancements in the fluidity of conversations. So certainly I am setting up a few days aside to really build my own productivity agent to help me arrange my calendar and also blocking out time for thinking. So these are really amazing tools and also very easy to get hands on with coding agents to build sort of your own productivity assistant. So your first knowledge worker agent is going to be managing your calendar, right? Yes. That's what you're thinking. In addition to trying all the fun voice stuff and speaking Japanese. So what are you using to build that and where are you with it today that you think is going to be quickly solved? Well, at Curson Clockhold for sure. This steel is app and I would love to have a nice UI and very good integrations with all the systems and tools I'm using. At the same time, I just really think the limitation today is time and attention. The reason why the first app is a calendar scheduling one is to really open up time to keep up with all the trends, keep up with all the research that's happening with AI, with use cases that people are taking these products and tools for. I am spending hours in just reading X or Twitter about how people are using MOBOT. This is the reason why I love investing in developer tools and infrastructure is once you give people creative tools and a toolbox. They can literally open up your imagination that you haven't even designed the use cases that you're thinking of or you're not designing for them but people are just going to take it and then run with it and then creating all these different and it's super interesting and an imaginable agent take use cases and workflows. So that's what I'm trying to set aside my time for and also what I'm trying to use agent for. I love that idea. Like as a journalist, I'm bombarded by email. You have no idea. I could check. There are 24,000 unread emails in my inbox and I literally could not even just managing that is like, was every six months I have to go through and clean it up and I would love an agent that did that for me but I feel like we're not there yet. Like the amount of effort I would have to do to train it and also trust it. I'd have to trust it because there are definitely things that come in my inbox that are important that I need to jump on but they're outliers. They're not in the bucket right and I don't trust that AI is designed for that to see the one thing the human mind is going to see and be like, oh no, I need to deal with this right now. So that's why I think I wonder whether 2026 is going to come to an average knowledge worker. I use AI all around the periphery but to actually hand it a task and say go do. I don't even know how an average knowledge worker would train their own AI for that. So what do you think is going to happen there? The example give is a great one on emails and parties. Like humans are just so great. I connect him thoughts and figure out and spoken context versus these LLMs. You already need to feed all the like have a clear go and purpose feed all the context in these and understand the things that are happened. Maybe not spoken in the email itself or on the calendar that are able to gather these information and bring sort of what should be the top party for you. That's why I think a lot of the email inbox agents are not quite there yet but there are certainly areas it can help with. When you're really trying to understand a topic, I'm sure you're getting a ton of inbound about carving certain things and that's where I at least find the most part of the lift is whenever I want to really figure out what has been talked about on the topic. Doing that kind of research which used to take multiple links and go into Google search and open up a bunch of tabs and coming up with a form the opinion of if this is top of mind, if what people have already talked about. Now you literally have either finger tips like I just listened to a deep research output and put a 11 laps reader and just listen on my way to a meeting or on my way to work. That experience has to take hours from you before where we can just do that on the go. I agree and maybe it's going to seep in a T-bag. It's a much better Google. It's a much better Google experience but it hasn't actually become something I can hand a task off to do. For sure. I write all my own stories. I have been accused of sounding like an open AI. I think open AI trained on me personally. I love a good ham dash. Which I think is a good thing that the actual creation still happens with human but these tools really help us to be a lot more productive and a lot more intentional and also a lot more expensive with our reviews. Every single time I'm again going to a meeting coming into a podcast is where I just spend hours to do deep research and just learning about things that has been explored and talked before or things that haven't been. So I think that's where the productivity again is happening but also the trust is being built. This is not a time I feel like we'll just be ready to hand over agents a ton of things. I'm fairly certain 2026 or maybe even 2097. It's still very much a co-pilot phase. And some of the pieces might go into auto pilot. If this really is about data entry, like IMS and Company Co-Reducto that turns PDFs and documents into structured data. You really don't want people to keep entering data. That's like looking at human region form. That's probably the job and the role that should go away first that turn these humans into something way more intelligent type of work than just sitting in front of a computer doing data entry from these overseas services providers. So those are the things I think will go out to the more autonomous versions first in 2026. So the average knowledge worker is going to be able to hand off boring tasks like that. I still have a lot of those tasks that I can't get AI to do for me. Collecting information and putting it in a different format. It's still pretty wonky. You know what I mean? Stuff like that. And so maybe 2026 it'll come along. I do have a beef though with Silicon Valley and even VCs to some extent that keep talking about agents as human replacements like full human job replacements. I think in my opinion from having experienced and played with these tools just as the year, you know, the couple years as this come out that that we're going to hand off undesirable tasks to agents. But I think that this whole oh, you're never going to have a salesperson again or things like that. I feel like that's just marketing so that these software companies can price their agents comparing them to labor. You know, okay, you can pay us $15,000 a year because you'd have to pay someone $100,000 a year. It's not really replacing a human. And so, okay, am I right or am I wrong? What do you think? Are these agents really going to replace a human? And are they going to do that in the next year or two? I think it really depends on the job and the role. It goes back to what I just said. It's like, you know, if this is a job really not viewed for human beings such as, you know, sitting from a computer doing theater entry or like being on the phone just talking about the same thing again and again, like explaining the same concepts or like answering the questions about where is my order? Like, how do I return this? Like, these are just inherently very cumbersome mundane and so crushing tasks that I wish, you know, agents can replace and elevate these humans into more intelligent type of work where there is still a large amount of knowledge work that really requires human interaction and human creativity to involve. And that's probably at, say, the last 20 percent of let's say customer service or like sales SDR BDR, like you still want to interact with a warm human voice. The understands sort of the human intent and also solve sort of a complex issue, build that relationship, build sort of the connection. So it's not going to be end on be all or like a binary answer. I think again, it really depends on the type of work. I feel like we're in agreement. It's they're going to replace tasks, but not human jobs in my opinion. Sure, there's still a few hangout human jobs like data entry from that hail back from the 1960s or whatever. But I mean, technology is always automated, those kinds of things. Correct. Like back in the day, remember the receptionist, you know, like companies don't have that anymore. And that was long before AI agents came in, right? Certainly. So I think tasks, not jobs, and that all of this jobs are going away. Talk is something Silicon Valley needs to cut out. It needs to stop doing. What do you think? You like them pitching that? I'd say certain jobs probably still go away and the job titles will change. Okay. Fair enough. Fair enough. Fair enough. I think we're mostly in agreement on this, but I don't know. So what is your most unhinged opinion? What is the thing that when you go to a cocktail party, you're like, this is my opinion. I'm standing to it and people are want to tell you you're wrong. I know these models are really creative. They're able to create in lightning speed in seconds, but I inherently still think creativity belongs to humans. And the best ideas will eventually still come coming out of human. So your unhinged opinion is that you're I'm putting words in your mouth. You're a little skeptical about AGI. So it's like, I think our AGI definition is reasonable to be skeptical about this thing that everyone's imagining. Yes. Yes. I think in the in the best form of AGI is where every single human are able to express our creativity and its maximum. And that's what I hope is what AGI brings us is, you know, there's a lot of intelligence that really helps us take care of the things. Either we don't like to do like it is not inherently pushing limits on the individual boundaries and really help each single human being expand our aperture domains knowledge and really reach the maximum creativity because I do think a lot of our day to day like we really just are in this crunch of getting things done. Like tell me how many times you feel you're creative and able to create something in the last year. Probably, you know, it's countable. It's not fair to ask me as I write for a living. So that's true. That's true. All creative moments. Yeah. But I'd even say like, if you have help to go through those 2400 emails every day, yes, you'll be able to create way more and probably be able to create more. I'll also push further of your best ideas. I think so. My unhinged opinion is I do think that LLMs have a limit. You know, I think they're interesting now. But I think that world models, which is where the in my un unscientific opinion is where the AI starts interacting with the real world. I feel like that's going to get us to this imagined future where you have robots, you know, picking apples and not humans anymore. That kind of thing. I think that LLMs are fascinating, but I don't think they're the be all end all. So would you agree with that? Do you think that LLMs are going to hit a limit? I also think it's a pretty agreed upon sentiment at this point that LLMs are good for certain things, but not going to be the answer for everything. I don't think they're going to turn into AGI. I do not think LLMs are going to turn into AGI. Yes. We'll need multi-modality. We'll need ways to interact with the real world. We can't just live in sort of token generation and next token prediction. So yeah, very much in agreement with that. Oh, see that's good. So the last VCI talk to you did not agree. Still just thought we were in the early stages of LLMs. And we are, we are, but I think we're starting to see their limits. I do. The other really interesting thing that's happening in 2026 is chip design. So this is like right up your alley. So we're now seeing AI, while we're seeing the early stages of a few startups that are using AI for chip design. I mean, the goal is like the AI is going to decide its own substrate level, you know? And iterate faster. Are you seeing that happening at all? Is it all just an idea for a few labs? I personally haven't spent a ton of time on this topic, but I've been just keeping up with the literature and the research our team has done. So I think this goes back to what I talked first that a lot of the AI workloads today are not running on the infrastructure that's built for it. And we're really pushing towards a future where there will be more purposeful mute workload specific hardware and software. That's either for just, you know, large scale inferencing fast speed latency. So like a lot of, again, on the chips level that they were seeing people are spending time researching what's needed for, again, pushing limits on getting performance out of these chips. So if you think about where, you know, LLM's and AI can really help with is this is like a waterfall takes years to design a chip but getting sort of into prototyping and to production. Like now we can really shorten that first phase of designing and prototyping. And this not just applies to chips. I'd say like this really applies to all software creation today is that shortened period of ideation and creation for the initial phase. So it's definitely coming to chip design. So one of the other things that's unusual about this particular stage and investing in AI is how quickly these companies are going from, you know, we just we're founded to whatever 100 million ARR, you know, in a tweet. So I'm just curious if you have any stories or examples of some of your fastest growing startups. And what, what is that? You know, for founders, what does that mean when you're going to get to these big numbers so much more quickly than has historically ever happened? I'll first caution the listeners and the viewers that not all AR are created equal and all growth are equal either. Like now we're literally talking about ARR revenue, run rate, GNB in one concept, especially on Twitter. And there's a lot of sort of missing nuances of the business quality sort of retention and durability that's missing in that conversation. So this is actually one part that's introducing a lot of anxiety to some like New York young founders is like, how do I do? How do I go from 0 to 100? I'm like, you don't, sure, it's a great aspiration, but you don't have to build a business that way to only optimize for the top-line growth. Like you can steal build a system of business that's growing 5, 2, 10 x year-over-year, which is again, unheard of. And also don't believe everything you read on the internet. Exactly. Exactly. They're using this number, but how they're constructing that number. There's a lot of variables. Totally, totally. So coming back to your question of like the businesses that we have seen that have reached those milestones, Cursor, 11 Labs fall, have all gone from 0 to 100, 200 million in a couple years or as short as, you know, a year. And these are, there's real reasons behind each of them is coding agents is actually the first agent that works in real life and producing really incredible results. And that's Cursor. For 11 Labs, voice agents plus synthetic voice has a ton of use cases that are lower hanging fruits, less prone to errors. And also again, talking about the crossing and county valley being one of the first modalities to do that. Like really jovy 11 Labs growth from 0 to I think last reported was 300 million by end of last year. And thought where they're really powering all the creative workloads from image to video. And we're really seeing the proliferation of open source where LLNs are really controlled by a few big labs and the sort of models on the image video side. There is a ton of demand smaller models. Yes. Yes. Also demand for variability. Like you don't want your design to all look the same. Like you kind of want different styles characteristics, but still have the consistent characteristics of low latency, high throughput. So that's a reason driving false growth in the past two years. So I'd say these are quite durable business models and reasons why these companies grow already fast. Okay, fair, fair. So not everyone grows that. Be careful with the number, believing the number. But what does it mean for a founder? I mean, normally it would have taken years and then you would have built up the structure to support that kind of revenue or those that many customers or whatever it's looking like for the business. They don't have that time. They're all of a sudden. They're just managing this. It may possibly sometimes without CFOs, without the financial boundaries that a CFO brings. So can you give me some examples of the kinds of things that start up, they're growing this fast face that you've never seen before. Yeah. I can tell you the hardest thing today, which again, shouldn't be a surprise, but it's just consistently being the cutoff big is like how do we hire not fast, but the right people that can really jump into this type of speed and culture, like these companies again, want to reach those milestones or all end their 100 people. And like you said, there's no like, you know, CFO or like CRO, like a lot of the growth happened with the initial teams and people stepping up. But then when we're thinking of, you know, how to grow the team further, like hiring best people that can move at the AI speed, they're just very limited talent out there right now, which is crazy to say given like, you know, there's this other side of fear of people losing jobs, like I've never seen hiring market like this in all of these AI companies that there's such shortage of talent. And they're just, you know, a lot of desire to find people who are able to move fast and also at the same time, there's also a ton of like unknown and open topics and problems that we have never faced before, like a lot of the deep fake how to counter that and legal and compliance requirement, like these are all the things that are in the new territories that we're consistently figuring out as the companies. So like those really requires a ton of creative thinking, like, you know, being also composed measure, also think about how, you know, the public audience and global audience will react to certain actions. So like those are always very nuanced topics that I've been having with these companies. They're not used to that. Suddenly one wrong word and, you know, people are up in arms or a mistake, cursor had a sort of a mistake in how they rolled out new pricing, right, because people to cope in arms, you know, the young company stuff that happens only they're not a tiny company anymore. I think it's hard. Yeah. Think about like a developer tool, IDE change pricing, becoming a huge deal three years ago. Right. Now it's a big deal. All right. So finally, what's one kind of start start up that you are on the search for that you'd you'd really love to fund with this 1.7 billion. Your team is it not right now. But when start up, you hope to find this year. It goes back to the word you just said search. I think there's still a ton of potential in search be a web search. Interesting. Yeah. Be a personalization type of search. I think LLM's needs tools. It really needs most up to date, accurate information. And that's a search problem. And the search infrastructure has evolved quite a bit, but still needs further involvement to make the search more personal, make the search also more expensive, faster to really keep up with the high throughput, high frequency, agentic search. And also accuracy too. Now we have further demand of what we're getting from the language models. We can't allow a single entry in the search results being wrong. Right. Like the hallucination problem or like, you know, the out of context that we're defining for this search query. Like we just have higher requirement for what we're hoping for from LLM's. And that's again coming back to a search problem. And I'm still looking for great team spewing in that. Okay. So someone handling LLM search with accuracy. If that's your startup. Come talk to me. Where would someone, how would listeners read you? They can reach me with DM on Twitter, Jennifer H. L.I. They can also email me. It's
[email protected]. Well, great. Thank you for talking to me today. And I think you're going to have a fun year. You've got quite the little pot to find startups and offer great terms for people. Thank you so much Julie. This has been a lot of fun. Equity is hosted by TechCrunch Senior Reporters and produced by Teresa Locansolo with Editing by KELP. Subscribe wherever you get your podcasts and see what's next at techcrunch.com/events. Thanks so much for listening and we'll talk to you next time.
PodcastTranscript AI