$688 Million AUM in 2 Years Investing in Data, AI, Web3 & Blockchain | Tomasz Tunguz

0:00Welcome to Sorcery. I'm your host, Molly O'Shea. Today we have Tomashtongas, a prominent venture capitalist and the founder and GP of Theory Ventures, a 6888 million dollar Aum fund. Theory focuses on investing$1 to$25 million in early-stage companies that leverage technology discontinuities to go-to-market advantages. Before Theory, he was a managing director at Redpoint Ventures, where he led investments into companies such as Looker, Expensify, Monte Carlo, Dune Analytics, and Customer. Throughout his career, Tungas has been recognized for his data-driven approach to investing and his very popular personal blog, TomTungas.com.

0:45Tomas recently raised a$450 million fund in November of 2024 for Theory Ventures' second fund, which marks a significant milestone for the firm. This follows the initial breakout fund, which launched in 2023 with$238 million. This is a very fun conversation where we go deep into everything AI and LLM business models. I hope you enjoy. Tomas Tungas, thank you so much for joining us. This is a conversation we've been wanting to have for quite some time. Pleasure is mine. Great to be here with you today, Molly. Thanks. To kick things off, where will we see the biggest shift for AI transformation in 2025?

1:28Well, I think we have to start with Jensen's quote from CES, where he said the IT department of the future is the HR department of today. In other words, the IT team will be managing vast numbers of agents. And I don't think we're there yet. The accuracy of the models, this is not quite there yet. The reasoning, the planning, the reinforcement learning is not quite there, but I do think the year, I mean, 2025, everyone's been talking about agents. I think you'll start to see agents that actually work pretty well. And that I think is challenging. I think it's challenging as an investor. It's challenging as an operator because it means reinventing all of your workflows.

2:05And that means learning to do your job in a completely different way. So where exactly are we in terms of AI adoption? There's been lots of profile funding rounds over the last few years. Are companies actually delivering any value? Where are we at? I think one of the biggest barriers to adoption to AIs that is really expensive. The ROI in most cases is not there. But like you, so let's start with the consumer. Let's organize it. In the consumer world, I think AI adoption is taking off in a really big way. You look at like 18 to 24 year olds, something, the Verge ran and surveyed 75 % of those people default to generative search as a place to go.

2:45So that's already happening in a really meaningful way. I think you have like OpenAI announced 300 million, I think, MAU on ChatGPT. So you're kind of approaching the billion user number in consumer AI. I think we'll definitely surpass that this year. And that's awesome, right? It means that Google search model is totally up for grabs, the ad dollars there, complete destabilization of the SEO, SEM market. That's really exciting. A lot market cap is now becoming loose. Within the enterprise, I'd say it's earlier still because the use cases within the enterprise are more challenging. Why is that?

3:23Well, the range of acceptable outcomes within the enterprise are much narrower, right? I remember I generated this image of a kitty cat on a fire truck and it was a Tonka truck. It wasn't a fire truck and the kitty cat was not it was a little you know five toes or whatever it was and that's okay for a consumer use case maybe i'm making somebody a funny happy birthday card but it was in the enterprise i really need to be a cat and i really need to be a fire truck and it has to have this brand on it so the range of acceptable outcomes is much narrower and we're starting to understand how to control those outcomes so you have things like open ai structured outputs which really kind of narrow the output JSON or the machine code that the computer or the AI system will produce and that will help.

4:07And we have these new architectures that handle the error, right? So just taking a step back, I mean, these AI systems have two properties that are hard to work with. The first is they're chaotic, which means they're a little bit unpredictable, and then non-deterministic, which means if you ask it the same question twice, you will not get the same answer. And that's very, very different in classic deterministic programming where you write a function add two numbers every single time you put it in one and one it will see two so what we're trying to do around like okay let's assume we have an ai brain inside of a program and we know that sometimes the answer is not quite right how do we manage that error there are lots of different techniques so there's uh we can talk about those techniques if you're interested but there are lots of different techniques and i think in 2025 going back to your first question we'll make some pretty material advances there where the predictability is much better.

4:59And that will lead into, I think, 2026, where we'll start to see pretty significant ROI coming from these agents. Let's talk about the techniques. Okay, so if you have, let's say you have a system that's 80 % accurate, right? If you have one step, it's 80 % accurate. If you have two steps, you have two times, you have two times 20 % inaccuracy, and three times through, let's just run the calculation, 0.8 raised to the third. That's the math that I'm doing right now, so I can't do it in my head. You have a system that's only accurate 49 % of the time. So you have this explosive error through multiple step processes.

5:38So how do you manage that error? Well, one technique is you set up an AI system that's then judged by a non-AI system. A non-deterministic system is then evaluated by a deterministic system, like a bunch of rules. Another system is you have two AIs, one that's judging the output of the other AI. A third system is what happens in Mechanical Turk. You ask three different AIs, maybe Cloud Gemini and OpenAI, to perform the same task. That's two out of three wits. That's state of the art. That's where we are. And we're trying to figure out, okay, well, for this particular use case, this is the right thing.

6:12Sorry, the fourth one is chain of thought, which is you ask an AI model, please be, gosh, what is the word? There's a word that's now like, you'll come back to me but anyway if you tell the model to be introspective and explain to you how it's thinking and pause before it thinks like you might tell a child it actually performs significantly better it will write out all the different steps and then it will give you a much better answer particularly for more complex tasks so those are four of the different techniques that researchers are playing around with and we'll see which ones ultimately work in which use cases but it's probably a combination.

6:51I know I should have these memorized, but I'm going to run through the hyperscalers names and some more. So we have Google, Microsoft, Meta, Amazon, Apple. We now have XAI, OpenAI, Anthropic. As we move towards a more exhaustive compute environment with inference booting up, advanced agentic ecosystems, where do you see the power dynamics between hyperscalers and these independent LLMs? I don't know if it was a pun or not, but the word power is incredibly important. So electricity is absolutely critical. I was talking to a friend of mine who spends a lot of time and energy, and he was saying, I was asking him, compared to an electric car, how much does a GPU consume?

7:37He said about 100X. So if you're building a data center, if you're building a power plant for a bunch of electric cars, typically around like five megawatts, if you're building a power plant for significant data centers, like five to 30 gigawatts. And so the major constraint, one of the major constraints today is just energy. So you see the reinvigoration of like Microsoft tried to start Three Mile Island again, the pressure on the nuclear regulatory agency in the United States to approve modular nuclear reactors. And so I think that that's critical. Just truly the energy will be absolutely essential.

8:13The next is just the CapEx. I mean, you have like Microsoft and Google and Meta each deploying something like 70 to 80 billion in capital expenditure, just buying GPUs and building data centers in just next year. And they were, I think, you know, 60, 50 and 45 this year so far. So amassing those GPUs. Microsoft is still compute constrained and they project, I think, all the way through 2025, they'll still be needing to buy GPUs aggressively because it just cannot serve service all of the demand. So whoever has the best allocation with NVIDIA, at least in the short term, is in a much better place.

8:49And then you have the models, right? The research that's happening there. There, I think it's anybody's game, but I would not have expected to be telling you here 2025 that Google has really been as aggressive and successful as they have been iterating through all of their different models like the Gemini Deep Research model where I can ask it to summarize. I was working on a project looking at the next generation ad networks and I asked it to summarize all the articles with startups on adweek.com over the last 12 months and provide a whole bunch of information. And it took 20 minutes, but it did it.

9:28Love what you're hearing on Sorcery VC? Unlock even more. With passes, you'll find exclusive bonus content from this podcast. Deep dives, behind-the-scene insights, and extra resources tailored for founders, investors, and startup enthusiasts. Head to passes.com to access everything you need to take your knowledge to the next level. Well, speaking of Jensen, Jensen predicts AI inference will be a billion times larger. This is, you know, setting the stakes, I'd say, low at this point. It's going to get really big pretty fast. how will this impact opportunities on the early stage side? How do you inform or how does this inform your thesis?

10:10Yeah, great question. I mean, I think you've seen a thousand X reduction in overall inference costs in the last three years. I think we'll probably see another thousand X reduction. There's a paradox called Jevons paradox that talks about how when the cost to do something decreases in technology, the demand actually far outstrips it. And so I think he's right. the reasons for this is if it's really cheap to run inference, well, you'll just do it all the time. And you won't feel guilty the way that I do when I spin up a big GPU job. But I think what it means is, and we're trying to figure out, how do you program computers in the future where you can infer at runtime and you can decide, okay, here's an input from a user.

10:54Do I use a deterministic system to interpret it or do I use a non-deterministic system to interpret it? Or when I'm a programmer, what part of my application is AI, what part of it is not? Today, it's really expensive. We have about 200 buyers that we call theorists around the firm. We asked them, what is the biggest adoption blocker to AI? About a year ago, it was security, but today it's cost. The office of the CFO is really pressing on AI leaders within their organization to show ROI and there's we talked about it's hard but it's also expensive. There's a, I was just looking at the analysis between the smallest models which are about four or two to four billion parameters and the very largest models which are about 450 billion parameters, you have a 100x difference in inference costs.

11:44There's a big push now to make the smaller models much more accurate which they are like the 5.4 model from Microsoft and the Gemma models from Google are incredibly small, incredibly potent, and it's just a year basically of that research. I think we'll start to see, particularly in the enterprise, small language models dominate. So what does that mean? I think overall there's one debate which is will enterprises buy off-the-shelf software or not? and if you need a custom crm are you hiring consultant are you not that that question nobody knows the answer to um i think the next question is will every company really become an ai company if a relatively modestly educated engineer can build software using ai you know why wouldn't a bakery uh also deploy it so those are some of the bigger questions we're trying to You mentioned doing some research on the advertising model side.

12:43This may be a little tangential, but how do you feel about the signal of perplexities move into the ad monetization space for their business model? And can they and the likes of Anthropic compete on the revenue side against the revenue and share market share of open AI and hyperscalers? Yeah, it's a good question. I mean, Google has dominated search, right? I think 90 % plus market share in the US. I think it's like 99 % in Europe and elsewhere. And the search ads part of the business is by far the most profitable because they don't have to split the revenue with anyone like they work with content ads or AdSense or Publisher Network, I think they call it now.

13:22But now you have hundreds of millions of people using Perplexity and OpenAI and Claude for search. And so it completely upends the ad model. I mean, what is search engine optimization, excuse me, in a world of AI? Nobody knows yet because the, you know, what is the right answer? Is it consistent? Is it not? What is the role of a marketer influencing the answer? How do you surface that to a user in a way that maintains the credibility of the overall result? Those are questions that we're really sort of thinking through. There's 250 billion in search engine marketing spend every year and so all of that is now totally up for grabs.

14:05There's another dynamic at play which is many of the younger generations are now starting with video as the dominant form of search rather than text. And so there's a shift like social media advertising for the first time ever surpassed search engine or SEM in 2024, roughly 250 billion each. So there might be a generational shift there. The ad models are totally unclear. I mean, working with AdWords, the way that I would build an ad for theory would be to create a list of keywords that might be relevant that people might search for. That doesn't really exist in the world of generative AI. One idea we've been kind of toying around with is kind of advertisers inject information into the context window, which is like the RAM, effectively, of a machine learning system and bid on those keywords.

14:55But it's really exciting and totally unpredictable where it'll end up. Yeah, we've been researching into that a little bit. And it seems like even the feeds going to the feed side, whether it's AdSense, they're constantly changing and they're reorienting them and making it really harder to push through and monetize like it was before. And the arbitrage business of advertising is just constricting. And you're seeing that with some of the public advertising players and where their revenue is going. Yeah. So we're talking about hundreds of billions of dollars changing hands and might be changing share.

15:29Yeah. It's going to be wild to watch, but we'll see what happens. So shifting over to AGI, because we all love AGI. I just think this is like a silly question to always ask. But if we are one to three years away from something like AGI, which could plausibly replace human labor for some set of tasks. And these tasks can range from like healthcare to, you know, operating a bakery shop or, you know, maybe some more like computer work. Do you think this is a Jevons paradox situation where efficiency gains lead to more consumption? Where are you thinking? Yeah, I think it's really exciting. I think, I mean, one of the interesting questions is like, what is AGI, right?

16:18It was like, is it talking to an AI that's just as sophisticated as a high school student? In which case, most models are 80 % as smart as a high school student. The open AI newest model has PhD level knowledge on many different topics, including math and programming. And so there's like the knowledge part, you know, many different kinds of intelligence. Knowledge retrieval is one. And then there's the reasoning part. I think once we get to a place where the AI is no longer like an intern but is more like an employee, then we'll start to see some pretty significant productivity gains. And one of the things that's missing there is long-term memory, right?

17:00Molly, you and I are talking and I learned something about you. Like, I don't know, your favorite kind of food is Italian and you really like penne. I'm going to remember that because it's memorable. And an AI system will forget like your goldfish's memory. And so we're working on that kind of long-term memory, which will be absolutely critical to having sort of longer term workers. But I don't think, I think the technology will be here much sooner than anyone can take advantage of it because all of us have to look at the way that we work, the way that we answer emails or make our outbound sales calls or put together presentations and learn again in a completely different way how to put them together.

17:37Like I was, I delivered a presentation last week and tried to do it in a completely new way. I went to one of the AI systems. I said, I need to put together a presentation in two days or whatever it was. And it needs to be 15 slides and here's the context. And I want you to write it in the format that my management coach taught me. And like here are a bunch of different bullet points. And then we iterated, iterated. And I said, okay, now it's just a bunch of images. And then I took all of those and I send it to Midjourney. And Midjourney is an AI that produces images. And so for each slide, the Midjourney produced the image from the content that the AI had produced from my ideas.

18:15And then we put it all together. And so the time to make that presentation was less than two hours where it might've been 20 or 30 before. And it came out great because the script that the AI produced that we collaborated on had a great hook and the transitions between the slides were brilliantly written and and then because we deliver it virtually i can you know i can basically read the script and so that's a completely different way than i would have built a presentation 10 years ago so there's one thing for the technology itself to reach a point where it works and then there's another for all of us to learn.

18:54This is how you, this is how you build a presentation in 2025. I know we're going to get to this question in a second, but have you ever done demos? I know we're going to go into your blog, but you are like a master product tester. And so even just explaining that out, have you ever demoed something like that? A little bit, a little bit. Uh, yeah, I think a lot of people have asked. I mean, I use dictation all the time. And then we've been playing around with lots of different kinds of software, but I think it's just really important to, it's really important to feel the technology and understand like how it works, right?

19:32Like different kinds of hammers feel differently in your hand and they're good at different things. But you're right. I probably should start like every Friday. This is my cool experiment of this week with AI. I'm sure your fans would love that. You have, you have many fans. get you on passes and get a subscription going. We're going to talk about your fund because this is really exceptional. You were able to raise two very large funds within just a matter of a couple of years. So your first fund, you raised in 2023. And then your second fund you announced in November, and this was a$450 million fund two.

20:15Because these two funds were raised in such a short period of time, how does that affect your investing period per fund? I can't imagine that you fully deployed fund one already. How does this work? We haven't deployed fund one fully. It means that we have commitments from investors that when we will start to raise fund two, that the money will be there. and I think it's I'm really grateful for it I feel to be just very lucky that there are lots of institutional investors who believe in what we're building and especially during a time both times when the fundraising market was extremely challenging because there's not a lot of liquidity that was being produced in the exit markets right and so as a part of that how big is your team to date and what is what is the actual construction of theory ventures look like today we are seven marvelous people small but mighty team and we have two great interns who just started a couple weeks ago um and uh there are three of us on the investing team and then the remainder work on what we call the intelligence team which is a research part of the firm and uh we build small very concentrated portfolios about 15 companies per fund and every Monday we get together and talk about updates on different research ideas so anyone can come in with an idea we had an intern last summer who said I think GPUs are not the architecture for AI they'll be ASICs or application specific integrated circuits and he wrote a 10-page paper on if you believe that premise here are all the downstream consequences for AI.

22:00And so we debated that and it makes for investing to be a lot of fun because people can be creative. They can read something on a Sunday or have a conversation with somebody on a Friday that keystarts an idea. Then we staff that research project and create a market map. How do you store all of the research? I know that because the fund is very data oriented, you might have your own systems internally, but how do you store all of this kinds of research, data, analysis, things you're going after? Great question. We store it all in very simple text files. And the whole idea is that AI is really good at processing raw text.

22:36So we store it all in text files and then we build systems on top. So some of those systems use AI to summarize and categorize and tag. And then we also build deterministic systems on top to do like long-term data analysis. ago? What is the average ARR of a Series A company that we've seen over the last 12 months, for example? I mentioned you're the master of product testing. I'd say you're definitely a Swiss army knife of product testing because I've followed your blog for many years. I feel like every time I read it, you're doing something else, you're trying something else, and it's really fun to ride along.

23:13I'm just really curious, like internally, like what is the operating system for that? What is the stack? And what are the different programs and tools that you regularly use? Yeah, I was, I mean, I was using a command line email client in the terminal. So I'm not using Gmail, but using this really, really old piece of software called Neomot for a long time. Because actually during COVID, I wanted to learn how to use the command line. And so, okay, why? Well, I remember, I remember meeting the Dropbox team at the seed when they came out of Y Combinator and they had commercialized like a Linux command line utility called rsync that allowed you to synchronize the status of two different folders.

23:57And they ended up building, you know, whatever, a$5,$10 billion company on top of that single protocol. And I think there are a lot of brilliant programming ideas that exist within the terminal that had been invented 30 or 40 years ago that are constantly surfacing and being exposed to different people. Like Superhuman Email Client is a great product. And it has a lot of what was built into MUT, which is a command line email client, or NeoMUT, which is a new version of it. And so like, you know, okay, why do this? Okay, let's say the first thing I wanted to do with an email was being able to hit a single keyboard shortcut and create a task from it.

24:39Like you email me, you say, okay, here's the agenda. I need you to comment on these three things. It's great. I hit F4. It creates a task on the command line and saves it to my task editor. Very basic thing, but at least it keeps me in the flow. I don't want to switch windows. The next thing I started to do is I started to play around with dictation. So I had carpal tunnel syndrome. Turns out you can speak three times faster than you can type. And so since 2012, I've been speaking to my computer. And now we're at a place where the accuracy rates are unbelievable. I mean, you play with like a whisper model, which is OpenAI species text model.

25:15I put it on my machine and then I think it's F7. I can hold it down and the computer will type. Then the next thing is, okay, well, take the output of that text, remove the ums, remove the likes, remove the oops, and reconstruct it so it sounds like it's in my voice, but the grammar is better. And so you're automatically answering those emails. And then anyway, so like one thing just, it's like a snowball or take a podcast and transcribe it and then summarize it. That's the project that we've been working on over two weeks ago. And so I think as long as you're playing around with the technology, you're constantly asking the question, like, okay, what's next?

25:53What's next? And then when you meet a founder, they're just like these little comments that you can make about like the friction around the technology that builds trust, which is, I think the basis of investment is building trust with people. Tomas, you're probably the most technical and involved investor I've talked with. So this is like extremely interesting. Theory's thesis is centered around data and you're data driven. I'm curious how you ended up on that as a differentiation for your fund. And how do you compete against the likes of other data driven funds like Insight Partners, SignalFire?

26:33How do you think about that? Yeah, I mean, I saw the power of data firsthand at Google. I just, Google provided us just, I mean, internet scale data and the kinds of products that we could build as a result of having exposure to that data, the kinds of questions we could answer just opened my eyes. And then I was really fortunate to get the job in the venture business and started working with data companies. And the idea there was that, well, maybe we could bring some of that insight, that level of insight from Google to other kinds of companies. And so I've been investing in data companies for a long time.

27:07And my graduate coursework was in machine learning. And so I also saw that happen at large scale. And now the basis of all machine learning is data. And so that's the core idea behind theory is every software company, every product that will be built for a user that's on a computer will have data as its fundamental building blocks. And so we invest in data systems, databases, databases, visualization. We invest in AI, both at the infrastructure and the application layer. And then we invest in blockchains because we view them as modern database systems with a very different and compelling architecture for certain use cases.

27:48Hey, it's Molly. If you enjoy our interviews, check out our newsletter, sorcery.vc, where we deliver a one-time-a-week top deals and tech headlines report and go deeper on our conversations. Subscribe to Sorcery today. And don't forget to subscribe to the podcast on YouTube, Spotify, Apple, or wherever you listen. The link is in the description to sign up. You more formally bucket this out in terms of the decade of data, AI as the new platform, and decentralized infrastructure as database. Could you break down each of these and maybe provide a portfolio company or two to better emphasize and just expand on that?