Welcome to the Peel. I'm your host, Turner Novak, founder of Banana Capital. Today's guest is Nathan Benaich, founder of Air Street Capital and author of the State of AI report. Nathan has been writing the State of AI for 8 years. It's a year-long effort on the biggest things happening in AI every year across research, industry, politics, and safety. We spend the next two hours talking about the biggest takeaways from his latest report, including the rise of reasoning, the surge in China's open source models, where AI is working in practice, the rise of sovereign AI, where he thinks value will actually accrue over the long term, if we're in an AI bubble or not, and how he's investing today at Air Street. A quick thank you to Nico at Adjacent and Dan at the University of Michigan for helping brainstorm topics for Nathan. A reminder, I publish new episodes of the Peel every week exploring the world's greatest startup stories just like this one. Check out the back catalog of over 100 episodes, including recent conversations with Marcelo Lera, co-founder of European unicorn Remote, and Kevin Hartz, co-founder of Eventbrite and seed investor in PayPal. Tune in over the next few weeks for guests like Garry Tan at YC, Jason Pudagant at Benchmark, Jake Sosch at Serville, and Duo Security co-founders Doug Song and John Oberheid. Let's talk to Nathan after a quick word from Numeral and Flex.
This episode is brought to you by Numeral. Numeral is the fastest, easiest way to stay compliant with US sales tax and global VAT. It's easy to set up and they automatically handle all registrations, ongoing filings, and their API provides sales tax rates wherever you need them with all the integrations you need. Numeral supports over 2,000 customers in both the US and globally, and they pride themselves on white glove, high-touch customer service. Plus, they guarantee their work, and they'll cover the difference if they mess anything up. They're fresh off a fundraise, closing a $35 million Series B from Mayfield, which they're going to reinvest into building an even better product. If you want to put your sales tax on autopilot, check out Numeral at their new domain, numeral.com. That's numeral.com for the end-to-end platform for sales tax and VAT compliance.
This episode is brought to you by Flex. It's the AI-native private bank for business owners. I use Flex personally and I love it because they use AI to underwrite the cash flow of your business, giving you a real credit line. The best part is 60 days a float, double the industry standard. Flex has all the features you'd expect from a modern financial platform like unlimited cards, expense management, bill pay that syncs with your credit line, and their new consumer card, Flex Elite. Flex Elite is a brand new Ramp-like experience for your personal life. A credit card with points, premium perks, concierge services, personal banking, cards and expense management for your family, net worth tracking across public and private assets, and a whole lot more fully integrated with your business spend. One card for your businesses, one card for your personal life, one card for everything. To skip the waitlist, head to flex.ai and use my code Turner to get an additional 100,000 points worth $1,000 after spending your first $10,000 with Flex Elite. That's flex.ai and code Turner for $1,000 on your first $10,000 of spend. Thank you, Flex. And now, let's jump in. Nathan, welcome to the show.
Thanks for having me, Turner.
Yeah, thanks for coming on. So, you put out this really interesting report. It's called the State of AI. You've been doing it for a while. I'll let you kind of explain how it all got started, and I think I gave people a little bit of context on kind of what we're going to talk about, but can you just kind of talk about this report that you put together, kind of how it got started, why you do it, all that stuff?
For sure. So, the State of AI report is an annual production. It's an open access document that I create in order to kind of disseminate the most interesting analysis across research, industry, politics and safety. And then we wager a couple of predictions every year in order to sort of cast things forward and see how we did the year after. The real goal is just to help people stay abreast of what's going on. There's just like so much stuff and you kind of don't know what is meaningful and most important. It also acts as like a good litmus test, I think, to see and sanity check to see like, hey, have we over-exceeded on progress estimations or under-exceeded or what are, you know, just how far have we come in the last couple of years. It's been running for about eight years now, starting in 2018 with me and Ian Hogarth who produced it together for a while and then been solo production for the last couple of years. And it's a good opportunity also to collaborate with people in the ecosystem and showcase, you know, good sort of example case studies of how businesses are using AI and how certain papers are leading through breakthroughs and what might be head fakes.
Yeah. I feel like there's a lot of data, a lot of numbers and like charts and a lot of visuals.
Yeah. Yeah. It's definitely not for the faint-hearted. I do try to write it in such a way that you could consume the headlines of every slide and then get a general sense of where things are going and then, you know, you can stop where your eyes get most transfixed. And so it's a bit more of like a buffet that you know must read end to end.
How much time would you estimate that you put into this thing because you do it once a year, like how much, how many like hours or like I don't know total days or however you want to quantify this.
It's a bit hard to give a clear answer on that because it is the result of just like everyday consuming of, you know, research, news, talking to companies. So it's very much like the result of my day job. But when it comes to genuinely producing slides, it's from early August until September.
And then you usually put out beginning of October it looks like.
Yeah. Yeah. We found like a good tempo around beginning of October. So like back to school season.
The report you put out in October, what were sort of the biggest takeaways? What we'll do is we'll throw a link in the description for people who want to actually like look at the whole thing, but if you're just like, ah, give me the, I don't know, the SparkNotes like quick version. What are kind of like the biggest things that have kind of been happening that you think people should know about?
Yeah, to me the probably four biggest things are one on research, which is clearly the move away from models consuming an input and just rapidly producing an output, but now going into like fairly complicated reasoning and tool calling. That wasn't the case even a year ago. It looks like, you know, you look at AI today and you're like, why wouldn't it be anything else than this? But this wasn't the table a year ago.
For somebody who's never heard of tool calling before or reasoning, like what does that mean practically?
Yeah. Yeah. So where maybe a little bit more than a year ago was, you know, a model had consumed basically the entirety of the internet, you know, people like to say, and then maybe some custom databases and has essentially like compressed and memorized all of that information. And it's basically like in sort of simplified terms, it can be thought of as like an API to all of that knowledge. And so when you input a query, it's kind of looking up its knowledge and then producing an output. It did not have access to the internet. And then like the first sort of tool call was do a web search, for example, for more relevant information because at the time what was very important were these like knowledge cutoff dates, which is basically like the date at which like the download of the internet was made. And so, you know, if something material happened after the cutoff date, it wouldn't be stored in the model because it didn't have access to real-time web search. You'd ask like who's the president of the United States and it would say Joe Biden and it would be like, no, he's not.
Exactly, or who is your creator?
What does it show for those usually if you search that?
Well, there's some interesting ones. I mean, you know, OpenAI would say, you know, OpenAI, and Anthropic largely the same, but the Chinese models would say weird stuff like they would sometimes say like, oh, my creator is OpenAI or my creator is this other company. And then, you know, that led to question marks on the Twitter sphere of, you know, are they training on outputs of American large models? You know, it's what's called like distillation basically. There was some FT headlines about this, you know, OpenAI were saying that this was the case and that was one of the reasons, you know, in a cohort of other reasons, why like US administration wanted to cut off Chinese access to American AI and things like that. So it's unclear if it's like actual schizophrenia or because like there's just a lot of examples of AI output that's generated by the leading labs on the internet and so statistically it's like more common to have that memorization.
Yeah. So essentially what this means is that it incorporates just the models with like what's on the internet. So the models can like learn and take in new data versus just what's been trained on and like captured in there and it's set forever. So these things can actually get smarter in a way and use current relevant news and information.
Yeah, I think it's even more than just being able to consume current and use news. It's like the designers of the system would ideally like separate memory, like storage of facts, from the ability to know how to retrieve them and how to like logically reason through answering a query. And so ideally you would want to learn the latter capabilities and not have to memorize facts in the ways of a neural network because then how do you selectively update them as information changes? And so yeah, the big push towards like tool calling, and you know, web search is a tool, you know, using an API could be a tool, using a software plotter could be a tool, is really to combine that with step-by-step reasoning where, you know, human annotators have explained in a very, very clear way, a structured way, how they would go about answering the question of like, how should I implement this stock trade given this is my goal and this is the ticker and, you know, these are my financial goals, etc., and anything you can really imagine. So if you repeat that enough times and you really learn like the logic of reasoning, then that logic can be abstracted away into many, many other areas and it's more repeatable than like just memorizing facts and memorizing like patterns.
So how is this kind of shown up if people are using a lot of AI products? Like where is this capability kind of shown up in what we're all using today?
The most obvious one is just when you enter a query like in ChatGPT, it'll say thinking and then it'll produce a result and then you can click on the thinking tab and then you'll be shown sort of simplified stepwise list of, okay, I think the user is querying this. Oh, maybe they're not. Oh, maybe I should go like disambiguate what they're asking. And then this is the step-wise plan to get to the answer. And I think it's interesting because it gives you like a vignette into what the model is doing to answer your question. So you can build trust in if the reasoning trace looks like what you would have done.
You can almost like troubleshoot it in a sense too.
Yeah. But then there's some other research around are these reasoning traces, otherwise called chain of thoughts, actually truthful? Like is the model actually doing that when it's telling you it's doing that?
Or is it actually just making it up?
Yeah, because in some ways it's like trained to recapitulate like human behaviors of knowledge traces. And so it might be producing a result, but like showing you what the human would have done because that's what a human would rate as a good reasoning trace.
So they've like learned how to lie maybe, like hide their hallucinations.
Yeah, there is some of that paper-wise that's come out. And I mean, to some degree it really depends how you measure these, how you measure that behavior, and maybe your measurement is wrong and therefore you're drawing the wrong conclusion.
Interesting. So you said that was one of the four. Is that probably like the biggest thing that's kind of happened?
I think so. Yeah. Because that's like unlocking computer use, which is like the model can interface with your computer and do a bunch of things with it. It's unlocking, you know, more complex like scientific and maths discoveries, which, you know, the AI world is very excited about.
Yeah. And then I say like the second big thing is like the rise of Chinese open source. I mean, it's only, you know, 12 months ago that, you know, DeepSeek moment happened and it was probably six months before that when DeepSeek, the Chinese company, had released like early versions of its model, this base model that was then tuned to do reasoning and resulted in this like DeepSeek freakout. And since then, you know, many more models have reached, you know, leaderboard headlines, in particular in world modeling and in vision. So basically anything to do with like making pictures or videos, long form videos, Chinese systems are very good at that. And then the like resurgence of Qwen notably, which is Alibaba's system that effectively like took the mantle from Meta's Llama initiatives where just a year ago, you know, the tech world was applauding Zuck and others at Meta for, you know, stepping into the fold and saying we will always do open source, it's good, you know, we've done this with PyTorch and other of our core technologies. And then of course like big vibe shift there and China really like stepped into the fold. And then particularly in the last 10 days, there's two filings for Chinese model companies, one MiniMax and one is called Knowledge Atlas company, which actually the regular name is like Zhipu AI or Z.AI. It's produced this model, it's called GLM. Anyway, like too many acronyms at this point, but the point being is like these are the two like first pure play large model companies that have gone public and they went public on the Hong Kong exchange for several billion dollars each and since like ripped. Yeah. So China has been first to market in this sense.
In the first to market in getting a stock that people can buy.
Yeah. Yeah. Well, Americans can't buy MiniMax. If you have US nexus, you can't buy it. I don't know, I haven't looked up the exact reason, but the other one I think you can.
Interesting. Are there like ADRs where basically they relist the shares? I think, I don't know exactly how that works.
Technically, there are no ADRs yet. It's just on the Hong Kong exchange.
So, one thing I wanted to ask you about, maybe this is a good time, like open source versus closed source. What is the big deal with that for someone who has no idea, like why is open source so important in the context of, what does closed source even mean? Like maybe give us a real quick 10-second and then what the importance is.
Yeah, there's a spectrum on this. So traditionally open source in the context of like regular non-AI software meant that the source code, which was basically like the end-to-end sort of book basically that produces the program, was available on the internet and could be reused. You could copy and paste the code, publish it and use the software technically.
Yeah. Yeah. And it would work.
Yeah. And you're allowed to use it sometimes for non-commercial reasons, sometimes for commercial reasons with or without paying a license. And the beauty with open source that people got excited about was that anybody on the internet can propose a suggestion, a change or discover a bug or produce a feature and then the maintainer of the project could basically approve or deny or provide suggestions. And there was like a moniker that a couple of years ago that was used to describe this and it's like the bazaar where everybody can kind of participate. And then in other ways, closed software is kind of like a cathedral where like you have one access point and no one can touch it. No one can touch the cathedral. No one can add anything to it. You can't influence how it looks or how it will be changed. It's the owner of it that decides. That's a closed source. You pay a license to get access to it.
You go and you offer an offering like a tithe to the cathedral to use the cathedral. You pay a tax.
Exactly. You pay tax in the form of a SaaS license.
And then this nomenclature just gets a little bit more complicated in AI because there are more moving parts. So for simplicity's sake, there's the training code, which describes how a model, how an algorithm should be run and what's the goal of the training, what objective are you trying to minimize, for example, like, you know, the dog-cat thing classification, you're trying to minimize the mistakes. Then you have the data set and then you have the model itself once it's been trained, which has weights and parameters, which are like settings effectively that are produced as a result of running data through an algorithm. And so what is it exactly that we mean when we say open source model? Is it the training code? Is it the data set? Is it the model artifact once it's been trained in the sense of its weights or the code to run it? So there's a spectrum of definitions. Right now a lot of people use like open weights to mean, well, I'm actually just releasing the resulting parameters, tuned parameters of my training run. This would be the equivalent of saying, you know, I'm basically giving away the result of hundreds of thousands of hours of GPU training or something like this. I'm not necessarily giving you access to the data. I'm not giving access to the training code or the pipelines to run it.
So you couldn't reproduce it.
You can just see the results of what I did. It's hard to reproduce the training run. Yeah. Because you don't have the data set, but you're getting the end artifact. So it'd be like, I don't know, like the Formula 1 analogy, you know, like the car that wins like the last race at Abu Dhabi in the year. Like, you get that. You don't get all of the learnings and recipes in order to get the car. You just drive the car. And so the community generally wants to have open everything because it leads to, you know, better reliability, more contributions from, you know, the ecosystem in terms of features and bug fixing and things, and then crucially like control over the entire system. So, you know, if a company decides to no longer publish open access and open source tools, then you didn't get rug pulled.
Because you just no longer have access to it or updated.
Yeah. Exactly. At least you have like the last timestamp released so if the system's good, you can still work with it. I think the second reason why folks want open source is because the training of these systems is expensive. You know, generally for anything that's like state-of-the-art, like, you know, millions, tens of millions, hundreds of millions, not many companies or certainly not universities or small groups have access to those resources. And so they can't train systems from scratch. So they rely on these getting like air dropped basically for free. And I think the last one is just the community doesn't want to have, you know, one, two, three companies the entire future of AI development because it is very limited on money, on compute resources, and on data set.
And then the reason you would be doing a closed source model is because it's probably easier to make money from it versus like you have to pay to use it and access it versus just giving it away and you can't charge as much or anything.
Yeah, I know. My two takes there are like open source is almost good when it's a necessity for your customer to buy your thing because if your customer is a developer, they probably, and the tool is catered towards them or some low-level like database or framework or something like that, they're going to want to peer under the hood and see how it works. And so if that's part of your sales motion, then you have to do open source. But I think I feel like rarely open source has been a competitive like tool that one uses because it's like good for business. It's almost like a necessity. Or secondarily, if the business is born from an open source project like Databricks with Apache Spark, like the thing is so complicated but very high value and customers don't want to have to manage it themselves. So like let's pay Databricks for the expertise because they wrote it and they're the best. And then the other part around like open source versus closed is considering that it's like Apple versus Android. Like at the end of the day, I think customers and regular people want a tool that's like really good, that's designed well, that's updated all the time, that's reliable, simple to use, and that's basically Apple. There is a period of time in like the early days of technology, I think, where a lot of nerds like to play with it and like mod it and change the background and like change these little things, but at some point it's like that community, I think it goes down over time because it's just a faff, like it's annoying. And I just think like, you know, when the last time was that I added a mobile phone number that was not an iPhone, like this, it's a green check mark in your iMessage. It's like weird. And I think this is because long-term people gravitate towards convenience and quality. And I feel a little bit of the same way in the open source versus closed source.
Interesting. Yeah, because it's probably you probably get to a point where as the technology continues to get better, you're just like, I just want this thing to work so I can do my value-additive thing that I'm doing on the model versus mess around with tweaking things because I just know that it works and it has what I need it to have. It kind of reminds me, it's kind of interesting how Apple, and I think of how Apple evolved. Like I used to think of Apple as like for old people, right? Like back in the 90s, right? Like your grandma has an Apple, one of those big massive like monitor things. It's like a fancy color, like it's like purple or something, and then it's like there's only one button on the mouse and it's super simple to use. Like you can't get a virus because like the old people, you know, they'll just get phished and like they get a computer virus or whatever. But it's like totally evolved over time to now where most people just use a MacBook product and you're kind of weird if you use Windows still. And I feel like people will reluctantly use them if they're working at a big corporation or something like that where there's like a work necessity. Yeah, it's just interesting how Apple's kind of evolved the products over time.
Yeah. And I think realistically as model capabilities have gotten better, like these things are so unwieldy and complicated and insane lifts with thousands of people like contributing in very tight control and I just don't know how you coordinate all of that on the internet with like randos on your laptop in cafes.
And is it probably, I mean, and there's tons of forks, right, where you basically take it, you make a tweak and then you republish it. So there's probably thousands, tens of hundreds of thousands of just like variations of different models that have been slightly tweaked over time, I'm assuming.
Well, there's that in the open source community. Yeah. But like to create a Gemini or like a GPT or something, I mean, there's so many, there's like basically all these different teams that are each responsible for like a different skill or capability. And each skill or capability has a set of evaluations and sometimes like hundreds of them, which is like ways that you test whether the model has succeeded or failed or what end of that spectrum for that specific capability. And then you have to collect data specifically for that and then just throw it into the soup and make sure that everybody else who's throwing in their capabilities and data and metrics and evals into the soup doesn't degrade each other's capabilities. And so like how do you coordinate that at scale on the internet like without being a centralized like company? I just think it's unrealistic. And I think the proof point to that is like no one has really managed to do it in the sense of no one's managed to publish like a model as good as, let's say, Opus 4.5 or GPT-5.2 like in an open source setting. I mean, people have come close but those are basically private companies that operate, those are like closed source model companies that just happen to open source as opposed to like true open source development which is like distributed contributions.
So it's almost like you need a dictator or like an opinionated like the PM, the person managing the whole process of training and releasing it.
So then I know we've been, you know, this has taken us a long time to get through the top four most interesting things, but what would you say is like the third or fourth most interesting thing?
I mean the third or fourth we can probably breeze through a little bit. The third is around like real revenue scale of products. Like it was only two years ago that, you know, the major labs were diminutive in revenue and, you know, there was that example that, you know, Jasper was making way, way more money than OpenAI was for very similar use cases. And this debate of is the model the product or should there be lots of other like scaffolding and sticky tape and UI around the model that would mean that it could extract more value for that targeted end user. And then that seems to have flipped as the core model is better and it's been kind of the interface for everything and then model vendors have created their own like sort of SaaS wrappers basically. So that's like tens of billions of dollars across major players. And then the last one I'd say, and there's many more, but last of my favorites at least, is just like the AI sovereignty marketing agenda has just gone like on full steroids. So this is, you know, the competitive positioning idea of nation states need to have access to energy, compute, data, and talent and models in order to dictate their own fate in the future of AI. And Nvidia has gone on, you know, a very aggressive marketing spree like selling this agenda to every single country that is willing to listen. And, you know, that's resulted in commitments of over a hundred billion dollars globally in different countries to build up data center capacity.
Just pick like a random country in Europe, like the fourth largest, like Italy or something in Europe, and like, I need to have my own sovereign AI strategy. What am I probably doing?
Well, this is kind of my point. It's like, what does that even mean?
I just keep invoking this, you know, Kanye West lyric of no one knows what it means, but it's provocative. It gets the people going. It's a bit of that. What they think they're getting is the ability to train and run AI models in their geographic vicinity in a way that cannot be unplugged or tampered with by any other nation state than themselves.
So like if you go to war with a country that a major lab is domiciled in, you don't get cut off from the ability to use AI.
That would be the idea. Yeah.
Yeah. And I could see how, you know, if AI continues to kind of get better, it becomes all software. Like you basically can't use software to defend yourself and maintain your sovereignty and like you'd be in an extreme disadvantage if you didn't have that. So I would be extreme if it's positioned like that to me, I'm, you know, spending a couple percentage of GDP on it. It's like I'm opening the checkbook probably.
Yeah. Yeah. But the challenge with it is that the hardware improvement cycles are very short. You still need to have like a bunch of software that runs on the chips to make them work. And maybe there's like a major update that gets shipped to them that's required to run new models on your hardware and, you know, the vendor decides not to ship it to you because, you know, we're not friendly with your country anymore. So, congrats. You got the hardware, but you can't run it. A bit like the whole spiel with the F-35. Yeah, we, you know, foreign nations can buy the F-35, but you can't fly it if you don't input the flight plan, which gets communicated to the US.
I didn't know that it worked like that. I mean, that's tricky. Yeah. How do you get around that? Or they, or like this just could be nerfed or...
And it's interesting from Nvidia's perspective because Nvidia has a customer concentration problem where basically Google, Facebook, I don't know, Amazon and OpenAI, maybe Anthropic is like most of their revenue, like I don't know, 90% of their revenue, something like that. So it's like, can you figure out how do I find a couple more, you know, nine-figure customers, right? Like get a couple more people that are paying me billions of dollars a year to get a little bit more diversification.
Yeah. Yeah. I think it's a great strategy, by the way.
I think they're right to do it. And you can see that like the next growth vector is, you know, data centers in space. So it's like what's always like the next town you can create, right, to keep selling.
What is sort of the end state of that? Because if you kind of go back maybe a couple years, if someone were to say we're going to build data centers in space, it just seems like just a word salad of just random buzzwords that makes no sense. Like does the entire universe just need to be like filled with like data centers, like power, like what, where does it end? Do you know like what actually happens?
I'm not sure what happens at the limit, but I could see, you know, a pitch around as we get more assets in space of various kinds and maybe we'll have space civilizations at some point. And so having compute there is just better than having it on Earth that we have to communicate to and from, which currently is cheaper than having data centers in space. Potentially even for like, you know, defense reasons, you might want to have compute in space for all of your space systems. But yeah, I'm not the space data center expert at this point.
Yeah, because like part of me is like this just sounds like it's hundreds of years in the future, but also maybe part of me is I feel like there's that quote of like less happens in one year than you think and more happens in 10 years than you think. So like with self-driving, my working theory with self-driving is like it's always going to be a couple years away, but I guess now it's actually finally here.
Yeah. Yeah. But also, you know, historically Nvidia has been exceptional at, you know, resourcing teams whether they're, you know, in companies or in academia who are exploring new things with their stuff. And, you know, before it was like shaders and graphics and then it became like AI with AlexNet in 2013. And then it was, you know, these AI labs, nonprofits that were started 10 years ago and then in bioarma. And now if some kids are launching, you know, your GPU in space, like I think Nvidia is pretty curious to know like does it work and if it doesn't work, like how can I fix it because who knows where the next growth frontier is. And, you know, if that's driving excitement for this stock, like even better. But I think like their net outlay in, you know, like figuring out if this works or not is pretty low for the potential upside they could get.
And it's not like they're losing money because they're like selling products they generate cash on, right? So it's like you can scale it up or down like, oh, the space data centers didn't work but we made $3 billion because everyone wanted to buy some for a year. It's like with crypto, like you go back to 2021, like the crypto bubble almost is like dwarfed by the AI surge.
Well, this is interesting because as you said, it's been dwarfed. And then crypto prices, you know, year on year now are like down whatever, like 20, 30% or so. I remember this time last year it was like 120 or something and now we're at like what, 95 at least in BTC.
97, just in Bitcoin's at 97. I saw a notification.
So it's going up. It's going back up.
Yeah. Like the Bitcoin miners have found something far more lucrative than mining Bitcoin now because I guess because of changes in the hash rate, etc., and energy prices, and that's like basically swapping their compute workloads from mining to AI. And so like last year we saw this rush of companies from, you know, Iris Energy, which is this like Australian business doing energy originally and then starting to do BTC mining or crypto mining and now is like transition to being a GPU company and same thing for Cipher Energy and then like Hut 8 and all these rando companies that are...
Have you heard of any of these?
Yeah, I mean you could look at them. There's like six of them and they've all like ripped like 100% or 50% last year. I mean, they're crazy volatile. So you got to have a stomach for it and who knows if any of them are really going to survive because they don't have the balance sheet for it. But as you know, traders are looking for volatile stuff to just like play on Robinhood. These things have been pretty popular as everybody's trying to front run. Oh, like, you know, Anthropic's launched a new, you know, 1 million TPU data center thing. Who are the other vendors that can profit? And so I think for that reason like, you know, BTC prices have also gone down.
This kind of leads into an interesting other kind of like phase of the conversation that I kind of want to ask you is like how do you feel about where we're at and kind of the cycle of excitement in AI? Like the maybe the...
A more simple way of phrasing it is like, are we in an AI bubble right now? Um, I don't know, how do you think through that as someone who's been in the space for a really long time?
The first thing I think about is if I were to have seen the capabilities that we have today, like this magic box you can ask anything and it basically gets anything right, and you showed me that 10 years ago, I'd have told you that is like absolute magic. And pretty much everybody in AI would have told you that's magic, that's not possible in a decade. And this is across just general, you know, question answering, but also video understanding, video generation, audio, like it's insane what we have today. I would have thought this would have taken way longer, and the future got pulled forward so damn fast.
And then we're still only in what, two or three years into this cycle of getting this like magic alien artifact with no, you know, instruction manual, no genius bar that tells you how you're supposed to use it. And the very companies that have developed these tools are figuring out how best to use it so they can educate everybody else.
And then you already see like vignettes into the future where companies that have adopted this and people that have adopted this have significantly higher productivity than they had before, as reframed as if you were to take this, that classic question of if I were to take this thing away, how pissed off would you be? And most people would be like pretty beeping pissed off.
And then also other things like, you know, cloud spend as a proportion of overall IT spend is still not ginormous, and that's been the case, you know, since like 10, 20 years, and that hasn't even factored in like AI distribution.
So I think we just have so much more to run if you just consider what we have today and educating everybody on how they can use this stuff. And also factoring in that the artifacts we have, at least if you look back three years ago, were a result of like a very, very small number of like highly technical nerds who like stumbled on something that really worked. And we're not product designers, we're not like behavioral psychologists, we're not, you know, like large-scale systems engineers or cloud or like, you know, cost optimization experts, etc. All this stuff, all these tasks that are super important, building like basically the internet, had not worked on AI, and now everybody's going into AI.
So I think there's far too much like money, resources, talent, and like genuine desire to use these things that progress can't get better.
Now does that mean like currently we're in a bubble? Like I don't actually think so. Uh, there might be like pockets of bubble-ish dynamics, but you know, the companies that are accelerating super fast in AI like Mag 7 are actually printing like tons of money. Uh, they have super healthy balance sheets. You know, forward like valuations are, I think, like 50% of the top names back in the late 1990s. Um, they're using their own money to fund a lot of these data center buildouts. I mean, there's some that are doing off-balance sheet credit, but I mean, it's like sophisticated buyers that are buying this. It's not Joe Bloggs that's, you know, betting their pension or something on, you know, Meta's data center buildout.
I think I saw a really interesting stat, too. It was that this was probably about maybe six months ago, nine months ago at this time. So this is probably even a different stat now, but it's basically OpenAI and Anthropic since the launch of ChatGPT had added more new revenue than like every other publicly traded software company, and this was, you know, a while back, so it's probably an even bigger diff, like it's probably even bigger now.
Yeah, it's astonishing. The valuation of these companies is just like, do people pay for your product and do you make money on it? Maybe that's a whole different question here. Like you can argue about the margins of some of these things, but like people are adopting the products and paying for them much more than if you don't have any AI capabilities in the product.
Yeah. Well, the margin profile looks like it's actually improved quite a lot.
So why so? Because I feel like even within the past couple months, I still see some of these headlines where it's like, oh, the vibe coders have negative gross margins and they're three months from bankruptcy or something like that.
Yeah. Well, I guess those are a bit different because they're like pass-through to model providers. So what I'm saying, you know, the margins are pretty good. I'm talking specifically about model makers, whether they're in like image or video or chat.
And I would say, I mean, without disclosing like specific names of companies, like some of the best ones are like at 60, 70, 80% gross margin on serving their models. And like, yeah, they spend a lot on training. Uh, but to some degree, like less than they did before, because whereas like pre maybe one year ago, the sort of quote-unquote recipe for how to build a large-scale model that worked really well was not clear or not written. And now I'd say like most people at the frontier would say like there is a recipe for scaling now. Like we sort of know how to do pre-training, the human mid-training, and then the post-training, and then RL, and then the eval, and all this stuff, kind of know how it works.
So we're getting more efficient.
A lot more efficient. And then the other thing is the quality of the model you get today compared to what it was a year or two years ago is actually far more useful. So then you can monetize it better because it's monetizable, like people are willing to pay for the outputs, and so you actually get faster payback. So you're more efficient. People pay for your thing because it's better. More people are educated on how to use it, so they want it more. Yeah. Then your general go-to-market is improving.
And there's some modalities that are cheaper than others, like text obviously probably the cheapest. Audio is actually not too expensive. Video and world models are very expensive.
You're saying to make or to also sell?
To like, it's expensive to do the training and then also expensive particularly to do the serving, the inference.
So it's just the pricing to consumers then, it's probably higher, like if you want to tap into a world model API, it's probably much higher than just text-based.
Yeah, exactly. Like to make a nice video on Google's like V3 or like the, you know, the super pimped out model, it's a couple dollars. So if you're selling to social media people and it costs a couple dollars per video, like it's a bit tough.
Yeah. But on the other hand is like if somewhere someone else, like an agency or a designer in Illustrator or like literally take a camera, record an expert doing the thing or like an actor, a dancer, whatever you want. What is like the cost of literally typing in 'make Nathan dance with an iPhone' for an Apple commercial versus like actually go make it? Like is it a thousand times cheaper and it also took two minutes instead of two months? Like I feel like that's part of the equation too, of like as the products get better, it just like, things just get so much more efficient where like you can literally have, it's like the person who's the marketing person at Apple instead of coordinating with the agency, they're just working with Nano Banana or ChatGPT or insert whatever tool they're using and just instead of sending an email to the agency going back and forth, they're just banging it with ChatGPT for a day. It's like, cool, this is what we would have got for a thousand times cheaper and literally like, you know, in 10% of the time, like we got it 10 times faster too.
Yeah. Exactly. So it's not particularly surprising that like, you know, ElevenLabs is ripping. I heard it's like, or you tweeted it was at $330 million in revenue and got $100 million in net new ARR in five months. It's like bananas.
Yeah. I saw $15 million in a day or something that they were tweeting about, which is, yeah, I mean that's run rate, that's pretty high.
And that's what people are doing, right? Is they say like, oh, we're at an $8 billion run rate because we signed whatever, like our Stripe balance hit the account today, if we times that by 365.
Yeah. I mean that would be mad sketch. Yeah.
Yeah. I mean it's kind of happening, isn't it, a little bit.
I think there's everything that's happening. I would say they're pretty above board, but there's certainly like some odd behavior everywhere.
But I think the magic here is just how much like latent revenue there was in these use cases that was only made unlockable as a result of like great capabilities, right? Because I still remember, you know, two years ago when Eleven started or when Synthesia started seven years ago, like it was not mega obvious that there would be like half a billion dollars worth of revenue for a product like this, you know?
Yeah. You'd be like, who gives a about, why would somebody want an audio model? Like there's no demand for that, right?
Right. Well, especially that there were audio models before. I mean, you know, Amazon had one, Google had one for TTS and speech-to-text.
Was it that they just weren't very good?
Yeah. I mean in London, like where I was, like I spent a lot of time, you know, it was always funny to me that I'd, you know, exit King's Cross station where, you know, Google DeepMind was, and you would hear the audio voice and it was like, this is clearly a robot and it sounds really crap, but there's a company like 300 feet from here that has a way better model for like the last 10 years, and like the Tube can't even use it. It's like we're still so early.
Yeah. Yeah.
So I'm curious then, are there areas within AI that you think are like a little bit overextended? Like when I just think about, because if you were to open social media and you'd read the commentary from investors, a lot of people say AI is in this massive bubble and things are going to crash. Are there like, so it sounds like you maybe don't necessarily agree with that, but are there certain areas where you feel like there's a little bit too much, you know, things have gotten a little too far ahead of themselves maybe? And I think maybe another way to think about it is, you know, with kind of just, we kind of talked about across the board like the products just keep getting better. Are there areas maybe the products aren't actually getting better at the pace of maybe like how excited people are getting about them?
In some areas like in science, I'm very excited about like long-term progress in AI for science, which is anything from, you know, helping human scientists process, you know, research papers, formulate ideas, figure out what experiments they should run, you know, test run those experiments and analyze the data. Like, you know, AI models are making like substantial contributions there, that's like a net positive for the world.
Clearly in biotech this works. There's a good business model for it if you develop drugs. There's an ecosystem of pharma companies that'll buy them. There's contract research organizations that'll run the experiments for you at scale.
And the industry has like transposed all this excitement over to like materials. And I think there's been a phase of several companies raising substantial money for materials. So much so that the venture dollars that have gone to materials companies is far greater than like US national funding for material science labs.
So it's like, you know, whereas the US government should be funding frontier like, you know, R&D, like venture dollars have stepped in there using many analogies from life sciences. But to me, like I'm very excited about that space as well, but I haven't done any investments there because I think it does not share a couple of the key features that is present in biopharma, which is like lots of pharma companies that have been structurally built to buy biotechs either outright or the drugs they've created. Like this is how biotech originated. This is like the tacit agreement that small biotech companies have with pharma companies. Like we take all the early risk and then if it works you buy us. That's how it works. Material science, not really.
And then there's not like this kind of industrial base of contract research organizations which are like industries that will run experiments for you. In the case of materials, okay, my AI model has popped out a bunch of different versions of like titanium or like some composite and I want to go make it. I can't send it to anybody to go make it. And so most of these companies are either doing like their in-house labs, expensive for all, you know, a whole bunch of other reasons, a bit inefficient, or they're sending it to academic groups who take like six months to make the thing, and then they're making like a tiny amount of powder, and it's like, you know.
And then the other thing is then you have like big companies like Meta that have been pouring in billions of dollars into material science and then releasing, you know, data sets and models and are really just constrained by synthesis, like actually making the suggested outputs of the model. And so I think materials is like rate-limited by all these things yet is seeing like irrational exuberance from venture investors as like the next frontier that's similar to AI and biotech.
Interesting. So what is an example of like one of these like materials companies?
I mean they're probably one of the most exciting ones out there because they have, you know, a very cool team, you know, two co-founders, one worked at OpenAI on ChatGPT and post-training, and then the other one worked at DeepMind doing material science. And then they're doing, you know, as far as like public reporting has chronicled, they're doing in-house testing of materials, and then, you know, using AI to scour the universe of potential materials, do optimization, and then using the internal labs to test.
Be like inventing like a new periodic table, like inventing a new element that goes on the table or something like that.
It wouldn't go so far as that, but it'd be like what's a new formulation of like a superconductor or what's a new formulation of like an alloy.
Like stronger steel or something.
Exactly. Yeah. That could be stronger and lighter and cheaper to manufacture and therefore could be great for lots of different things, or like a material that you could make, you know, an iPhone screen that you could then go outside and not get blinded and it would just work in, you know, with direct sunlight, like things like that.
That could be amazing. Like I can totally see how like you make these awesome products, like let's use AI and make like, I don't know, like how do humans have wings and they can like fly or something? I don't know. Like I can see now like that's, the TAMs are on these things are insane.
Yeah. No, I mean I think it's very inspiring.
The question is like, you know, would you pay, you know, almost close to public market valuations for material companies, for a private business that's just started to do this on the basis that it could be like the next frontier lab in this space and follow the same fundraising dynamics?
I mean, that's kind of the role that a lot of venture investors play is like the momentum identifying and trading almost in a way. So, I guess it's just like, is that your strategy? And if that's your strategy, then that's what you should probably do if like that's what you're telling your LPs is what you're doing. So,
Yeah. Yeah, for sure. I think it's the strategy that one should do if, you know, one scaled to billions of dollars of assets. Like I think there is probably no other way to move that volume of money into ideas that could be big enough. What you could say on the founder side is like amazing that, you know, entrepreneurs have the opportunity to go pursue their dreams with like insane balance sheets on ideas that should have been funded by the US government a couple years ago.
But, you know, at the same time, if it's your dollar that's getting spent and you look and you're a little bit more pragmatic, which is what, you know, LPs tend to be, it's a bit of a stretch.
Yeah. And it's just like so fascinating when I talk to people that were in the industry in like the 90s or whatever and they're like, 'Yeah, I had to sell 60% of the company for like a million dollars and there's like a 2x liquidation preference and it was like a real product, real business. It was profitable.' And then today it's just like, you know, 'I interned at an AI lab and I raised $8 million to go build, you know, a material science thing that... Yeah. It's like I'm going to make stronger iPhone glass.' It's like the contrast is pretty insane.
Yeah, it's tricky because it's all about like risk-reward. You know, you take a lot of risk if the reward upside is big, but if the starting valuation is billions, then the reward is not necessarily capped. But I mean there's some reality. I mean as soon as you actually become a materials business then you get valued as such or...
Yeah, but the argument that people will use is that the outcomes are so much bigger today and that with generally, with like, I mean with a lot of AI software it's like replacing the work, right? So like your customer might historically only be spending 5% or 2% of their revenue on software but they spend 50% on labor. So in theory, you could maybe capture like half of their labor budget. So like your revenue potential goes from 2% to 20% of their revenue, which is, you know, 10 times bigger outcome. So you can afford to pay that 10 times higher entry valuation.
So I don't know, you can argue it where it works.
I think that's true on the outcome size. I would just rather not have to pay the 10x on the front. So the return is 10x bigger. But, you know, to each their own, I guess.
You come in at like, you let the multi-stage platform funds do that after you, and then they help support what the capital needs to like, you know, make it towards the public markets after you've invested as an investor.
That would be ideal, but I mean, you know, some of the mega funds are doing this already with these mega fundraises where, you know, the company announces, you know, a multi-billion dollar price tag, but there's been three tranches before that where, you know, brand-name mega fund got in first. You know, what for a seed manager is a high price but for one of them is like whatever, and then there's been two markups since then in this short period of time. So there's a bit of that.
And you can almost like manufacture those in a way where like you put in, you know, 2% of your fund at 200 million post, and then you put in another 1% of your fund in a follow-on round with one of your LPs who leads the round, who for them it's also a very small percentage of their portfolio, and then you have an extremely small round where the company sells like 2% of the business at a billion post-money, and everyone kind of looks good on paper and nothing's changed in the two months since the first round happened.
I mean, this is exactly what's happening sometimes.
And it is, I don't know, it just strikes me as pretty unhealthy. And sometimes the rationale is, well, employees want to know that their equity is worth a lot of money, and so, you know, they'll get the equity early and then there'll be this big write-up. So it's almost a way that companies can offer like lower percentage ownerships in their business in the form of options, but because the markup is so high, like the paper value is like already in the millions, and so then they can try to compete with offers at the big labs. But obviously dangerous because of all the liquidation preference features that are present in these extended valuation companies, not least growing into them. So that's a little bit scary to me.
Yeah. And for people who don't know, liquidation preference is liquidation preference. And it's basically in a lot of these rounds that companies raise, if you... any dollar that comes into the business, if there's a liquidation preference, it's usually one times. Sometimes it might be two times, three times even depending. But typically, if you're in your standard venture round, whatever dollar you raise, those investors kind of have right of first refusal on any exit. So if you've raised a billion dollars, the company needs to be worth a billion dollars and all that cash goes to them first. And then beyond that, the employees, the founders, etc. get paid.
So, if you take a job at a company that's raised a bunch of money, I mean, there's a ton of nuance to this, but it's probably like whatever amount you've raised, like the valuation needs to at least be that price before you get anything. Like, the investors generally get kind of a right of first refusal. It's not always the case. You should probably ask when you're kind of talking to them about taking an offer and like figuring out exactly how that works. Google it. There's tons of YouTube videos that explain it pretty well. There's like full podcasts where people explain the nuances of this stuff, but the price, like just because you work at a unicorn, might not actually be worth that much at the end of the day. So,
I mean the other thing is whether your stocks get vested immediately once the acquisition happens if you stayed for a shorter period of time than is like required. So, if you get acquired in your first year and you're supposed to stay four to earn all your stock, like do you earn the four on the date of the acquisition or not? And that's like been particularly topical in these like pseudo-acquisitions of shell companies and things.
Yeah, I did actually have that with one company that got acquired. The founder specifically negotiated all of his employees like fully vested, which I think a lot, I mean a lot of this comes down to who you work for. Like who is the founder of the company that you're joining? Like do you trust them? Not just in can they build a good product, are they commercial, can they sell things, can they grow the company, is also like are they going to treat you fairly? I think is like an under-discussed topic that is probably pretty important to think about nowadays.
But so then we kind of maybe alluded to this a little bit, but in terms of, so we could maybe talk a little bit about Air Street and, you know, the fund that you run and just kind of like what kind of things do you invest in because we talked a little bit about maybe what you might not be participating in. So maybe it might be interesting just like how did you get into this originally? How did you start investing? I know you've been doing it for a long time.
Yeah. I started getting interested in venture capital in college because I started in 2006 and that was the era of like Dropbox, Skype, Twitter, SoundCloud, Facebook, etc. All these like tools that, you know, I just used as a regular college student and get excited about them or read about the history and was classic, you know, kid in a dorm room like thought of something and it became big. And at the time I was majoring in biology and I got the chance to work at the Whitehead Institute, MIT in Boston, and that was a place where they did the whole genome sequencing project. There were some venture funds in the area that were, you know, hunting around labs for new, you know, ideas and drugs and innovations that they would then like spin out into companies.
And I was just really excited about this, like working on new technology, new frontier ideas, and then try to like make something out of it for the utility of, you know, real people and companies.
And after I did my PhD in the UK from 2010 to 2013, I was really convinced I wanted to somehow play a role in the startup ecosystem. Didn't really know where to start or what would be a good fit. And so I think naturally like being part of a venture firm would be a good sort of aperture to like everything because you get this like unfair calling card, you know, that you can use to talk, you know, enterprises of different types, meet entrepreneurs who tell you things that, you know, hedge funds pay lots of money to GLG for.
You get to learn of successes and failures. And then I could maybe find what I found most like kind of fulfilling long-term. And honestly like what I found was I like working with brilliant people. We're trying to invent the future and particularly this intersection of like AI, science, engineering and building real companies.
I became even more interested in AI like in 2014-15. That was around when DeepMind was acquired in London and it was not that far away from where we were, and I had some friends working there, and it really captured like the interest of machine learners at the time that their skills could be used for practical things. At the time, you know, DeepMind had shown Atari and it was like the first time that a computer could solve this video game and discover tactics that humans had never found. And I think that nugget of hey, like if you could have a learning machine that could amalgamate more experience than any human on the planet and learn from all these experts, then there's certainly going to be like solutions to every problem in the world that we haven't yet discovered but that the computer can help us discover.
It's interesting like the classic like, you know, ChatGPT or Claude Code like 'do this thing for me, blah blah blah, make no mistakes.' It's like kind of like the meme but it's almost like real. It's like solve chess for me or like solve world hunger, go.
I would challenge like anybody who's interested in technology and just like world progress to like watch The Thinking Game on YouTube, sort of the contemporary story of DeepMind and all the stuff that they've done, and walk away from that and not think this is probably the most exciting thing that could possibly happen on planet Earth and like I want to be involved in it somehow.
I've seen it. I just have not watched it yet. But it's been kind of on my like, oh, I should convince my wife on a Friday night when we're gonna watch something to watch it instead of like, you know, something else.
It's well worth it. I mean, the narrative is great. I mean, I think, you know, Demis is like one of the, you know, perfect sort of like characters that you want to win. He's like, he's intellectually curious, like brilliant, you know, like too modest and just super driven to invent new things and do it in his lifetime and feels like this irrational sense of urgency of especially when he's describing like why the hell did you sell DeepMind, you know, like looking back.
It'd be worth trillions of dollars at this point.
Yeah, exactly. I mean he sold it for 500 million pounds which at the time was like insanity. Now companies are raising that right from day one. And he's in this taxi like describing like, look, like if I gave you the opportunity to have like, you know, infinite compute and infinite money and resources to like accelerate, you know, potential invention of AGI in my lifetime, like there's nothing more that I would trade than being able to use my, you know, fruitful years when my mind is still working, I have the energy, I have like, you know, the physical capacity to go do this. Like I could have waited, you know, five, ten more years and like struggled here and there because nobody wanted to fund this stuff at the time, but like I wouldn't trade those five, ten years for tens of billions of more dollars.
Yeah, that's huge. It looks like it's on YouTube. Yeah, I'll throw a link in the comments or in the description for people. So I guess you made this transition from you were working at a venture fund. It was 0.9?
I worked at a firm called Playfair Capital before that which is pretty much a family office, and then I moved to 0.9 in 2017 until like late 2018, and then had the first closing of Air Street Fund One on like January 2nd, 2019.
So, what was that transition then like for people listening that are curious how raising a venture fund goes? Because I think a lot of people kind of think, oh, this guy has like millions of dollars and just investing his own money because he's like rich because his parents gave him money. Like how does it typically go raising a venture fund? Like maybe talk us through how that went.
Yeah. So it was basically like playing pinball with my eyes closed is how I'd describe it. Like in the sense that, you know, you have this goal in pinball, like get the ball in the hole, you know, and then you've got like the pins and you have to position the pins to get the ball, but you have no idea how to position the pins because your eyes are closed. And that's like the best analogy I have for like what it was like, you know, meeting prospective like LPs of any kind that would be interested in 2018, like a solo GP, first-time venture fund manager, like early-stage focused, small fund by virtue of those features, and then also focused on Europe and then a little bit of North America.
And then AI too, like AI was not hot also. It's like you should be doing VR or crypto maybe.
