Nathan Benaich on the State of AI: 90% of research uses Nvidia, compute index up 100x in 3 years, and EU's $1.1B plan is a house for ants

of our major distributors talking about watches um because he was super into them. Anyway, our next guest is already in the ream radio waiting room. We have Nathan from Airirst Street Capital coming in to the studio. Nathan, good to see you. How are you doing? Welcome. Great. I was worried it's light where you are.

I was worried I was going to keep you up late because you're usually in Europe. Uh give us a little introduction on yourself and and backstory and then tell us where you actually are if you can. Incredible jacket too. Jackets are a great sign of respect in our culture. It is nice jacket. Yeah. Thank you. Thank you.

It's life at the AI frontier. Yes. Um yeah, so I founded a firm called Air Street Capital uh in Europe. I invest in AI companies at the very early stage. I've been doing this for about 10 years now. Um and spent kind of half my time in the US and in Europe. Um, thank you. And you're in the US now? Yeah, I'm in New York.

Oh, nice. Very nice. Yeah. Uh, well, on vacation or investing? No, I've actually been doing a bit of a state of AI uh launch tour, if you will. So, we did a meet up in San Francisco last Thursday for the report drop. Yeah. And then did one in New York last night. And what is the state of AI? Yeah.

So, before we get into that, I have one is AI good? Yeah. One question. How how competitive are round AI rounds at the early stage like preede seed uh today?

I know I'm sure they're getting more competitive, but I think historically if you could write a million-doll check at the idea stage, you you you had a uh quite a bit chance to build a real position. Yeah. Yeah. Yeah. I mean, in that sense, I think the game has like materially changed.

um there's a bit of a bifurcation like you have the sort of large lab model companies that are raising you know significant amounts of money and I think those are also potentially good investments but they uh they definitely cater to a certain type of venture firm.

Um, and then you have the other category which is potentially more pragmatic entrepreneurs going for less of the general purpose AI but more looking at kind of what use case can they solve with a general purpose AI system and uh and those ones are a little bit more modest but in general yeah you've seen you know inflation of of round sizes but also actually interest um from customers to buy this stuff and so founders just want to take a bigger swing want to go faster and that's what's driven a bit of the bit of the round size increase I feel like a couple years ago, the meme was even if you're a seedstage company, as long as you slapped AI on there, you could justify like a hundred million dollar raise because who knows, we might need to train a model.

We might be doing some real AI, uh, not just a rapper company here. Has that dissipated as it's become clearer that to win in that game, you need hundreds of billions of dollars.

And so your like your actual cost structure should look a lot less like I see GPT4 as like a capex expense basically like a huge training run and for most AI companies it's going to look a lot more like opex until they have a customer they don't really have a big token bill and so why do they need $50 million at at seed?

I ask myself this question a lot.

uh at the end of the day to some degree you know as an investor you're not the one deciding this you know the market sort of pumps what they will um but I I think the I think the biggest um frame change and you can see this documented in the state of the report the last eight years that have been doing it is a lot of the capabilities that you get out of the box of a model today whether it's you know language reasoning image video biology um if you showed that to you know me or others 10 years ago we would have probably said this is magic and like not possible.

Uh and so I think there there is just so much you can do with what's available today that it's almost a gift to the startup community that now they don't have to blow you know their three four five million seed budget on trying to build an AI system collect the data and then you know throw it over the parapet to a to a puritive customer who says you know what I don't care I've got other priorities what is this AI thing doesn't even work anyway.

Yeah. Um so now you can run your whole seed budget on product market fit experiments.

So it's almost like back to kind of OG seed investing which is less about you know what the initial starting idea is but how capable is a team to just uh run a lot of experiments and find something that works and then uh and then you know rip it when it does. Yeah.

What uh what's the headline KPI that you like to center on in a state of AI report? I mean, you can focus on market caps. You can focus on revenue growth. You could even look at uh gigawwatts and and electrical numbers. You could look at number of H100 equivalents that are being built out.

Like there's so many different ways to ground the shape of the AI buildout. Like what do you like to focus on? Yeah, frankly, I mean the report is about 300 pages, so the honest answer is all the things.

Um but uh but so what we started doing uh a few years ago is this compute index where I basically combed around the internet to try to find uh all documented cluster sizes for various companies and we've been charting just how much of an expansion uh you know has happened.

Um you know there in the last 3 years it's basically probably like a 100x increase. Mhm. Um the other cool thing we've done is like mine all open source uh research papers and programmatically determine which chips are used in those papers.

So we can give people uh like a leading edge view for what researchers like in terms of hardware. And so with that we've shown I think the only evidence to this uh to this I've seen which is that 90% of AI research papers rely on an Nvidia chipset. Yeah.

And through that kind of work, we've seen, you know, AMD start to grow a little bit as well. Um, out of nowhere, of course. It does seem like Nvidia or AMD is really picking up this year. We've seen this in Cluster Max from semi analysis.

Uh, they kind of did their, I don't know, year of efficiency or in some ways and started squashing bugs.

And then there's just immense pressure when there's, you know, hundreds of billions of dollars going into the next the next phase of the buildout that uh, you know, uh, Nvidia's margins are going to be in the in the crosshairs for sure. Yeah.

And strategic alliances with uh, with OpenAI that I think is is quite interesting because there's uh, you know, if you read through the uh, through the news, you'll see that actually OpenAI has warrants in in in AMD. Yeah. Uh at a strike price of 600.

And so there's like some interesting incentives there to try to, you know, get AMD stuff to really work.

Um in some degree it's not as similar to like the relationship, I suppose, between Anthropic and AWS where Trrenium uh and and AWS's hardware is clearly behind um you know, the bigger competitors like Nvidia and so they get sort of like it help desk support from uh uh from Enthropic to make the stuff work. Yeah.

Uh do you think we're in a bubble? Um there's there's probably local bubbles, you know, all the time. Uh I think in like our little like ecosystem of of AI, uh there could be one in this like capex um you know, build out and all the interconnectingness um with various companies.

Um I think as soon as you see like Wall Street Finance get excited about something, you have to think a little bit why they're getting excited. Yeah. Um but you know at the end of the day I think um the outputs of this bubble if there is one are really really productive.

I mean back to the point around like this technology is magic. There's there's so much value that's getting created and we've documented some of this with uh with our mutual friends at ramp looking at how much uh your you know US companies are spending on AI services and that's going up the retention is going up.

You know, some of our investments in like 11 Labs and Cynthsia and others are just like growing because it's genuinely useful stuff. Yeah, it feels very different than What about What was your reaction to that Deutsche Bank report on Chad GPT spending growth?

I was reading I looked at it and it felt hard to judge because we're looking at summer basically. Yeah. And um but how did how did you uh process it? I mean, yeah, astronomical numbers.

But I guess the question is like is like is uh is the is the TAM for paid AI subscriptions in the consumer or proumer realm uh saturating? And I mean we're certainly seeing messaging from OpenAI that says we want to do agentic commerce, we want to increase the monetization of our free users.

We want to get into advertising potentially affiliate revenue API business.

like it it doesn't just feel like there's an unlimited pool of people that will pay $200 a month for, you know, a a great product, a magical product, but uh there just might not be that many people worldwide that are like sign me up for the $200 a month plan. Yeah.

So, I mean along that notice, I was also interested in this topic.

So uh I surveyed about,200 AI practitioners uh for the report and all this data is open access through it and I was really surprised because 95% of the people who took the survey uh use AI in their personal life and their professional life and it's like 75% of them pay out of their pocket for it and 10% of people pay 200 bucks a month or more.

Wow. Um and uh and like this is a pretty you know educated crowd in North America and in Europe. Um but I was surprised at those numbers. And then the second thing to your point around like agent commerce, I think that's huge.

Like we have some data in the report that shows that um uh conversion of e-commerce that's derived from kind of agent conversations converts at a higher rate than uh direct. Mhm.

Uh and so you can imagine that like the the brands that sort of jump on this earlier and create content that agents are using to uh to best learn like what the right solution is to serve to their like uh you know human uh uh like the human that they're serving are actually going to probably grow their revenues faster.

It's like this recursive cycle particularly as you know more and more labs and companies want to do this like reinforcement learning to improve um improve personalization. Yeah. Yeah. So, I think there's a lot of like adjacencies that one can uh one can navigate into.

Actually, uh a week or so ago, the EU announced their 1. 1 billion plan to ramp up AI and key industries.

Uh what was your uh obviously people in uh the west coast were mock you know kind of mocking it uh considering that you know the whole the whole plan was less than uh same Alman's carpet uh you know their meta's recent uh post single but um but yeah I my my my reaction was like that's that's kind of the the takes were funny but like at least like there is a plan at least there is evidence of right um and I I guess like when you read it and reacted to it, I guess like what was your immediate thought on what you wish the EU was doing?

Yeah. I mean, my reaction wasn't too similar. Uh there's been like a meme kind of bandied around in some of our friend group which is back to the Derek Zoolander like is this a house for ants? Yeah, exactly. Is this a plan for ants? But like does but like like Europe did not build their own Google and it was fine.

I mean they they like missed out on it. But like the natural state of the market equilibrium was that Google was built in America and America absorbed the market cap and European investors could go invest in Google if they wanted to and Google eventually built data centers to serve Google products locally.

But there was never like a we need a we need our own thing. I is is the thinking around like national AI changing because I feel like there's a world where even if you're using some like locally trained model like if you're going to chat. com like it's still open AAI wins. Yeah.

Uh so there was uh an initiative I suppose I'll call two examples. So there was one in Europe uh which called Quero which was like a government um kind of funded initiative to try to build a search engine that best represented European interest because allegedly Google wouldn't because it's an American company.

And obviously no one knows about it because that flopped. And then you know the second example was the French government not wanting to buy Palunteer. Oh yeah. For many many years as you guys will know pretty well.

And uh and eventually they caved in after burning the same amount of money they'd probably have spent on Palunteer and got it anyway. Mhm. Um and yeah, the latest example of all this um uh sort of infrastructure buildout in the US and in China is now every country wants to be sovereign with regards to AI. Yeah.

So so here's an interesting and like like is Putin the least AGI pilled world leader. Like I I I was trying to get research on how his clust on their clusters over there. They were like, "We have 800 GPUs and we're running them hot. " And I and so like they're basically calling the world's bluff on AI.

We'll be fine without did say that he actually wanted to have AI supremacy about a year or two ago, but but admittedly maybe there's less evidence of that. I mean to some degree also like you know Yandex was a pretty amazing Russian search company and then they've since le left and relisted on the NASDAQ as Nebius.

Nebas. Yeah. and one of the clouds that's really crushing it because they're they have really good technology. Yeah. Um but yeah, I think this the sovereign thing is a little bit of uh like a meme or a misnomer.

I mean, I think it's great for um the suppliers to the technology because they have yet another like TAM to sell to.

But, you know, if you're a non-American nation and you buy Nvidia chips and you buy American technology and, you know, Chinese technology and you just install it in your country, I don't necessarily see how that makes you sovereign because these things need updates. They can be switched off basically. Yeah.

They have a relatively short life too. That too. Useful life. Exactly. So, I I think sovereign is a little bit more of like a a an alignment of technology agenda and political agenda. uh than it actually is uh the true meaning of the definition of like we control our fate when it comes to intelligence and running it.

Okay. Moving forward, uh a lot of people are nervous about cracks in the global economy, cracks in the AI economy. Uh what are you monitoring? What like like what what are you uh what do you think is the most important metrics or data points or stories to be really digging into these days?

Yeah, I mean I think it comes back down to like the return on investment, you know, cases of companies becoming more efficient, launching better services. Uh I think the reason why I do this at all is because I think AI unlocks new use cases that weren't possible before.

Um and uh and so for so long as we see those examples, you know, propagate through the economy and it's still very early, um I think we're we're in a good shape.

And we only talked about software here but there's so much in you know national security and defense that's uh that's uh that needs to be solved and many of those problems whether it's like you know electronic warfare autonomy rely on AI and so huge town there also in Europe and we've got plays there with uh you know with Delhi and a few others uh companies in Europe.

So uh I think it's like uh it's full speed basically.