# The Supply and Demand of AI Tokens | Dylan Patel Interview

https://www.youtube.com/watch?v=LF3aUIM57uw

[00:00] What used to matter a lot was execution.
[00:02] was very, very difficult and ideas were cheap.
[00:04] Now, ideas are cheap and plentiful, but execution is very easy.
[00:07] So, really only the good ideas are the ones that can justify the spend on super cheap implementation.
[00:30] You told me this incredible story about how your own team's use of tokens has changed dramatically this year.
[00:36] Yeah.
[00:36] Will you tell that story and what it is teaching you about what's going on in the world?
[00:41] Last year, we thought we were heavy users of an AI.
[00:42] Everyone's using ChatGPT, everyone's using Claude, everyone's got, you know, I'm providing whatever subscriptions anyone wants.
[00:49] On the order of spend of like tens of thousands of dollars for our firm.
[00:53] This year, the spend is just skyrocketed.
[00:56] And it really started in late December with Opus that included Doug, who's
[01:00] president, Doug O'Laughlin.
[01:02] He's very much like leading the charge in the sense of like non-technical people using AI for coding.
[01:06] Um and so, he's basically pulled the whole firm slowly over time.
[01:13] I think he's been the the leader in doing that.
[01:14] Obviously, the engineers were using it anyways, but spend in January just started to inflect and rocket and rocket and rocket and rocket.
[01:22] Um we signed, you know, an enterprise contract with Anthropic and it's gone to the point where now, um I think when I last talked to you, it was 5 million spend rate.
[01:29] It's actually 7 million spend rate now.
[01:31] That was last week, by the way.
[01:33] A lot of that is just the usage, right?
[01:35] What's What's really, you know, people people who are have never coded before are using Claude code and spending thousands of dollars sometimes a day.
[01:44] But, across a firm, we're spending $7 million a year now on Claude code at the current rate um versus our salary expense being in the neighborhood of $25 million.
[01:53] So, you know, we're north of 25% of spend on Claude code as a percentage of salary.
[01:59] If this trajectory continues, then you know,
[02:02] we'll spend more than 100% by the end of the year.
[02:04] Which is a bit terrifying.
[02:07] Thankfully, I don't have to decide between people and AI because our company's growing so fast.
[02:11] It's, you know, more so like, okay, well, I don't have to hire nearly as fast and I can spend a lot more on AI and it works and we just grow faster.
[02:17] But I think other folks will start to reckon with the fact that huh, if this person can do the work of five to 10 to 15 people uh using Claude code, then all of a sudden, I should probably cut people.
[02:29] But right now, I think the use cases are so broad.
[02:34] For example, one thing is we have a reverse engineering lab in Oregon that we've been building for a year and a half.
[02:37] We have a bunch of, you know, fancy microscopes, scanning electron microscopes.
[02:41] The whole purpose of this is you reverse engineer chips.
[02:43] You get the architecture out of it.
[02:45] You get the materials that they're using to manufacture and this is some of the data we sell.
[02:49] This is a very slow process of analyzing that data.
[02:53] Instead, um one person on the team, they've been able to spend with a couple thousand dollars of Claude tokens, they've been able to create this application that is GPU accelerated, runs on a server that we have at CoreWeave, and anytime we
[03:03] send it an image, it's able to take the picture of the chip and overlay where every single material is.
[03:08] Oh, this part is copper.
[03:11] Oh, this part of the gate is uh tantalum.
[03:12] This part of the gate is germanium.
[03:13] This part of the gate is cobalt.
[03:15] And so, you can do a finite element analysis of the entire stack up of the chip very, very quickly, visual with a dashboard GUI.
[03:22] It's everything. Few thousand dollars of Claude.
[03:24] The person previously worked at Intel and he said that was an entire team's job to build that and maintain that.
[03:29] Now, rack that up across, you know, the entire firm, it's it's insane.
[03:33] Another example that I think is super fun is Malcolm.
[03:36] He's an economist at a major bank before.
[03:39] Um their economist department was like 100 or 200 people.
[03:44] What he built was the most incredible thing ever.
[03:46] He piped all of this different data, you know, FRED data and all these other data sets, right?
[03:51] Employment reports and all these other things from various APIs.
[03:53] We signed a couple of contracts with folks to get API access to data.
[03:56] Pulled it all in, started running regressions, started looking at the impact of various economic revolutions on the economy.
[04:03] Um from a deflationary inflationary
[04:05] perspective.
[04:08] The BLS has this entire um Bureau of Labor Statistics has this entire like set of like 2,000 tasks.
[04:12] And so he did that with AI.
[04:14] Which ones can be done by AI, which ones cannot, and grading them across a rubric.
[04:17] You know, about 3% are doable now with AI.
[04:20] Um and so he's created this like metric so that you can measure things that can be done by AI, what what the massive deflationary uh you know, what the cost of being able to do those with AI and therefore the deflationary aspect of it.
[04:31] You know, output can go up.
[04:32] It's called phantom GDP is what he's called it.
[04:33] Phantom GDP.
[04:33] Output can go up, but cuz cost falls so much actually GDP theoretically shrinks.
[04:38] So he created this whole analysis and a brand new benchmark of uh language models.
[04:40] Um a set of evals across 2,000 different evals, right?
[04:45] He did this all by himself?
[04:47] He did this all by himself, yeah.
[04:48] And he's like, "Dude, this would have taken the team of 200 economist a year."
[04:53] And he's just like he's like completely cracked out on Claude.
[04:54] He's like, "Everything has changed."
[04:56] How do you think about it as a business owner going from close to zero to 25% accelerating towards whatever percent of total spend?
[05:03] Like at what point are you like, "Whoa, I need to put the brakes on
[05:06] and be careful how much we're spending.
[05:09] Maybe we don't need to spend on the most cutting it on Opus 4.7 which came out today.
[05:12] Maybe I can throttle it back to something that's a little bit cheaper.
[05:15] Ultimately, like I'm in the information business, right?
[05:17] That that is, you know, we sell analysis to sell.
[05:18] We do consulting.
[05:20] We create data sets.
[05:22] I don't see why this wouldn't be completely commoditized on a pretty rapid basis if I'm not constantly improving.
[05:29] My first product that I was selling as a data set actually it is, you know, like there's more people trying to do it.
[05:33] Now we've made it constantly better and better and better and more detailed.
[05:35] And so therefore it sells a market.
[05:36] But the way we were doing it in 2023 is not terribly different than, you know, it it it's it's basically what everyone else is doing now.
[05:45] If I don't move up the bar, then I will be commoditized.
[05:48] If I don't move fast enough, I will also lose my edge.
[05:50] So, the question is, yes, AI commoditizes things, just like it commoditizes software.
[05:57] Those who can move fast and keep control of their customers and keep providing them an awesome service and keep improving the service won't shrink, they'll grow.
[06:03] They'll grow faster.
[06:07] Those who are incumbent and not doing anything, they're going to lose.
[06:10] And so, it's a bit of an existential, like if I don't adopt AI, someone else will and they will beat me.
[06:16] Uh and another easy example is the energy space.
[06:18] So, we've had a few energy analysts for a couple for like a year now.
[06:21] We've been trying to build out this energy model.
[06:22] It's very complex.
[06:24] Energy's data services market is something like $900 million.
[06:27] So, obviously a huge market for me to try and break into, but it has, you know, we really hadn't broken into the energy data services business despite a year of having multiple people on the team.
[06:34] Um then cloud code psychosis hits.
[06:38] One of the people who leads the data center energy and industrial sort of business at Sentiment Analysis, uh Jeremy, it hits him.
[06:44] And now all of a sudden, in 3 weeks, um he spent a lot.
[06:47] He was spending like $6,000 a day.
[06:48] It was an insane amount, but he scraped every single power plant in the US, every single transmission line above a certain voltage, um and created this entire mapping of the entire US grid, as well as a lot of demand sources, all from various public sources of data.
[07:03] Um and we've shown it to and and we built and it's got like this dashboard where you can view and
[07:08] check.
[07:10] You can see all the micro regions of the US where there's power deficits and surpluses.
[07:12] Um all of these details, built in a handful of weeks.
[07:15] We started showing some of our customers who buy our data center data set, but our energy, like traders, we showed some of them and they're like,
[07:23] "Wow, how long did this take you?
[07:24] This is really good.
[07:26] This is better than XYZ company."
[07:27] Or and then we like get dig deeper.
[07:29] XYZ company has 100 people and have been working on this for a decade.
[07:32] Now, obviously our thing is not fully robust as robust, but in some ways it is better.
[07:35] I'm going to commoditize these energy services companies, data services company.
[07:38] Who's going to come commoditize me if I don't move faster?
[07:40] And so, the question from a business owner's perspective is, "Yeah, I'm spending a lot, but what is that spend getting me?
[07:45] Is it getting more revenue?"
[07:47] Yeah.
[07:49] >>> [music]
[07:50] >>> Most software companies try to maximize your time on their app to juice engagement.
[07:53] Ramp does the exact opposite.
[07:55] Ramp understands that no one wants [music] to spend hours filing expense reports, reviewing expense reports, and checking for policy violations.
[08:01] So, they built their tools to give that time back [music] using AI to automate 85% of expense reviews with 99% accuracy.
[08:05] And since Ramp saves
[08:11] companies 5%, it's no wonder that Shopify runs on Ramp, Stripe runs on Ramp, and my business does, too.
[08:16] To see what happens when you eliminate the busy work, check out ramp.com/invest.
[08:22] OpenAI, Cursor, Anthropic, Perplexity, and Vercel all have something in common.
[08:26] They all use WorkOS. And here's why.
[08:28] To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SCIM, RBAC, and audit logs.
[08:34] That's where WorkOS comes in.
[08:36] Instead of spending months building these mission-critical capabilities yourself, you can just use WorkOS APIs to gain all of them on day zero.
[08:43] That's why so many of the top AI teams you hear about already run on WorkOS.
[08:48] WorkOS is the fastest way to become enterprise ready and stay focused on what matters most, your product.
[08:54] Visit workos.com to get started.
[08:57] Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready work using your firm's own templates, contacts, and standards.
[09:04] Send Felix an email like, "Take these comments and turn them for me."
[09:08] Or, "Update my tracker with the context of these emails."
[09:11] Or, "Run the ability to pay math on this buyer."
[09:13] And Felix sends back finished PowerPoint decks, Excel models, and sourced research.
[09:19] Felix works the way your team already does, delivering work quickly and accurately around the clock.
[09:22] Learn more at rogo.ai/felix.
[09:27] Are you worried that in the limit, the people that control capital and investing capital were often hiring you for for what you do, will just say, "Well, we have analysts, too, who are really smart about this."
[09:36] Like, we'll just build this ourselves.
[09:37] Like, if it's getting that easy, at what point does it just all pool into the investment firms that stand to gain the most cuz they have the most leverage on top of the data or the insights that that they glean.
[09:48] First of all, any information services business, obviously, I don't generate as much value as my customer does from said information.
[09:56] Uh because if I sell you information for a dollar, you're only buying it for a dollar because you know that information helps you make a decision that lets you make more than $1.
[10:03] And so therefore, you have you have you you have made more money off of me than I did from the information myself.
[10:09] An investment fund, these investment funds all have their
[10:12] own information services, you know, especially like the super like the Jane Streets of the world and the Citadels.
[10:17] They're they're really detailed on their data.
[10:18] And yet, um these sort of folks also purchase data from us and continue to do so and continue to grow with us because I think there's just some some it factor, right?
[10:27] We move faster, we're more nimble, we're a smaller team that's focused on just one specific thing, uh AI infrastructure and and the huge revolution that causes in AI, um and tokenomics and all these things.
[10:40] And and we sort of really see where it's headed and so we're moving faster and building faster.
[10:47] Um I think investment professionals just would you know, yes, they'll try and build some of the stuff we do and um more likely they'll just buy the data from us and it's cheaper for them to buy the data from us and then to build and then build on top of it than it is to build it themselves.
[11:00] But ultimately, some may try.
[11:02] I feel like every conversation I have with you, what I'm always getting at is just supply and demand of tokens.
[11:06] Like, that's the thing that's interesting to me in the world right now.
[11:09] What has this experience taught you about the demand?
[11:11] Has it changed your
[11:13] view on the demand side of that equation?
[11:14] Just feeling it viscerally yourself?
[11:17] If we take a step back and look at the macro lens, right?
[11:18] Anthropic has gone from 9 billion revenue to what?
[11:20] They're at 35, 40 billion now.
[11:22] Probably by the time this airs, 40, 45 billion.
[11:24] Who knows?
[11:25] AR.
[11:27] Their compute has not grown to the same degree.
[11:31] Um and if you do the calculations and you assume they didn't decrease their research and development compute, they clearly didn't they released they have mythos, they have office 4.7.
[11:37] So they clearly didn't decrease their research compute spend.
[11:42] So ultimately what they've done even if you assume all incremental compute they've gotten is gone towards inference, their margins are at a floor of 72%.
[11:49] In reality some of that incremental compute they've got probably went to research and development it maybe higher than 72% gross margins.
[11:57] To be clear at the start of the year they started there was a there was a leak by someone from their funding some some of their funding round docs someone leaked it.
[12:03] 30 something percent gross margins.
[12:06] Where on earth does a business like this grow margins like that?
[12:08] And it's in principle right?
[12:10] Their demand is so high they're able to cut back on usage limits, rate limits, all these things.
[12:16] What really matters is having an anthropic rep and having an enterprise contract with them and getting the rate limit increases that you need because otherwise tokens are ultimately super super in demand.
[12:25] Whoever whoever can pay for them anthropic has the same problem right?
[12:29] Like I mean not problem it's it's just the reality of how capitalism works.
[12:32] Yes, people are spending sending them 40 billion dollar ARR in tokens and but those tokens are generating way more than 40 billion dollars in value.
[12:40] Various businesses will have different value generation per token but as we get more and more intelligent what really matters is access to these most intelligent tokens and leveraging them at things.
[12:50] You as a person deciding what is the best way to leverage these tokens to grow business and generate value.
[12:57] Because a lot of folks will want tokens and generate tokens but the shitty SAS startup in in SF who is using Claude to generate you know their software product is not necessarily actually creating a ton of value and therefore they're going to get priced out of tokens.
[13:12] Soon enough.
[13:15] Are you at all surprised that I I had this
[13:16] experience just today where on the flight here I got rate limited out on something.
[13:20] I saw 4.7 came out and what I immediately wanted was like to be on 4.7 that second and I was all I it just I couldn't think about using 4.6 anymore now that this 4.7 is out.
[13:30] I was perfectly happy with 4.6 for the last many weeks.
[13:34] It's amazing.
[13:36] Are you surprised that people are so insistent on going to the most expensive leading edge thing to the degree they are?
[13:42] Without a doubt.
[13:44] One of my funniest memories in the past month and a half is myself and a buddy of mine Leopold being on our knees in front of an Anthropic co-founder begging him for access to Mythos and then pretending it doesn't exist.
[13:58] [laughter]
[13:58] Cuz we knew it existed we're like please give us access and he's like I don't know what you're talking about.
[14:03] What was your reaction to that rate card or that eval card coming out?
[14:07] It was rumored in the Bay Area you know, we sort of like knew it was supposed to be really good but um if you just look at the benchmarks and obviously benchmarks change over time Mythos is
[14:17] potentially the biggest step up in model capabilities in like 2 years.
[14:21] I think that's really really an an important detail that you know, and it it's so good that they're like don't want to release it even though they're they they already announced the price to their people that they did a selective release for cyber foreign it's like five or 10 x the token cost.
[14:35] They just don't want to release it because they're worried about the like impact on the world and they're releasing a shitty worse version Opus 4.7 to us and they explicitly said in the model card hey we actually preferentially made it worse at cyber.
[14:47] I don't know if you read that.
[14:49] Whoever you are if you have enough capital you should get a freaking enterprise cloud enterprise Anthropic subscription where you pay per token not with these like subscriptions cuz then you won't get rate limited much and then you must you need to figure out how to leverage those tokens to the highest value task um and make money off of it because ultimately what you're doing maybe maybe like a year from now or 2 years from now the business is actually just arbitraging tokens, right?
[15:13] The tokens are amazing but let's figure out what direction to point them in and then 3 or 4 years from now the model will know, you know, what
[15:17] to do with the tokens and how to make the most value.
[15:20] You know, you can you can look at this retroactively pick any benchmark.
[15:24] The cost to hit a certain capability tier used to cost X and now it cost 1/100 or 1/1000 of that.
[15:31] Deep Seek, for example, on GPT-4 was 1/600 the cost.
[15:36] And since then, the cost have fallen further for GPT-4 class models.
[15:39] Of course, no one gives a crap about GPT-4 class models.
[15:43] They want the frontier because the frontier lets them create the economically valuable things, but GPT-4 class models can still be used in like stuff.
[15:48] And so, people are using them in some like tiny use cases.
[15:51] It's just the cost have fallen so fast.
[15:54] It's it's not really what's driving the demand.
[15:55] What's driving demand is is all these new use cases.
[15:57] Yeah, current 4.6 Opus or 4.7 Opus tier models a year from now, my spend for the same exact quality of the model would probably be like 70k.
[16:11] I bet you it'll be a 100 times cheaper.
[16:14] Irrelevant because I'm going to be using a way way way better model which can do way way better things.
[16:17] Uh Anthropic
[16:19] Mythos is more expensive as a model, but it spends a lot less tokens to do the thing.
[16:22] And therefore, it is actually cheaper in most tasks than 4.6 Opus because it's just way more efficient even though each individual token is smarter.
[16:31] When I last saw you, Mythos had just come out, maybe the day before or something, or the the card had just come out.
[16:37] And you said something like uh it actually made you feel like a little scared.
[16:39] It was so good.
[16:40] What did you mean by that?
[16:42] Anthropic's whole like goal in 2025 was and and even a lot of 2024, they're like, "Hey, by the end of 2025, we need an L4 software engineer uh in our model."
[16:53] And and they by and large achieved that with 4.6 Opus.
[16:56] What they didn't say is that, you know, and and if you look at Mythos and if you compare like the benchmarks, it's like an L6 engineer.
[17:01] So, L4 is like pretty new.
[17:04] L6 is like quite well experienced.
[17:07] I think Anthropic said that the model internally was available in February.
[17:11] So, in 2 months, they've gone from L4 engineer to L6 engineer.
[17:15] Uh what's next?
[17:18] Um you know, when you when you think about
[17:20] the model progress, it's only accelerated.
[17:21] Anthropic's release cadence has compressed.
[17:23] Open AI's release cadence has compressed.
[17:24] Why?
[17:26] Because these models generally to make a better model, you need a few things, right?
[17:28] You need amazing compute.
[17:29] Compute is very expensive and it has a time scale that we, you know, we track and it's like, you know, it's growing, but like you know, it's it's sort of set in stone for the next, you know, short short term.
[17:37] It's like kind of set in stone what you've already signed.
[17:40] Um and there will be delays and shifts and some somehow you can find a little more, but it's generally pretty set in stone.
[17:45] There's amazing researchers that people are paying tens of millions of dollars for.
[17:46] And then lastly, there's implementation.
[17:50] Historically, it's been very difficult.
[17:51] If I have an idea, now I have to implement it.
[17:53] Implementing is hard.
[17:54] Now, ideas are there.
[17:58] Implementation is very easy.
[17:59] It's expensive, but it's very easy.
[18:02] So, how do you how does one decide what ideas to implement?
[18:04] And it turns out if your implementation is just so much easier, now you can just implement more ideas and move on the treadmill faster and faster and faster.
[18:12] Whether that is AI model research and scenario model release cadence has shrunk to down to two months from where it was six months before or
[18:20] hey, I want to I want to take every power plant in the US and every transmission line and model it and run regressions and see the micro supply and demand.
[18:26] I can also do that.
[18:28] The idea is cheap.
[18:29] You know, which idea makes sense, which idea is worth the capital that you have to spend on the tokens cuz the implementation is there.
[18:35] It's It's that's the I think the key learning and if implementation costs continue to tank, which they are.
[18:43] Um we don't even have meet those yet.
[18:45] It's only been, you know, a handful of hours since Opus 4 7 launched, but you know, my team is pretty excited about it internally.
[18:50] What now comes to the world?
[18:53] Uh it's a complete reordering of how like economies work.
[18:55] What used to matter a lot was execution was very, very difficult and ideas were cheap.
[19:01] Now, ideas are cheap and plentiful, but execution's very easy.
[19:05] So, really only the good ideas are worth are the ones that can justify by spend on super cheap implementation.
[19:11] So, are you actually scared or are you just is it just does it just introduce some uncertainty that's hard to grapple with?
[19:18] Uncertainty is there, um, but I do I do
[19:22] think that causes some fear in terms of how does society reform itself?
[19:27] How does one exist in a world where actually any, you know, your ability to implement something is not actually that important.
[19:36] Your ability to just choose the correct idea for AI to implement and then your ability to sell that idea or sell what the AI has implemented is what matters.
[19:45] Your ability to garner capital towards that is what matters.
[19:48] And going back to the point of like it's very important to have the newest model always.
[19:51] Who's going to have access to the newest model?
[19:53] Anthropic's project, I know it's not called earwig, but I troll Anthropic people by calling it earwig.
[19:57] Um, glasswig.
[19:59] Anthropic earwig, you know, where they only release mythos to certain companies for cyber, that's just going to be something that continues.
[20:07] Models will have less broad and less broad deployment.
[20:09] I know I know OpenAI and Anthropic and all these people are like, "We want to have great AI for everyone."
[20:15] AI is very expensive.
[20:18] Who's going to pay for the trillion dollars of infrastructure?
[20:19] People who have money and can can build useful
[20:22] things with AI. And then you don't want
[20:24] people to distill your model, so you
[20:25] don't release them broadly. Uh, you
[20:27] release them to a fewer and fewer set of
[20:28] customers. Those customers are also now
[20:31] wrestling over the tokens unless
[20:32] Anthropic jacks the you know, they could
[20:34] double their pricing on Opus and I would
[20:36] continue to pay and I bet most users
[20:37] would continue to pay.
[20:39] I bet that wouldn't solve their
[20:40] humongous capacity problem they have.
[20:42] So, then the question becomes where does
[20:45] this cycle end where, you know, token
[20:47] usage and therefore the benefits of
[20:49] those tokens, the additional value
[20:50] generated on top of those tokens
[20:52] aggregates among fewer and fewer and
[20:54] fewer companies? I don't have mythos.
[20:56] You know who has mythos? Top freaking
[20:58] banks. Um, now they're only using it for
[21:00] cyber security, but at some point I can
[21:02] envision a world where, hey baby, I,
[21:03] because I have an enterprise Anthropic
[21:05] contract and because
[21:07] Anthropic people kind of like me,
[21:08] they're willing to give us like slightly
[21:10] earlier access or slightly higher rate
[21:12] limits or something for a model. I hope
[21:14] that's what happens. And then my
[21:16] competitor, whoever that is, doesn't
[21:18] have that and I'm able to crush
[21:20] them. There are people who are like Ken
[21:22] Griffin of Citadel is like super well
[21:23] connected and super rich and he's like
[21:25] he he just signs, you know, who knows?
[21:27] He goes and signs a deal with OpenAI
[21:28] Anthropic that's like, "Yeah, I'm going
[21:30] to get access to your models um and I'll
[21:32] buy the first 10 billion dollars worth
[21:34] of tokens each year. So whenever you
[21:36] release the model, you know, I'll spend
[21:38] the first 10 billion tokens and then
[21:39] everyone else can get the model after
[21:40] that." And it's like, "Okay, well, now
[21:41] what does that do?" Well, now he's going
[21:43] to crush everyone in the market. And so
[21:44] that's just an example. Could be cyber
[21:46] like Anthropic is worried about, "Oh,
[21:47] now I can hack people." Could be
[21:48] information services business like
[21:49] myself where I crush someone else. I
[21:50] think, you know, it it it's it's it's
[21:53] such a broad base. We don't know what
[21:54] these models can do. Anthropic doesn't
[21:55] know what these models can do. No one
[21:56] knows what these models can do. It's up
[21:58] to the end user to figure out where they
[21:59] can leverage the tokens to see what they
[22:00] can build and imagine, which is
[22:02] tremendously productive and uplifting
[22:04] for humanity, but then what happens to
[22:06] the concentration of resources and usage
[22:08] of it?
[22:08] >> Presumably right now robotics or robots
[22:11] consume relatively zero tokens versus
[22:14] everything else. Do you see what's your
[22:15] view of that? If that's like a second
[22:17] demand curve that could start to ratchet
[22:20] There's a new startup every single day,
[22:22] you know, within a mile of here trying
[22:23] to build something interesting in
[22:24] robotics.
[22:25] >> So there's this concept of software only
[22:26] singularity, which is that the world
[22:29] has, you know, AI singularity but only
[22:31] in software and now what about the rest
[22:33] of the world? Vast majority of the world
[22:35] is
[22:36] physical. You can see the world orient
[22:39] around hardware not software. That's
[22:40] actually why I think software only
[22:42] singularity is like just a blip and not
[22:44] like a you know, we we do get everything
[22:46] else because once software's super easy,
[22:48] what makes robots really hard? It's like
[22:49] programming microcontrollers and
[22:51] actuators and controlling all this stuff
[22:52] is very difficult. Right now the
[22:54] interesting thing thing about models AI
[22:56] models is they're actually really
[22:58] inefficient in learning. It's just we're
[23:00] able to give them so much data that
[23:02] they're able to learn and surpass us in
[23:04] certain ways. Robots currently the robot
[23:06] models
[23:07] VLAs vision language action models which
[23:10] is very popular right now is probably
[23:12] not going to be the thing that
[23:14] ultimately scales beyond they're
[23:16] inefficient in data
[23:17] and we can't scale the data for them
[23:19] fast enough. There is going to be some
[23:21] way to large scale pre-train robot
[23:23] models where just like humans see all
[23:25] this data throughout their lives and
[23:27] what's interesting is humans
[23:29] the reason why we're so good is we're
[23:30] sample efficient one example we're good
[23:33] and so applying that to robotics so once
[23:34] you once you have this software only
[23:36] singularity implementation is super
[23:37] cheap anyone can start to build these
[23:40] people can start to build models that
[23:41] now robots are actually useful and so I
[23:44] think in the next 6 to 18 months we'll
[23:46] start seeing real breakthroughs in
[23:47] robotics that enable few shot learning
[23:51] i.e. there's a pre-trained robot model
[23:53] and now there's a robot that you have
[23:55] hired or bought or whatever you showed a
[23:58] few examples and it's able to do it. You
[23:59] show it how to stack these two things or
[24:01] you tell it hey this can can actually
[24:02] like balance perfectly you know and it
[24:04] and it starts doing these things.
[24:06] >> messy done
[24:06] >> [laughter]
[24:07] >> one shot
[24:09] No trust me I've spilled many a time.
[24:12] So I think I think robots will
[24:14] you get few shot learning right now you
[24:16] know there's a lot of companies doing
[24:17] robots for like you know advertisement
[24:19] or robots for like simple stuff like
[24:20] that but it'll be like oh folding
[24:22] clothes now but it's going to get really
[24:23] niche like robots just for cleaning
[24:25] chalkboards
[24:26] and it's a rental service or you know
[24:28] it'll be it'll be a model package that
[24:30] you download onto your standard robot
[24:32] that then does that right and and you
[24:33] pay for that you and anyways there'll be
[24:35] a huge explosion in physical good
[24:37] acceleration and and deflationary
[24:39] effects there but and so that's that's
[24:41] ultimately going to keep token demand
[24:44] going crazy and I I don't think token
[24:45] demand slows down personally.
[24:46] >> Did you learn anything else about the
[24:48] world based on Methuselah's results and
[24:50] how it was built? My way of asking like
[24:52] the you know if you break down the the
[24:54] components of of scaling laws like the
[24:56] pre-trained ones.
[24:56] >> is a materially larger model than prior
[24:59] models. And so, yes, it is a much larger
[25:01] model. Now, whether or not it's it's
[25:04] what chip it's trained on is not really
[25:05] relevant. It's the scale. And obviously,
[25:07] you know, to 100,000 Black Wells is
[25:09] equivalent to hundreds of thousands of
[25:11] prior generation chips. TPUs and Tranium
[25:13] have their different release cadence, so
[25:14] it's not exactly like mirrored
[25:16] one-to-one.
[25:17] Um but ultimately, yes, Mythos is a
[25:18] significantly larger model. It's proof
[25:20] that the scaling laws still work. Um
[25:22] everything about it shows that the trend
[25:23] line continues of models more compute
[25:26] into model makes model better. And along
[25:27] the whole way, it's not just more
[25:28] compute into model makes model better.
[25:30] Along the whole way, we're also getting
[25:31] these compute efficiency wins, which are
[25:34] you know, as as all this research
[25:36] compute that the labs are spending is
[25:37] actually turning into if I want X
[25:39] capability to your model every 6 months,
[25:42] that cost or every 2 months, that cost
[25:43] is dramatically decreasing. But then, if
[25:46] I scale it up massively, I get a
[25:47] humongous capability jump, as well. And
[25:49] so, yes, it's it's proof that this is
[25:51] still happening. Google and Anthropic
[25:53] are not heavy heavy users of GPUs on the
[25:55] training side, but OpenAI, they'll
[25:58] they'll start having their new class of
[25:59] models. I think they're taking a more
[26:01] sensible, principled approach to scaling
[26:03] uh in small steps. Anthropic really went
[26:05] for a huge jump. We'll see better and
[26:07] better models throughout the year. And
[26:08] the release cadence is only going to get
[26:10] faster. We've gone a long way in the
[26:11] conversation with saying almost nothing
[26:12] about OpenAI, which would have been so
[26:14] strange.
[26:15] >> So, so this is this is the interesting
[26:16] thing. Everyone's like, "So, okay, so
[26:17] Anthropic's just won, right?" You know,
[26:19] they had Mythos in February. They never
[26:21] even released it cuz they didn't feel
[26:22] the need to. They're already sold out.
[26:23] Their revenue's already adding 10
[26:25] million dollars a month. Um and then,
[26:27] you've got Opus 4 7 today, all before
[26:30] OpenAI's, you know, um alleged spud
[26:33] release, which, you know, media such as
[26:35] The Information and others have have
[26:37] posted about. So, clearly, Anthropic is
[26:39] in the lead, right? And OpenAI is
[26:41] cooked. What's interesting is because
[26:43] Anthropic has such bounds on compute,
[26:46] and they can only grow it so fast, and
[26:48] sort of to the point of, you know, you
[26:50] know, Daria Daria used to gloat about
[26:52] how OpenAI was being too aggressive on
[26:54] compute and Anthropic was more sensible
[26:57] in their scaling and now Anthropic is
[26:58] like, we should have I wish we had
[27:00] a lot more compute. OpenAI is able to
[27:02] pay the bills perfectly fine. In fact,
[27:04] they've raised a ton of money to get
[27:05] incremental compute
[27:07] in addition to the irresponsible levels
[27:09] of compute that they were buying from
[27:11] Oracle and Coreweave and SoftBank and
[27:12] all these people and Microsoft. Uh, you
[27:14] know, such as Trainium. Now they're
[27:15] getting Trainium as well from Anthr-
[27:17] Amazon. Um, so so they've done this like
[27:20] insane thing on compute and they need
[27:21] know they also know they need more. But
[27:23] what's interesting is if you were to say
[27:25] Opus 46,
[27:26] you know, let's ignore models getting
[27:28] better over time. Let's just take
[27:29] diffusion of this technology. You and I
[27:32] may get jump on the model immediately
[27:34] day one, but other businesses take time
[27:36] and they take time for people to learn
[27:38] and the spark of, oh Claude
[27:41] psychosis moment doesn't hit everyone at
[27:43] the same time. And so by the end of the
[27:45] year, let's say a 46 Opus tier model the
[27:47] economy would spend $100 billion on. I
[27:49] don't think that's unreasonable. It's
[27:50] spending 40 billion right now. That's
[27:52] like a linear extrapolation.
[27:53] >> It's a linear extrapolation, not a not
[27:55] an exponential. To get the exponential,
[27:57] you need the better models. Anthropic
[27:59] won't have enough compute to do that.
[28:00] And so and and presumably OpenAI and
[28:02] Google will hit that tier soon enough.
[28:05] Whoever hits that tier next, sure
[28:07] Anthropic may get to charge 70 plus
[28:09] percent gross margins, but if OpenAI
[28:11] hits it next, they charge 50% gross
[28:13] margins, they still get all of this
[28:15] incremental demand and probably they
[28:16] also won't have enough compute to serve
[28:18] all the users. And so
[28:20] sure maybe Mythos is a model where if
[28:23] the world had enough compute it'd be
[28:24] $500 billion of revenue or something
[28:26] crazy. There is such demand for these
[28:28] tokens and such limitations on compute.
[28:31] You know, and we see this with H100
[28:32] prices skyrocketing and the useful life
[28:34] of these GPUs continue to extend. It's
[28:36] pretty clear even the tier two lab is
[28:37] going to be sold out of tokens. Let
[28:39] alone the tier one lab. The tier one lab
[28:41] will have better margins,
[28:43] but the tier two lab will be sold out
[28:44] and probably the tier three lab will
[28:45] also be close to sold out. Economic
[28:47] value that the best model can deliver is
[28:49] growing faster than our ability to
[28:51] actually serve those tokens to people
[28:53] via the infrastructure. And so this gap
[28:54] will continue to grow and the model labs
[28:56] will continue to have expanding margins.
[28:58] Until people in the hardware supply
[28:59] chain infrastructure supply chain are
[29:00] like, wait, no, why don't I just jack up
[29:02] my margins? So suffice to say I think
[29:03] the assessment today or your assessment
[29:05] of the demand side is completely
[29:07] explosive in your own particular example
[29:09] here at SemiAnalysis, but just more
[29:11] broadly that as people fall in call it
[29:13] AI psychosis, as people fall into this
[29:15] experience of what they can do, the
[29:17] implementation difficulty going
[29:19] completely away. I've certainly felt
[29:20] that.
[29:21] You know, my own token spend is just
[29:23] through the absolute roof just in the
[29:24] matter of weeks. So that that feels like
[29:26] a pretty good assessment. Anything we're
[29:27] missing on the demand side?
[29:29] >> If you don't use more tokens, you'll
[29:30] never escape the permanent underclass.
[29:32] Just expand on that. So either either
[29:34] you use more tokens and you generate
[29:36] economic value outsize economic value
[29:38] for the use of those tokens. Um a lot of
[29:40] people are doing it the boring lazy way.
[29:41] Oh, I guess I'll just work one hour a
[29:43] day instead of eight hours a day and
[29:44] I'll have I do most of my job. That's
[29:46] the boring way. The cool way is I'll
[29:48] still work eight hours a day and I'll
[29:49] I'll do eight x the work and maybe I'll
[29:51] make five x the money.
[29:53] Um maybe not you can't do this with a
[29:55] job obviously. There's people who have
[29:56] multiple jobs. Um there's people who
[29:58] like start companies and start selling
[30:00] stuff. Get that economic value on on
[30:02] this AI before everyone is using it and
[30:04] it's table stakes. Uh because it's still
[30:05] not table stakes. If you don't use more
[30:08] tokens and generate the value from them
[30:10] and capture that value. These are
[30:11] there's three different problems here.
[30:12] Using more tokens, generating value from
[30:14] those tokens, and capturing value from
[30:15] those token uh from the value that you
[30:17] created from the tokens. Uh if you don't
[30:18] do these three things, you'll never
[30:20] escape the permanent underclass. I.E. as
[30:23] models continue to skyrocket in
[30:24] capability and the concentration of
[30:25] resources potentially happens. All
[30:27] right, let's talk about supply. What is
[30:29] going on? Like how would you describe
[30:30] the frontier of what's changing or what
[30:33] is changing at the frontier of supplying
[30:36] the entire stack that's required to
[30:37] serve all these tokens as the demand
[30:39] curve explodes? As demand skyrockets,
[30:41] prices are going up for everything on
[30:43] the supply side. Um whether it be the N
[30:45] GPUs,
[30:47] uh their prices are going up. In
[30:49] addition, their useful life is
[30:50] extending. H100 prices look like this.
[30:53] >> Yeah, exactly. There's people who have
[30:54] argued GPUs' full lives are less than 5
[30:56] years. Complete nonsense.
[30:59] Um there are clusters now re-signing 3
[31:01] or 4-year-old Hopper clusters re-signing
[31:02] for 3 or 4 more years. Um there's A100
[31:05] clusters that are re-signing for another
[31:07] couple years. So, the useful life is
[31:08] clearly not 5 years. It's maybe even 7
[31:10] or 8 years. Um arguably. We we don't
[31:13] know yet. We'll see. We'll see when
[31:14] Hopper gets there, but it it's clearly
[31:16] not 5 years. So, the useful life is
[31:17] extending and the prices are going up on
[31:19] that renewal.
[31:21] So, in effect, the gross margin was not
[31:23] 35% on a cluster. It's beyond that. Um
[31:26] so, margins are expanding in the in the
[31:28] cloud layer. Margins are um extremely
[31:31] healthy on the hardware layer with, you
[31:33] know, Nvidia still charging 75 or
[31:35] whatever percent gross margin. As we
[31:37] move down the stack, memory obviously
[31:39] margins have skyrocketed there. Places
[31:41] like optics and logic, there are large
[31:44] prepayments um and margins are growing
[31:46] slowly. Um more so the companies that
[31:49] are making chips like Nvidia are paying
[31:52] huge prepayments. So, in effect, the
[31:53] cost of capital or timing of cash flow
[31:56] or return on invested capital is going
[31:57] up even if the gross margin isn't. And
[31:59] you see this across the whole supply
[32:00] chain. You see ASML is completely sold
[32:02] out and they need Carl Zeiss to expand
[32:04] faster. Everyone along the chain
[32:07] everyone's either sold out and margins
[32:09] are going up or they're getting
[32:10] prepayments increases the return on
[32:11] invested capital cuz the invested
[32:13] capital is lower. And so, this is like a
[32:15] consistent trend across any part. It's
[32:17] it's even like, you know, a PCB. It to
[32:19] make a PCB requires copper foil. And
[32:21] that copper foil is sold out and people
[32:23] are making prepayments for it. It's like
[32:25] anything and everything that like has a
[32:26] pulse and is like sold out, people are
[32:29] like jumping to get more incremental
[32:30] supply and fighting over the supply for
[32:33] the years after. As your business scales
[32:34] up, everything gets more complex,
[32:36] especially your compliance and security
[32:38] needs. With so many tools offering
[32:39] [music] band-aids and patches, it's
[32:41] unfortunately far too easy for something
[32:42] to slip through the cracks. Fortunately,
[32:44] Vanta is a powerful tool designed to
[32:46] simplify and automate your security work
[32:48] and deliver a single source of truth for
[32:50] compliance and risk. There's a reason
[32:52] that Ramp, Cursor, and Snowflake all use
[32:53] Vanta. It frees them to focus on
[32:55] building amazing, differentiated
[32:57] products, [music] knowing that
[32:58] compliance and security are under
[32:59] control. Learn more at vanta.com/invest.
[33:04] I know firsthand how complex the tech
[33:05] stack is for asset management firms. And
[33:07] seemingly every new tool and data source
[33:09] makes the problem even worse, adding
[33:11] more complexity, more headcount, and
[33:12] more risk.
[33:13] Rydeline offers a better way forward,
[33:15] one unified platform that automates away
[33:17] that that automates away that complexity
[33:19] across portfolio accounting,
[33:20] reconciliation, reporting, trading,
[33:23] compliance, and more, all at scale.
[33:25] Rydeline is revolutionizing investment
[33:26] management, helping ambitious firms
[33:28] scale faster, operate smarter, and stay
[33:30] ahead of the curve. See what Rydeline
[33:32] can unlock for your firm. Schedule a
[33:33] demo at rydeline.ai.
[33:35] What do you think are the most important
[33:37] bottlenecks? Like typically in economic
[33:39] history, when there's this kind of
[33:40] demand, supply reorients and rises very,
[33:44] very quickly to meet the demand. It
[33:47] seems like it's almost impossible for
[33:48] supply right now in this moment to keep
[33:50] up. You know, famous last words, every
[33:51] every shortage is followed by a glut
[33:53] historically. But what are the most
[33:55] interesting bottlenecks to you on across
[33:57] the supply side?
[33:58] >> Supply chains are usually very fast to
[34:00] react. Um
[34:02] one unique thing is that our supply
[34:03] chains now are more complex than ever,
[34:05] and the things we're building are more
[34:06] complex than ever, and therefore the
[34:07] lead times
[34:08] are longer. Um and it's not like we
[34:10] haven't seen 18-month long lead times in
[34:12] other industries. It's just
[34:15] building incremental supply didn't take
[34:17] years. Um
[34:18] and this is the case with memory, right?
[34:20] Memory can only grow capacity, you know,
[34:23] low double-digit percentages a year,
[34:25] right? 20s, 30% a year, um even less for
[34:28] NAND, a little bit higher for DRAM. Even
[34:29] though the demand signal was very strong
[34:31] at the end of 2025, the memory companies
[34:33] immediately sort of started reacting.
[34:35] None of that incremental capacity really
[34:37] gets here until the second that they've
[34:38] decided to do in addition to the typical
[34:40] 20 to 30% you know, they can stretch a
[34:43] little bit, but really the true
[34:44] incremental supply doesn't come till 28,
[34:46] which is a very unique thing. Even if
[34:48] they wanted to build as fast as
[34:49] possible, it doesn't come till 28, early
[34:52] late 27 at best. And so, the result is
[34:55] memory prices have you know, gone
[34:57] through the roof. And guess what?
[34:58] They're going to double and triple
[34:59] again. Um at least on DRAM especially.
[35:02] People are like, "Oh, the memory story
[35:03] is overplayed. Everyone gets it." And
[35:04] it's like, "No, no, no, you don't get
[35:05] it. DRAM will double or triple from here
[35:08] still."
[35:09] Because
[35:10] that's that's how much capacity is
[35:12] required and they have to steal capacity
[35:14] from somewhere else and the only way to
[35:15] steal capacity from somewhere else in a
[35:17] in a capitalist economy is demand
[35:18] destruction via higher pricing. We're
[35:20] not like rationing stuff here. And so,
[35:22] ultimately that's what's going to
[35:23] happen. And so, margins continue to go
[35:24] up. Um
[35:26] I think logic also has humongous
[35:29] capacity problems too. TSMC just had
[35:30] their earnings. Uh they keep up in
[35:32] CapEx. Ultimately, you know, it takes
[35:34] them quite some time to build fabs. Um
[35:36] they're trying to do everything they can
[35:37] to squeeze every little output out of
[35:39] every fab that they have. But
[35:40] ultimately, they're not raising prices
[35:42] fast because they're good people it
[35:44] seems like.
[35:45] Um you know, single digit price
[35:46] increases instead of you know, triple
[35:48] digit price increases like the memory
[35:49] guys have had. And so, you ultimately
[35:51] have like this like market where you
[35:52] know, TSMC's a great company, but are
[35:54] they are they actually going to extract
[35:55] all the value? I mentioned things like
[35:57] copper foil, glass fibers for PCBs,
[36:00] lasers. These are things that are like
[36:01] well understood in niche supply chains,
[36:03] but they're very very tight. Um
[36:05] and ultimately upstream, the
[36:07] semiconductor wafer fabrication
[36:09] equipment supply chain is one that like
[36:10] I still think is it's gone up a lot, but
[36:12] it's still very under appreciated. TSMC
[36:15] CapEx this year they say 56. Uh we've
[36:17] had 57.4 billion since January.
[36:20] Um and we may up it slightly more just
[36:22] cuz we see some some ways that they can
[36:24] get incremental CapEx. But what people
[36:26] aren't focusing on is what does that
[36:27] mean next year? What does that mean the
[36:29] year after? And it turns out 3 years
[36:31] from now TSMC is going to spend a
[36:32] hundred billion dollars on CapEx.
[36:34] Maybe 2 years from now, right? Might be
[36:36] 28. Sincerely, they may spend a hundred
[36:38] billion dollars on CapEx in 2028. And
[36:40] people like just can't fathom that. But
[36:42] what does that mean for their downstream
[36:43] supply chains?
[36:45] You know, companies like Lam Research or
[36:46] Applied Materials or ASML or their
[36:48] further downstream supply chains like
[36:50] MKSI and and all these other companies
[36:52] the tail whip just gets whipped harder
[36:55] harder and harder and ultimately that's
[36:57] a shortage if you know TSMC wants to
[36:58] spend a hundred billion dollars in 2028,
[37:00] which is a real possibility. I think
[37:02] people would think that's insane, but
[37:03] that's a real real possibility.
[37:05] >> What about other parts of the chip
[37:06] ecosystem where GPUs have been
[37:08] completely dominant? What about like
[37:09] CPUs or ASICs or things that start to
[37:12] pop out as both opportunities and
[37:14] bottlenecks beyond just like Nvidia's
[37:16] GPU dominance? I mean ASICs are
[37:18] obviously taking off, but I'll sort of
[37:19] pivot away from AI chips to talk about
[37:21] these other things. There's a project we
[37:22] did on FPGAs and it turns out there's a
[37:24] hundred twenty FPGAs per per
[37:27] next generation rack
[37:29] AI rack and then like what about all the
[37:31] FPGA names? CPU wise all these
[37:33] reinforcement learning environments plus
[37:35] all the slot code you and I are
[37:36] generating that is now running on some,
[37:38] you know, Versel instance or whatever it
[37:40] is
[37:41] or some AWS instance or some bucket that
[37:43] we've spun up. All of that requires CPU
[37:46] and so CPUs are completely sold out and
[37:48] demand is skyrocketing there. Yeah, help
[37:50] people understand the role that CPU
[37:51] plays in everything. Yeah, so there's
[37:52] two there's two main reasons why you
[37:54] need tons of CPU. One is when you're
[37:55] doing reinforcement learning,
[37:58] the CPU is very critical to that. So so
[38:01] before you would throw all the
[38:02] internet's data into the model, train
[38:03] it, spit it spits and it it's it's some
[38:05] stuff out. Now you train all the world's
[38:07] internet's you put all the internet data
[38:08] into the model, then you put it in this
[38:10] environment. This environment is like
[38:11] hey model,
[38:13] try this out and it tries stuff out
[38:15] tries a bunch of different things and in
[38:16] the end there is an environment which
[38:19] scores whether or not what it tried out
[38:21] is successful and it grades it. And
[38:23] these environments can be anything. It
[38:24] can be, "Hey, check if the text was
[38:27] outputted in the right way, structured
[38:28] output." It could be very simple stuff
[38:29] or it could be very complex stuff. Um
[38:31] and people are starting to get into very
[38:32] complex things, right? Like, "Hey, I
[38:34] want you to open this file, change it,
[38:37] edit it, update it, submit it to this
[38:39] website. I want you to open up this
[38:40] physics simulation from Siemens and edit
[38:43] this CAD model." So, the environments
[38:44] can get more and more complex and those
[38:46] environments run on CPUs. They don't run
[38:48] on GPUs. They don't run on ASICs. The
[38:50] ASICs run the model that takes the input
[38:52] data from the environment, runs it
[38:55] through the model, the model creates
[38:56] outputs of various different
[38:58] trajectories, right? Ways that it think
[39:00] it could solve it um in different
[39:02] instances.
[39:03] Those trajectories are graded {slash}
[39:05] scored and the ones that are successful
[39:07] you train on and you update and you
[39:09] reiterate and you iterate iterate
[39:10] iterate. And so, CPUs are very useful
[39:12] for that one. And then, once you have
[39:14] these great models and you're deploying
[39:16] them, those models are generating code.
[39:18] They're generating useful output. That
[39:20] useful output, it doesn't go from a GPU
[39:22] straight to the human brain.
[39:24] Um it goes from a GPU or an ASIC through
[39:27] to, you know, a deployed app that you're
[39:29] deploying somewhere that actually just
[39:30] runs on CPUs. So, that's another area
[39:32] where there's a lot of demand and things
[39:34] are sold out um in a large large way. As
[39:37] you continue to assess and try to be the
[39:39] world's best informed person on both the
[39:41] trajectory of supply and demand, what
[39:43] are things that you wish you knew to
[39:45] make that understanding that you don't
[39:47] know? I think the hardest area for us
[39:51] um and for everyone is understanding
[39:53] tokenomics, economics of tokens.
[39:56] Um
[39:56] I think we have a really tremendously
[39:58] like good insight into how much it cost
[40:00] to run infrastructure, what the cost of
[40:02] tokens are, what the cost of models are,
[40:03] what the margins of these labs are.
[40:06] But, the usage and adoption is what's
[40:08] really difficult to model. You know,
[40:09] continuously, right? We We have these
[40:12] like We had like crazy In January, we
[40:13] had crazy estimates for February.
[40:14] Anthropic smashed them. How do we
[40:16] calibrate this model? What are the data
[40:17] sources for this? February, uh we had
[40:20] crazy assumptions for March and then
[40:21] they smashed them. And everyone sees the
[40:23] number of 10 billion and they're like,
[40:24] what the How do they add 10 billion
[40:25] in revenue? Who is using all these
[40:27] tokens? Why are they using them? What
[40:29] are they building with them? And then
[40:30] more importantly, with what they're
[40:32] building with these tokens, how is that
[40:33] actually diffusing into the economy and
[40:35] what value is that generating? Cuz it's
[40:36] not really something that you can
[40:37] capture in any
[40:39] any GDP statistic, right? All of the
[40:41] value of the tokens that I use get
[40:43] transformed into better information,
[40:45] which I then sell at a discount to what
[40:47] people used to sell information for
[40:49] relatively because
[40:51] and therefore that information is now
[40:53] making its way throughout the economy
[40:55] and and people are making better
[40:56] investment decisions or better
[40:59] competitive decisions but if they're
[41:00] semi-direct company or data center
[41:01] company or hyper scalar. And now how how
[41:03] much what what is the value of this and
[41:05] what is that what is that done to the
[41:07] economy? It's clearly by every
[41:09] subjective metric amazing.
[41:11] But where is the phantom GDP? What is
[41:13] the phantom GDP? How do we track the
[41:16] real economic value cuz cuz the GDP
[41:18] metrics are not, you know, accurate if
[41:21] you were to say, what is the GDP that
[41:23] Dylan Patel is making? It's tiny
[41:24] compared to what the value that I think
[41:26] is being created. And so ultimately,
[41:28] what is the value being created by these
[41:30] tokens? Not on a basis of, you know,
[41:32] just simple, you know, what is the
[41:33] knock-on effect, right? What is the
[41:35] knock-on effect of all the things that
[41:36] these things are doing? And I think
[41:37] that's the real
[41:39] question and challenge that's hard to
[41:41] measure. I think we've got a tremendous,
[41:44] you know, reading on the supply side of
[41:45] things. I think we've got a tremendous
[41:46] reading on even a lot of the demand side
[41:48] signals but it's it's what is the value
[41:50] these tokens are generating? That's hard
[41:52] to quantify and measure. I hope we get a
[41:54] chance to do this like every 3 months
[41:55] because this changes so quickly.
[41:57] What do you think's going to happen
[41:58] next? Like when I when I come back 3
[42:00] months from now and we're in San
[42:01] Francisco together again, what do you
[42:02] expect?
[42:03] >> Large-scale protests.
[42:05] Really? Yeah, I think there'll be a
[42:06] large-scale protest against Anthropic.
[42:09] And up in AI. Expand on that a little
[42:10] more. Uh people hate AI.
[42:12] Um, AI is less popular than ice, less
[42:16] popular than politicians. Confused how
[42:18] Pew surveyed this, but apparently, AI is
[42:21] less popular than politicians. You know,
[42:22] with Anthropic adding so much revenue,
[42:24] that's going to start causing business
[42:26] changes downstream. People are going to
[42:28] get more and more scared of AI. They'll
[42:29] start blaming more and more of their own
[42:31] problems and things that are, you know,
[42:34] global, you know, have been deep-seated
[42:36] problems for a long time. Those will
[42:37] bubble up and be blamed on AI. Um,
[42:40] probably some politician or some social
[42:43] media people will start to be able to
[42:44] take, uh,
[42:45] influence or will be able to start
[42:46] taking and weaponizing AI against
[42:49] people. You look at the comments of news
[42:51] articles where Sam Altman had a Molotov
[42:53] cocktail thrown at his house twice in
[42:55] like 2 weeks. They're like, people are
[42:57] cheering it on.
[42:59] Uh, and this is just the beginning. So,
[43:00] I think I think we'll see large-scale
[43:01] protest against AI in 3 months. What is
[43:04] the counterweight to that? Like, how
[43:06] should the AI industry head that off?
[43:08] First of all, Sam Altman and Dario have
[43:10] to stop getting on interviews. They're
[43:11] so uncharismatic.
[43:13] I don't know what they're doing. Every
[43:15] interview they do is like, well, normal
[43:17] people are going to hate you even more.
[43:19] Like, Sam being on Tucker Carlson
[43:20] probably made all Republicans hate
[43:22] OpenAI. And same with Dario, they just
[43:24] have no charisma. I think that's first.
[43:26] Two,
[43:27] they need to start showing uplifting
[43:28] things that can be done with AI.
[43:30] Um, three, they need to stop talking
[43:32] about how the capabilities are going to
[43:33] change the whole world constantly,
[43:35] because then people are going to get
[43:36] fear of that capability, cuz they have
[43:38] no connection
[43:39] >> to use it. There's no connection to it,
[43:41] either. Like, the average person doesn't
[43:43] know an Anthropic employee. Average
[43:44] person doesn't know an OpenAI employee.
[43:46] Average person doesn't know who these
[43:47] people are, what their goals are, and
[43:49] they just view them as like this like
[43:51] sneaky cabal of like 5,000 people at
[43:53] this company that are going to change
[43:54] the world and automate all the jobs and
[43:56] and destroy society. That's what they
[43:58] view it as. And and as people who are
[44:00] funding the building of all these data
[44:02] centers and and power plants that are
[44:04] going to pollute the world, right? They
[44:05] don't quite understand what's happening.
[44:06] You know, they have to stop talking
[44:07] about the future thing that's going to
[44:08] happen and only talk about present, how
[44:10] uplifting AI is. I think it's a huge
[44:13] reorg and rebranding needs to be done. I
[44:15] love doing this [music] with you. Thanks
[44:16] for your time. Awesome, thanks.
[44:22] Your finance team isn't losing money on
[44:23] big mistakes. It's leaking through a
[44:25] thousand tiny decisions nobody's [music]
[44:27] watching. Ramp puts guardrails on
[44:29] spending before it happens. Real-time
[44:31] limits, automatic rules, zero
[44:32] firefighting. Try it at ramp.com/invest.
[44:35] [music]
[44:37] As your business grows, Vanta scales
[44:38] with you, automating compliance and
[44:40] giving you [music] a single source of
[44:41] truth for security and risk. Learn more
[44:44] at vanta.com/invest.
[44:47] Ridgeline is redefining asset management
[44:49] technology as a true partner, not just a
[44:51] software vendor. They've helped firms 5x
[44:53] in scale, enabling faster growth,
[44:55] smarter operations, and a competitive
[44:57] edge. Visit ridgelineapps.com [music] to
[44:59] see what they can unlock for your firm.
[45:02] Every investment firm is unique and
[45:04] generic AI doesn't understand your
[45:05] process. Rogo does. It's an AI platform
[45:08] built specifically for Wall Street,
[45:09] connected to your data, understanding
[45:11] your process, and producing real
[45:12] outputs. Check them out at
[45:13] rogo.ai/invest.
[45:17] The best AI and software companies from
[45:19] OpenAI to Cursor to Perplexity use
[45:21] WorkOS to become enterprise ready
[45:22] overnight, not in months. Visit
[45:24] workos.com [music]
[45:25] to skip the unglamorous infrastructure
[45:27] work and focus on your product.