YT YouTube2Text

YouTube2Text → Transcripts → Why AI Demand Is Outrunning Compute Supply

Why AI Demand Is Outrunning Compute Supply — Transcript

by a16z · 14,393 words · 2,242 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00when the history of the 21st century is
0:01written, you know, there was like the
0:02Victorian age. I think this will be like
0:04[music] the age of Elon and Jensen
0:06because they are fundamentally altering
0:08the fabric of human society and
0:10civilization.
0:10>> What happens if there's like a massive
0:12supply shortage?
0:13>> Every time you've had a real profound
0:16new technology, you get a bubble because
0:17the markets get really excited and they
0:20get ahead of themselves. Things get
0:21overvalued. That overvaluation leads to
0:24an overbuild. One of the things that I
0:26think has been correct but ineffective
0:29is this idea that we need to stay ahead
0:30of China.
0:31>> You're opposed to data centers. Well,
0:33you know what? It's probably the best
0:35thing that has ever happened to
0:37workingass Americans. We are
0:38re-industrializing America and it's
0:40awesome.
0:41>> Assume that you're right. There's not a
0:43physics reason why this can't work.
0:45>> An increasing fraction of the world's
0:48compute is going to be in orbit. This
0:50sounds crazy, but asteroid mining is
0:52going to be a very real thing. It has
0:54more gold, silver, platinum, every
0:56[music] precious metal in it that exists
0:58in the earth's crust.
0:59>> Every LP conversation that we have
1:00starts with like, how's this all going
1:02to go wrong?
1:06>> Gavin, uh, you've been out here hanging
1:09out on the West Coast over the summer
1:11and you've been talking about the fact
1:13that you're like trying to find someone
1:14to make you to give you like a a bearish
1:17case, like to make your sentiment more
1:19negative. Um, have you found anybody?
1:22>> No. And I ask everyone, my standard
1:24question is, can you tell me one
1:27quantitative data point in your business
1:30that's getting worse? Just one. That's
1:33my standard question. And it's at least
1:36in July and August, I haven't been able
1:38to find a single person. Now, if we're
1:40being honest, you know, Anthropic is um,
1:43you know, in a quiet period, so maybe
1:44they've slowed down a little bit, but I
1:47do think the rest of the world
1:50has accelerated. You know, OpenAI is
1:53clearly accelerated. Open source, I
1:54think, has accelerated more. And then I
1:56do think Grock, particularly after
1:58Grockbot, has had a pretty experience,
2:01has had a pretty dramatic acceleration.
2:04And so AI overall, it accelerated in
2:06July. It accelerated in August, and it
2:10can't keep accelerating forever, but
2:13it's just kind of wild that, you know,
2:15public stocks have kind of fallen out of
2:16bed over the last, you know, two months.
2:19And I mean,
2:21you know, it's uh that, you know, they
2:23you can you can drown crossing a river
2:25that's on average 2 ft deep. And so, you
2:27know, there's not a lot of action at the
2:29index level,
2:30>> right?
2:30>> But some of these AI names are in pretty
2:32significant draw downs. And they b they
2:34bounced a little bit um in August, but
2:37still pretty big draw downs and things
2:39are broadly accelerating.
2:41>> Yeah.
2:41>> It's um you know, our friend Eric
2:43Fishery did a podcast with Patrick
2:44Oanessy and he said maybe everyone wins.
2:47>> Yeah. you know, Anthropic wins, OpenAI
2:50wins, SpaceX wins, Meta wins. Um, you
2:54know, Google wins by selling a lot of
2:55TPUs. Um, open source wins, NeoClouds
2:59win, inf you know, inference cloud, uh,
3:02the inference clouds win on top of the
3:04Neoclouds. Um,
3:06>> applications win.
3:07>> Yeah. Every Yeah. And that kind probably
3:10maybe not all applications applications
3:12that I think execute well and navigate
3:14this
3:16but that feels like a very possible
3:19scenario to me and there's so much zero
3:23someum thinking in the world and by the
3:25way on anthropic what is
3:28my hypothesis would be if you're
3:31anthropic one I think they probably tred
3:34up and cleaned up some accounting
3:36>> yes definitely
3:36>> you would you'd rather do Yes.
3:39>> So you rebased and now you're comparable
3:41to OpenAI.
3:42>> Yeah. In terms of revenue added, like in
3:44terms of the definition and now I think
3:46kind of revenue added.
3:47>> Exactly. So you kind of rebased and then
3:51they you know they did their testing the
3:53waters. Um [clears throat]
3:56and then you know I would hypothesize
3:58because they've executed well probably
4:00the next disclosure is a reaceleration.
4:04And then there's always this kind of
4:06funny game between the frontier model
4:09companies. They always have more
4:10advanced checkpoints. Anthropic is
4:13clearly waiting for OpenAI to release
4:15Astra.
4:16>> Yes.
4:16>> And then it's like the next
4:17>> the next day.
4:18>> Here's Fable 5.1.
4:20>> Yes. Exactly.
4:21>> Magically and just happened to be
4:22available several hours after Astra.
4:25>> Yeah.
4:25>> So I think they're being thoughtful um
4:28and you know heading heading into this
4:29IPO and everyone is shooting at them.
4:32>> Yes. everybody's shooting at them and
4:33they're in a quiet period so they can't
4:35really shoot back. Um, and so it's, you
4:39know, there's a lot of gamesmanship, but
4:41I do think
4:42having OpenAI anthropic be public
4:46companies is going to be helpful for the
4:47market just cuz it's,
4:49you know, it's such a uh powerful force
4:52and a lot of public investors, you know,
4:54you hear, oh, you know, Sarah Frier said
4:57this at an all hands meeting and it's on
4:58the cover Wall Street Journal. Okay,
4:59we're going to put that into our model.
5:01>> Yeah. And it's just a lot I think it'll
5:03be better for them to be public. I am a
5:06little um you know Anthropic is now in
5:10their culture interviews saying how
5:12would you feel if the equity went to
5:13zero?
5:14>> Yeah.
5:14>> Because we're looking for people who are
5:16mission aligned.
5:17>> Yeah. Mission not mercenary. Yeah.
5:18>> And and that's great. We we want we want
5:21missionaries, but we also want people to
5:23make money. And at the end of the day,
5:25you can't afford the compute you want
5:27for your mission if you go if the equity
5:29goes to zero. Like I'm no expert, but
5:32I'm pretty sure on that
5:34>> in that I do think they are
5:36>> they're like the accidental enterprise
5:38company.
5:38>> Oh, for sure. Oh, yeah. They're kind of
5:40like the accidental everything.
5:42>> Enterprise is just a byproduct of like
5:44the the the mission, the objective at
5:46the end. Yeah. Whereas I think open eye
5:47is a little more commercial and
5:48obviously SpaceX a little more
5:49commercial.
5:50But all of these companies like let's
5:52just let's let's just say um let's just
5:54say they have 10 gigs of power and
5:58they're allocating eight to inference
6:01and let's just say they're monetizing
6:03that inference at you know whatever um
6:08you know 60 60 billion a year. Um, so
6:11$480 billion a year in revenue,
6:14>> which is like a on a revenue payback
6:16basis would be like a one-year payback
6:18on a revenue basis, not a gross profit
6:19basis.
6:20>> Yeah. On a revenue basis. Yeah. Yeah.
6:22>> Um, and I've tried to use conservative
6:23numbers. You know, people seem to think
6:25and open are both monetizing at hundred
6:27billion dollars a gigawatt today.
6:29>> Yeah.
6:31Let's say they have a big research
6:32breakthrough and they decide, "Wow, it
6:35is to our long-term advantage
6:38to go from eight gigs allocated to
6:40inference, two gigs allocated to
6:43training to 8 gigs on training and then
6:46your revenue just went from 480 to 120."
6:50And I think they your annualized revenue
6:53and I actually think they would do that
6:55>> make that decision.
6:55>> Yeah. And this is just something that
6:58like public markets are going to really
7:01have to get used to.
7:03>> Yeah.
7:04>> It as you say, open AI may be a
7:06different animal. And I do think like
7:09the realities, you know, everybody
7:11everybody has these ideals about how
7:13they're going to manage their business,
7:14then they go public and the stock is
7:16volatile and it really impacts, you
7:18know, employee morale, recruiting,
7:20retention. So, I'd be surprised if they
7:23did such a dramatic cut, but a lot of
7:26the revenue is kind of under their
7:28control based on what checkpoint they
7:30release.
7:31>> Yeah.
7:32>> Where they price um along this, you
7:35know, kind of paro curve and then how
7:37much they allocate between training and
7:38inference. So, it's just
7:42it's going to be, you know, Meta and
7:44Google and these kind of internet
7:46companies. It was just it was pretty
7:49smooth fundamentally even if the stocks
7:51were volatile.
7:51>> Well, there was no like massive
7:53trade-off they had to make in terms of
7:54the cost or infrastructure to serve
7:57revenue side. Like they were totally
7:59separate
7:59>> 100%.
8:00>> Yeah. It's it's fascinating. Um so, you
8:03know, if you go back to Eric's point of
8:05like it's all going to work like I
8:07actually think that's a great point.
8:08Like I I I describe it differently. I've
8:10had this conversation with LPs a lot cuz
8:11every LP conversation that we have it's
8:13probably the same for you starts with
8:14like how's this all going to go wrong
8:17>> and it's like what's what's going to
8:18crash and I'm like this is the this is
8:20the and like oh are the are the large
8:22models screwed or the labs screwed
8:24because of open source and I'm like this
8:26is this is this is all wrong like this
8:27is not an or thing it's an and thing
8:29right like this is an and thing um
8:32Frontier is going to work really well
8:33like N minus one models are going to
8:34work really well open source is going to
8:36work really well um there's going to be
8:38a bunch of application companies that
8:40work really well. Like the clouds are
8:42probably going to be fine. They're
8:43probably going to work really well.
8:45>> Like the five lab companies are probably
8:46going to do really well.
8:48>> Yeah. And Nvidia is
8:51at the center of all of it.
8:53>> Yes. Yes. They're probably going to do
8:54pretty well.
8:55>> Yeah.
8:58The last um 26 years have taught me not
9:00to bet against Jensen.
9:02>> Yeah. He's he's he's in a pretty good
9:03position here. Um I want to come back to
9:05that. The the point that you made about
9:07training verse inference is an
9:08interesting one. It seems to me like the
9:11labs will decide to take all incremental
9:15profits and probably much more than
9:18their profits and invest them in
9:20training for a long period of time.
9:22Would you think that's fair? Like this
9:24is very different than like the clouds,
9:26you know, cuz like the the cloud like
9:28the internet companies and the clouds,
9:29they just end up being supply demand
9:32driven and they generate tons of profit
9:34and they can still grow a certain amount
9:37like but they don't have some maybe with
9:39the exception of Meta like some big
9:41long-term bet that's like a multi-year
9:43payoff.
9:44>> Yeah, I think it's important to kind of
9:45be precise. They I for sure I don't
9:48think they will generate free cash flow
9:50anytime soon. I think they're going to
9:52generate a lot of operating cash flow
9:54and then they'll use that to buy a lot
9:56of, you know, GPUs,
9:59um, XPUs, whatever, whatever we're going
10:01to call them. Um, or maybe they
10:04subsidize heavily. Like we do know that
10:06that's happening at the labs.
10:08>> Subsidize what heavily
10:10>> their first party products. So token
10:12consumption of their first party
10:13products. So like they're doing all this
10:15research and they're spending a lot on
10:17data on compute
10:18>> and the first products that are like a
10:20heavy subsidy
10:21>> products today, right?
10:22>> Yeah. So it's 8 gigs of inference and
10:24two gigs is for internal research and
10:26then you know two gigs is actually
10:29training.
10:29>> Yeah. Exactly.
10:30>> Um and you know including probably the
10:32inference that goes into post- training.
10:35>> Yeah. I don't I I think given the belief
10:38systems that they all seem to have about
10:41scaling laws
10:43which continue to hold I don't think any
10:46of them are going to be that focused on
10:48generating free cash flow and you've
10:50seen right we saw Satcha blink.
10:53>> Yes.
10:53>> And Satcha really regrets that I think.
10:56>> Yeah. Yeah. um you know he kind of
10:57blinked I think it was last year
11:01you know he gave that great interview
11:03for Davos and they asked him about all
11:04the capex and he said I know I'm good
11:06for my 80 billion
11:08>> right
11:08>> and and I think they blinked a little
11:11they slowed down they regret that and
11:14then Daario famously he went on a
11:16podcast and he made and he said listen
11:18some people are being super
11:19irresponsible with their spending and
11:21it's a hard decision because if you
11:24don't spend enough you could lose a lot
11:26of shares But if you spend too much, you
11:27could go bankrupt. And like those are
11:29both bad things, but bankruptcy is worse
11:31than losing shares. So I'd rather be
11:33conservative. And he was conservative.
11:35And OpenAI was aggressive. And now
11:37OpenAI is back in the game.
11:39>> And SpaceX was aggressive.
11:40>> And SpaceX was aggressive.
11:42>> And so, you know, like there are clear
11:44high ROIs on those independent of supply
11:47demand mismatches that are happening.
11:48Like clearly that seems to be the right
11:50decision short-term and long-term.
11:52>> Yeah, absolutely. I mean, we we
11:53calculate, you know, Nebius um and
11:56Corweave both gave some interesting
11:57disclosures, but you can kind of get to
12:00a 9 to 10 month payback for Nebius
12:04because you know, okay, you bring on a
12:06gig, it costs 50 billion. You get you
12:09can get an upfront payment for 50 to 60%
12:11of that for customers. Yeah.
12:13>> So now, you know, you're talking about
12:1525 or 30 billion and then you can
12:17monetize it if you put it into the spot
12:20market.
12:20>> The spot. Yeah. at a spot spot paybacks
12:23are probably much faster than nine or 10
12:25bucks.
12:26>> Yeah, you got to assume like a smoothed
12:27out level like two two bucks, three
12:29bucks even with that. It's very very Now
12:31you can get like five bucks or eight
12:32bucks and Yeah.
12:33>> And then SpaceX cuz they build these
12:35really big clusters and and I think at a
12:37really important point is they bring
12:39them on fast.
12:40>> Yes.
12:41>> They have an even faster payback and
12:44they can monetize at you know higher. I
12:46I have tried to shift um you know to
12:49think of pricing and you know per
12:51megawatt rather than per GPU because it
12:53seems like it's where where the world
12:54>> world is but like SpaceX the payback
12:57feels well inside of that.
12:58>> Yes.
12:59>> And I just in my career as an investor
13:04there haven't been that many
13:06opportunities where you have companies
13:08that could deploy tens hundreds of
13:11billions of dollars and get sub one-year
13:14paybacks.
13:14>> Yes. And it's kind of crazy. And then
13:17also like we should also talk if
13:19particularly if you're buying Nvidia
13:21GPUs to a lesser extent TPUs, you can
13:23finance these.
13:24>> Yes.
13:25>> And there's a very sophisticated, you
13:27know,
13:27>> Yeah. very low cost of capital to
13:29finance them today.
13:30>> Yeah. And everybody's, you know, worked
13:31up about, you know, circularity and it's
13:33like, well, I don't know. Um, I know a
13:37lot of smart people who work at
13:39Blackstone and KKR and Apollo
13:43and they're the ones that are financing
13:45>> the ones who are financing it at a
13:46relatively low cost
13:47>> at a relatively low cost. And I think
13:50one reason that's happening is useful
13:52lives just keep getting extended and has
13:55these models get better and better and
13:57better and the ROI on token spend goes
13:59up, you know, the monetiz monetization
14:02rate per gigawatt goes up. So, I mean,
14:06the true equity payback like might be
14:11way inside of a year.
14:12>> Yeah. Exactly. Exactly. Yeah. And look,
14:14there's a case you could make that the
14:16prices actually of all the stuff go up,
14:19which could make the the supply side
14:21economics even more compelling, right?
14:23Like, you know, so on the supply side,
14:27like that's the dynamic today. Like, it
14:28just is what it is. Like, there's a ton
14:30of data points out there that paybacks
14:31are within a year.
14:32>> Yep. Um I think it's actually
14:34interesting to think about the demand
14:35side too because the knock would be well
14:38in all these cycles you get some
14:40overbuild and then that you know
14:42destroys the economics of the supply
14:43side. The demand side today like what
14:46are we monet like the monetization of
14:49these companies which are doing call it
14:5280 billion of revenue or something in
14:53that direction um is on the back of what
14:57like 30 million actual heavy paying
15:00users like re getting real value. I'm
15:02talking about like developers like
15:04>> I might take the under on 30 million
15:06>> so call it yeah actually what we see
15:07inside our companies is you know
15:09obviously there's a power law in which
15:11companies are spending a lot on tokens
15:13like old banks are probably spending 1%
15:16very techforward companies are spending
15:18high single digits but if you actually
15:21look at the sort of the the power law of
15:24what's happening of the actual engineers
15:26in those companies the highest spending
15:28engineers are spending 10 or sometimes
15:31is 100x more than the median engineer.
15:33And so, yeah, your 30 million is
15:35probably way overstated. It might be sub
15:3710. And so, there's this question of
15:39like where are we at in diffusion?
15:41There's one and a half billion knowledge
15:43workers. Like, it feels like we're
15:44nowhere on the demand side and we're
15:46massively supply constrained.
15:48>> And what are I'm just curious across the
15:50A6Z portfolio if what are your best
15:54companies spending on tokens per month
15:58relative to human compensation? What
16:01rough range?
16:02>> Oh, high single digits, some at 10%,
16:04like some of the very AI native ones
16:06like 10% plus. And so, you know, and and
16:09then old economy companies are spending
16:11the ones that are probably doing a good
16:12job like 1%. So, it feels to me like
16:16when I look at the supply demand
16:17characteristics, it's like supply stuff
16:21people say, is that sustainable? Well,
16:23like when you pair it with the demand
16:24stuff, I it feels it feels specific.
16:27Like there could be things that
16:28disappoint us in terms of like diffusion
16:30into the real economy,
16:32but it feels like over a 10-year
16:33stretch, like we're nowhere.
16:35>> Yeah. Absolutely nowhere. And I just
16:38>> my So at a trade is our internal token
16:41consumption has gone up 100x from the
16:44month of March. March through August.
16:46100x our token spend. And we just got
16:51access to uh Grockbot Enterprise and
16:55with two people using it like it looks
16:58like it token spend might 10 or 20x in a
17:03month.
17:03>> Yes.
17:03>> From August.
17:04>> Yes.
17:04>> Like like I
17:06>> But but it's actually extremely
17:07valuable. Like we have some heavy
17:09Grockbot users here and like it is very
17:13productive use. Like this is not like
17:14wasteful tokens, but
17:16>> yeah, I was and and listen like I I try
17:19super hard. I you know I always when I
17:22use AI, I just remember when my parents
17:24like I was trying to get them to shift
17:26to an iPhone and an iPad and like you
17:29know get them used to it and like you
17:32know and they did a good job. I give
17:34them loads of credit and but you know
17:36I'm 50 years old you know like how old
17:39are you David?
17:40>> 42. 42 and you see these like 23-y old
17:42kids and just the way they use AI,
17:44they're just fluent and native in it. I
17:47just feel like maybe in a way that no
17:48matter how hard I try, I will never be
17:50and I'm trying really hard. But, you
17:52know, like we got cloud code, I try I
17:55you know, I built some stuff, did some
17:57cool stuff and in like I don't know
18:01three minutes of type creating Grock
18:04bots, I had much better versions of
18:06everything I created, you know. You
18:08know, so I went on this Patrick Oannessy
18:10pod podcast like 5 months ago and I
18:13said, you know, like I love having a
18:15podcast summarizer. Everybody's like,
18:16"How'd you do it?" I was like, "Well,
18:17just use AI and do it."
18:19>> Yes. Pretty simple.
18:20>> It takes 10 seconds and Grockbot.
18:23>> Yes.
18:23>> It's amazing and it's so good.
18:25>> Yeah.
18:25>> And then, you know, a Substack
18:27summarizer, an X summarizer, um an Xs
18:30sentiment tracker for topics and stocks.
18:34>> Yeah. And like that all of those would
18:36have taken me I don't know hours working
18:40with cloud code and they each took 7 to
18:4512 seconds with Grockbot and it's
18:48better.
18:48>> Yeah.
18:48>> So to to me Grockbot does feel like
18:51another um at least for me like kind of
18:54chat GPT moment because Claude code
18:57>> like I can see in the data was it was
18:59powerful. I did some really cool stuff
19:01with it that was like empowering and
19:03this is neat.
19:04>> Um
19:06>> you know like family calendar apps,
19:09things like that.
19:09>> Yeah.
19:10>> Um but this is just 10 seconds and it's
19:14way better than what I was able to do.
19:16>> Yeah. Yeah. Yeah, the cloud code thing
19:17like was obviously the shift in coding
19:20and you know our our most sophisticated
19:24engineers you know were doing whatever
19:2520% of their code you know with with
19:28with AI to like you know whatever 90
19:30plus% and so now I think everything you
19:33described in what you built with cloud
19:35code or codec
19:36>> is still kind of reactive
19:39>> in a way right like it's it's still you
19:40know it's like summarizers yeah
19:42preparation it's all like knowledge
19:44enhancing which is part of your job, but
19:46it's not actually doing the work for
19:48you.
19:48>> Yeah. And now you have a Groc bot that
19:51says, "What are the recommended
19:52actions?" Yes. Exactly.
19:54>> Based on everything the other bots have
19:56learned today.
19:56>> Yeah.
19:57>> What recommendations do you have for me
19:59today? And that for sure is like
20:01>> and it it was so easy to build. Um I now
20:04have it. So I'm I'm like horse racing
20:07all these which is like I have uh uh
20:09Crockbot doing it, Codeex doing it. All
20:12the like action taking for just I want
20:13to know
20:14>> make me better at my job. Look at
20:17everything I do. Give me give me
20:18recommended automations you can do. I
20:20have Town doing it as well which is one
20:22of our companies very good at it.
20:24>> Um but and we're like kind of on the
20:26bleeding edge of trying to do this
20:27stuff.
20:28>> Just wait till everyone does this stuff
20:30>> and then and then when we actually click
20:32like yes go just automate this.
20:33>> Yeah. It feels like that's sort of
20:35endless token
20:36>> and but I do we should acknowledge like
20:38the the history of financial markets,
20:42>> you know, dating kind of back to like
20:43the South Sea bubble is whenever you get
20:46this transformational new technology.
20:48Um, I actually went on a podcast, I said
20:50I thought the South Sea bubble was
20:52connected to like the invention of
20:54longitude and the ability to sell. Turns
20:55out it was not. [laughter]
20:57It was just it was kind of like a more
20:59of a tulip episode. But like every time
21:02you've had a real, you know, profound
21:04new technology, you know, whether it's
21:06the automobile, the TV, the radio,
21:08internet, the PC, um, railroads,
21:12>> steel mills, you get a bubble because
21:14the markets get really excited
21:16>> and they get ahead of themselves. Things
21:19get overvalued. That overval
21:21overvaluation
21:23leads to an overbuild. And then
21:25particularly if you're funding it with
21:27debt um and and even today a majority of
21:30this is still being funded out of
21:31operating cash flow which I think is
21:32really helpful. Um you know debt funded
21:36built buildouts they demand immediate
21:38ROI not an ROI in three years.
21:40>> Yeah. You can't be off in the time. You
21:41can't be off off on the time, but I'm
21:43just more, you know, like I um, you
21:46know, I talked to Jazz about how Watson
21:48wafers Jazz I guess and Patrick Watson
21:50wafers are these fundamental constraints
21:52and just that the buildout is so big and
21:55we're so early that we are
21:58it's like impacting the raw productive
22:01capacity of so many industries. you
22:03know, now you know, everybody in
22:05everybody in copper, there's like an AI
22:07thesis and like we're going to have to
22:10like think about it to like fill the,
22:13>> you know, if if
22:15>> 10% of what we just talked about comes
22:17true, you know, we're in this acute
22:18shortage with, I don't know, several
22:20million people are driving a crazy
22:23global compute shortage. What happens
22:25when that's 500 million? And you know,
22:28how many copper mines do we need to
22:30build to like support this? Yeah,
22:33>> it's kind of a wild thought. And so like
22:35these fundamental constraints, I think,
22:37are slowing us down.
22:38>> And I
22:39>> and I think that's good. I actually
22:41think that's good for society. And I
22:42would now say rates and regulation, you
22:45know, real rates are going up. Yes.
22:46>> And it just is what it is, which makes
22:48sense because we're like investing a
22:49lot. So it makes sense um that real
22:52rates are going up. And then regulation,
22:54man. It's it is like I'm kind of shocked
22:58at what's happening in America.
23:00>> We're in a really bad place.
23:01>> Yeah. [clears throat]
23:02And just, you know, I had this exchange
23:05with with um Schulto for from Anthropic
23:08and and and Daario on X last weekend.
23:12You know, Daario said, "Hey, I don't
23:14think I've been negative. You know, I've
23:15written I've written two essays. One was
23:17positive, one was negative." So being
23:2050% negative and particularly when it's
23:22like a terrifying negative
23:25>> like an existential
23:27>> an existential negative everybody might
23:28be out of out of a job like that Eleazar
23:31Yukowski guy says if we build it
23:33everyone will die and it's like how
23:35about if we build it like we're going to
23:37cure cancer we're all going to live
23:39forever. I thought one of the best
23:40things Dario said was like what we need
23:42to do is stop talking about curing
23:43cancer and actually cure cancer.
23:45>> Actually cure cancer and actually make
23:46breakthroughs like
23:48>> but just somebody like that my favorite
23:50line in the Bible is the truth shall set
23:52you free.
23:52>> Yes. But the only person who can the
23:56only group that can tell the AI
23:58industry's truth is the AI industry.
24:00They need to just start telling the
24:02truth. Hey when we Okay, you're opposed
24:06to data centers. Well, you know what?
24:07It's probably the best thing that has
24:09ever happened to workingclass Americans.
24:12>> Yeah, exactly.
24:13>> You know, it's like going to college
24:14might be significantly NPV negative now
24:18because you can go learn how to be an
24:20electrician, a plumber, an HVAC tech,
24:23and make ungodly amounts of money. Yeah.
24:26>> So, this has been amazing for
24:27working-class Americans. We now have a
24:29lot of data that particularly with
24:31behind the meter power generation, when
24:32a data center goes in,
24:35it transforms a town. like tax revenue,
24:38it doesn't double. It like 10xes and it
24:42is re revitalizing all of these like
24:45dying small towns all over America. And
24:48listen, we're getting we're getting much
24:50better at addressing the environ
24:52environmental stuff. Generally, they use
24:54natural gas, which is a pretty clean
24:56fuel.
24:57>> The the water the water consumption
24:58thing is totally debunked. It's totally
25:00debunked. Yeah.
25:01>> It's nothing. It's nothing. So these are
25:03like really really really good and
25:04they're having a really positive impact
25:06on the world. That's without even
25:08considering things like curing cancer,
25:10but somebody needs to tell that story.
25:12>> It's now and and I think the problem
25:14with it now is like the burden of proof
25:16is on not curing cancer, but actually
25:18delivering some real tangible everyday
25:21American benefits beyond using chat, you
25:24know, or Grock to like answer your
25:25questions or substitute
25:27>> for a search engine, right? It it does
25:29feel like we're pretty close to that. Um
25:33yeah, it does. And and by the way, like
25:36one of the things that I think has been
25:38correct but ineffective is this idea
25:41that we need to stay ahead of China.
25:43>> Like it's like it is true. Like I'm very
25:45much like I'm a patriot. Like I believe
25:46that. But it's way too abstract. Yeah.
25:48The abstract for the average American.
25:50Like
25:50>> nobody's worried about China invading
25:52America.
25:52>> Yeah. Exactly. Like they ocean is really
25:55big.
25:55>> Yeah. like affordability and like how is
25:57this going to change my life for the
25:58better or worse, right? And so
26:00>> I think there's a pretty immediate
26:01impact you could feel like I my favorite
26:03is, you know, Lowden County, Virginia,
26:05which is like the highest uh highest per
26:08capita income uh zip code in the US or
26:12county in the US
26:14>> and it has the highest density of data
26:16centers.
26:17>> Yeah.
26:17>> And they and they make a tremendous
26:18amount of tax revenue from data centers.
26:20Like we should we should do this
26:21everywhere.
26:22>> Yeah. It was actually very funny. a
26:23someone very opposed to data centers
26:25said, "Oh, you're for data centers. I'd
26:26like to see them put in the highest
26:28income zip code and the highest, you
26:30know, income county." And they're like,
26:31"Actually, the highest income zip code
26:34in America and the highest income county
26:36has the highest per capita concentration
26:38of data centers." So, we've done that
26:40[laughter]
26:40>> and it worked out really well.
26:42>> Yeah. But, you know, hey, don't bother
26:43me with the details. I'm on to my next
26:45talking point.
26:45>> That's good. That's good.
26:46>> And all those talking points, it's
26:48tragic. Like there is an organized CCP
26:50funded campaign. I think against data
26:53centers here in America like I think a
26:55lot of it gets laundered through Tik Tok
26:58and it's just tragic because the other
27:00thing that's happening is this is
27:01re-industrializing America. The
27:03combination of having the straight of
27:05foremost closed which is amazing for
27:06America. You know natural gas here is
27:08two or three bucks.
27:09>> It's now 25 bucks
27:12>> in Europe and Asia or 20 bucks or
27:13whatever it is. And natural gas is an
27:16you know important input to the cost of
27:18electricity which is an important input
27:20to almost all manufacturing processes.
27:23And so we have a huge cost advantage for
27:27that basic input now.
27:29>> And you have that happening and you have
27:31this kind of data center boom happening.
27:33We are re-industrializing America. And
27:35it's awesome. This is what everyone in
27:37both parties has wanted for a long time.
27:40>> Yeah. Exactly.
27:41>> Like bring industry back. small towns
27:43that were left behind by the steel mills
27:45closing. Well, data centers are bringing
27:47them back.
27:47>> Yeah. But somebody has to tell that
27:49truth. I mean, I try to do it on every
27:50podcast, but like I'm just a dude.
27:53>> Yeah. And like your audience is the tech
27:55audience that that already believes
27:56you're you're preaching the choir, if
27:58you will. Um, but yeah, the story the
28:00story I met Meta is probably doing the
28:01best job of telling that story, I would
28:03think.
28:04>> Yeah, it seems.
28:05>> You know, and I think one reason it's
28:07really wired into Meta's DNA. So, one of
28:09the first things they started doing as a
28:11public company I don't remember if it
28:13was on their first attorney's call, but
28:15Cheryl would run through Cheryl Samberg
28:18would run through 10 or 15 very specific
28:21small businesses that had started using
28:25Meta's advertising products and the
28:28impact it had on that business.
28:29>> Yeah. you know, this cake bakery in De
28:33Moines started, you know, worked with
28:35Meta and, you know, it was it was it was
28:38two women who were single mothers
28:40working by themselves and now they have
28:4315 locations. They employ 50 people.
28:47>> Yeah.
28:47>> And this has been amazing for De Moine
28:49and it's been transformative for them.
28:52>> Yeah.
28:52>> And they would just run through that
28:54every time. And and I do think the
28:56entire AI industry um like I'd love to
28:59see, you know, everybody SpaceX,
29:03Enthropic, OpenAI, Google, Meta say,
29:06"Hey,
29:07>> here are real businesses and real
29:09Americans and like either name the
29:11business or get permission to if you can
29:13name the American or anonymize it." This
29:15is a really positive thing it did it it
29:17had on their life.
29:18>> Already very tangible. Yeah.
29:19>> Yeah. Same. Nvidia, AMD, Broadcom, all
29:22of them.
29:23>> Yeah. just run through specifics because
29:25the truth will set you free but only if
29:27you tell it.
29:27>> Yeah. Exactly. Exactly. Yeah. So it
29:30seems more likely than given that fact
29:32pattern if you go back to just the sort
29:34of macro situation that we're in that we
29:37we underbuild on the supply side.
29:39>> Oh yeah. For for like through 28. And
29:42and by the way like there's no capacity
29:43available with all the forecast builds
29:46that will happen through 28 which are
29:48probably now going to be delayed given
29:49the political dynamics they have. So,
29:52um,
29:52>> everybody's worried about over supply.
29:54I'm like more worried about
29:55>> massive massively under supply. Yeah,
29:57exactly. Which, which Okay. So, then if
29:59that's the scenario,
30:01like you could see a scenario where you
30:02see, you know, big price increases
30:05actually to access the intelligence.
30:06Yeah.
30:06>> Which is the opposite direction of where
30:08everybody thinks this is going to go.
30:09>> Yeah. Well, Dorcash had a wild point. I
30:11forget what it was, but he was positing
30:14>> um I forget the
30:15>> like the cost of a token could go up 10x
30:16or something like that. Yes. Yeah,
30:18>> which is crazy, but like we do live in a
30:20supply demand world.
30:21>> Like it's conceivable if the demand goes
30:24massively. And by the way, the whole
30:26premise of this that's happening so far
30:28is that there's a massive amount of
30:30consumer or user surplus being
30:32generated, right? So like why do people
30:34select the frontier tokens when they
30:36could use the cheaper tokens to do most
30:38tasks? There's many reasons why, but
30:41like the biggest one is because there's
30:42a tremendous amount of surplus even if
30:44you're using the frontier tokens, right?
30:45Absolutely. And so yeah, what happens if
30:48there's like a massive supply shortage?
30:50Well, I think that would be the, you
30:52know, kind of funny the consequence of
30:55like the these like data center
30:58degrowthers
31:00um
31:02may be like real compute inequality
31:06where big companies and wealthy people
31:09can afford compute and then you know two
31:12years from now they'll be on about that
31:13and it's like well that happened because
31:15of you. Yeah.
31:16>> You know that happened because you
31:17wouldn't let us build data centers.
31:19>> Yeah. And by the way, we've we've seen
31:20this, right? Like the path to a lowcost
31:23product delivered to consumers in a mass
31:26market is advertising. It takes a long
31:29time to build an advertising business.
31:30>> Yeah.
31:31>> Um as we've seen with all the, you know,
31:33consumer internet businesses that we've
31:34invested in over the years.
31:35>> Um and so there may be a disconnect in
31:37the period where you can't actually
31:38offer that.
31:39>> Yeah.
31:39>> And that would be a terrible outcome.
31:41>> That'd be a terrible outcome for the
31:42world. Nobody wants that. So we need to
31:43build a lot of data centers.
31:44>> Yeah. Exactly. Exactly. Yeah. like a
31:46compute in inequality like future that's
31:50that's not a good that's not a good
31:51future for anyone which is another
31:52reason open source is so important
31:54[gasps] and just one of the things um
31:57you know I you know I had uh Grock make
32:00me make me like a meme of that like
32:03three-headed dragon and one of the heads
32:04is like kind of confused about like all
32:06of the really like stupid
32:09>> bearish AI narratives but people have
32:11this idea that open- source tokens are
32:14free they're
32:16And it's like it takes the exact same
32:18amount of compute.
32:20>> Yeah.
32:20>> All else equal to make an open source
32:23token as a you know Frontier token for a
32:26comparably sized model. Now there's a
32:29lot of nuances there but that's broadly
32:31true.
32:32>> It's just a question of what are the
32:33margins that are charged on top of that.
32:36And even then, the Kimmy license,
32:39something that I don't think a lot of
32:40people appreciate is the Kimmy license
32:43stipulates a 30% um share of any
32:46revenue.
32:46>> Yeah. Yeah. Yeah.
32:47>> So like Kimmy has taken a 30% cut of all
32:50the revenue generated on its and this is
32:53because it's open weights, not open
32:54source.
32:54>> Yeah. Exactly. Yeah.
32:55>> Yeah. But it's also extremely token
32:57hungry too, right? So it's more it's
32:59it's far even we're talking on a token
33:01basis, but on a task basis, it's far
33:03more inefficient. So it's very costly.
33:05>> Yeah. And I just always like Jensen,
33:08he's a great patriot, great American.
33:10Like we're so lucky to have we're lucky
33:12to have him and Elon like and I think
33:14like you know kind of when the when the
33:16history of the 21st centurion 21st
33:19century is written you know there was
33:20like the Victorian age I think this will
33:22be like the age of Elon and Jensen.
33:24>> Yeah. because they have they they are
33:26fundamentally altering kind of like the
33:28fabric of human society and civilization
33:31with AI SpaceX making humanity
33:33multilanetary Starlink you know bringing
33:36lowcost internet access to the poorest
33:38communities in the world which is
33:40amazing um which is you know something
33:43that people don't talk about but it's
33:44like an amazing you know you talked
33:46about consumer surplus that is an
33:48amazing surplus
33:49>> there was never there there was never
33:51going to be an economic case to build
33:53internet access in those places because
33:55of the cost
33:56>> and the willingness to pay and now you
33:58could
33:59>> without and any incremental internet
34:02capacity like is not going to be built
34:04in a traditional sense on Earth. It's
34:06going to come from space and so like
34:07that is a huge that is a huge unlock. I
34:08agree.
34:09>> It's a good thing but like we're you
34:11know we're like you know we should we
34:13should all be grateful for them because
34:15I do think that you know they're you the
34:17they're making the future as exciting
34:19and inspiring as possible. say we are in
34:22this supply crunch. Um it's so funny
34:25when whenever I talk about SpaceX and
34:27it's it's obviously near and dear to
34:28both our hearts. Um you know I I say
34:31like first of all the orbital data
34:33center stuff it's not like big buildings
34:36in space like it's helpful to actually
34:37think of it's like the size of an
34:38airplane.
34:40>> People are picturing like the Death Star
34:43like or the Pentagon floating around in
34:45space. That's not what it is at all.
34:47>> Yeah. It's a It's you know whatever the
34:49size of an airplane, right? Rack of 72
34:51whatever chips.
34:52>> Yeah. It's it's like five of us standing
34:55together is kind of roughly
34:57>> is like the wings
34:59solar wings.
35:00>> Yeah.
35:00>> And then you keep it in a suns
35:02synchronous orbit.
35:03>> So you have the radiator always in the
35:06shadow of the rack.
35:08>> That's how you cool it.
35:09>> And it's I like I can't it's very hard
35:13for me to engage. you know, there's all
35:14these people on X and they're like, I am
35:17a physics PhD and I this is impossible.
35:22Um, [laughter]
35:22and actually there's there's there's a
35:24friend who's another investor who
35:25actually is a physics PhD who had many
35:28um arguments with him and he's like, I
35:30am a PhD and this is impossible. And
35:32then he goes to the SpaceX day and you
35:35know he talks to the SpaceX engineers.
35:36He's like, well, I was wrong. And so
35:38like if let's say you're an astrophysics
35:42PhD, you are brilliant. You're hanging
35:45100 IQ points on me. Have you thought
35:48about this for an hour? Have you thought
35:50about it for 10 hours? Have you thought
35:52about for five hours? Cuz you have
35:5410,000 of the world's smartest engineers
35:56at SpaceX who've thought about this each
35:58for hundreds if not thousands of hours.
36:01And the sum of that working with like
36:04very sophisticated, you know,
36:06engineering tools is it's a solved
36:09problem. And in their minds, it's
36:10dramatically simpler and easier.
36:12>> Yeah.
36:13>> Than a Starlink satellite cuz a Starlink
36:15has to have the phased arrays and move
36:16around.
36:18>> I think it's like So, okay. So, assume
36:20that you're right. I say it's like
36:23physics. There's not a physics reason
36:26why this can't work. Costwise, it seems
36:30really imposing, but kind of the history
36:33of the Elon companies is the cost curve
36:36gets dramatically better. Like when we
36:37first invested in SpaceX, you know,
36:40Starlink like was not commercially
36:42available and like we had all these
36:45questions about how the economics would
36:47proceed over time. The same on the
36:49launch side, the same with the Model 3.
36:51Like I I I just have to think that that
36:53will get solved paired with the fact
36:55that we're going to have massive under
36:56supply self-inflicted on Earth.
36:59>> Uh it feels clear to me at a minimum it
37:01will be swing capacity.
37:03>> Yeah.
37:03>> And you know in the fullness of time
37:05maybe it will be larger.
37:06>> Well no it's really simple like if we
37:08use 50 and it is the people the question
37:11people should be asking about orbital
37:12compute which is the one SpaceX is
37:14focused on is Starship reusability.
37:17>> Yes. Because the math is like let's
37:19let's just say it's 50 billion a gig and
37:22let's just say 35 of that is it. Yep. So
37:24that's the same and maybe it grows a
37:26little because it's it's going into
37:27space. The rest is power, cooling,
37:31labor, all sorts of things that you
37:33don't need in space because you have the
37:35so you have the solar panel and the big
37:38radiator. Um [clears throat]
37:41and that call that's 15 billion and
37:44that's probably inflationary here on
37:46Earth.
37:46>> Yeah. Because [clears throat] labor
37:47fundamentally feeds into that. We just
37:48talked about what's happening to, you
37:50know, electrician. Um,
37:52>> yeah. Comp. Yeah.
37:53>> Yeah. Electrician
37:53>> materials are all going to go.
37:54>> Yeah. All of it. Yeah. We're going to
37:55have Yeah. We're going to run out of co,
37:57you know, we're we're the the copper
37:59bulls are, you know, focused on like
38:01copper shortages. All of it.
38:03>> Yeah. So that 15 billion is
38:04inflationary.
38:06And so what you have to compare it to is
38:08the cost of launch. And with Starship
38:10reusability, that goes to under a
38:12billion. So the economics just instantly
38:14flip. Now, you're always going to train
38:17on Earth. There will always be
38:19advantages to having, you know, GPUs
38:23right next to each other. Like there
38:25are, you know, speed of light
38:26limitations are a real thing. Latency
38:28matters. So, data centers on Earth,
38:30they're not going anywhere. I think
38:32they're going to continue to be very,
38:33very valuable. But an increasing
38:36fraction of the world's compute is going
38:39to be in orbit. And you know, Elon said
38:43that he and Jensen have co-designed a
38:44Reuben rack
38:46>> and they're it's gonna launch in the
38:48fourth quarter of 27.
38:49>> Yeah.
38:50>> And let's just say let's just say he's
38:52off by two quarters.
38:54>> Yeah.
38:54>> I mean, that's that's 2028.
38:56>> Yeah. That's still okay. That's pretty
38:57soon
38:58>> that, you know, as Brad Gersonner says,
38:59like nobody's really paying attention to
39:01this and it's like kind of happening in
39:03plain sight. And it kind of to me solves
39:06for something, you know, mids single
39:08just billions today,
39:10>> which by the way, you know, is like
39:12that's just like keeping share constant.
39:15>> Yeah. Exactly.
39:16>> You know, of like what's happening with
39:18>> not presumably taking any share on on
39:19Grockbot.
39:20>> Yeah. Yeah. From from three billion. And
39:22by the way, man, I would just I'd
39:23probably take the over with Grockbot.
39:25Yeah.
39:26>> I bet it's like
39:27>> changing by the day just based on my own
39:30usage and the number of people who are
39:31hitting their usage limits. And then you
39:33are starting to get from you know
39:35Grockbot like hey we're servers are
39:38overloaded every once in a while and
39:39like they have a lot of compute. Um so
39:42it's just like okay you don't want to
39:44debate orbital data centers
39:46>> no problem. Well like Starlink mobile
39:48like they have a pretty clear credible
39:51plan
39:51>> for how that's going to work and that
39:54you know wireless is you know call it
39:55another 8 900 billion of revenue that
39:58they address. So your yeah your mobile
40:00plus your broadband whatever it's call
40:02it like close to two trillion of a
40:03market
40:04>> and [snorts] then you have a really
40:06rapidly growing AI AR base.
40:10>> Yeah. AI AR you've got the cloud you
40:12know the sort of the cloud business.
40:14>> Yeah. Um so I don't think great you're
40:16an orbital computic no problem. It
40:19doesn't matter.
40:20>> Yeah. Exactly.
40:21>> We don't even need to. We could just
40:22look at things that are happening today
40:24with terrestrial compute, with cursor,
40:26with Grock, with Grockbot. By the way, I
40:29think X ads are, you know, we have
40:31telemetry.
40:32>> They're also growing.
40:33>> You know, I would expect at some point
40:35you'll have like a Starlink
40:37Grockbot
40:39um Xadvertising [clears throat]
40:41bundle. You know, kind of one of the
40:42ways Google built their cloud business
40:44as they bundled it with ads and like,
40:46hey, we're, you know, maybe you're
40:48bundling the ads with AI, but why not do
40:50that?
40:51>> Yeah. Yeah. I actually like the AI
40:53position that they're in because it's
40:55like heads you win, tails you win in the
40:57sense that their first party business is
40:59growing very fast and they they caught
41:01up to the frontier like very quickly.
41:04Yeah.
41:04>> Um and so they've made the very
41:07aggressive compute investments to enable
41:09that first party work.
41:11>> Um and that's the kind of heads you win
41:13and like tails you win. Say they
41:16overbuilt their capacity for what they
41:18need for inference or training. they
41:20have a very compelling sub six month
41:22payback on the compute side um you know
41:25with like massive scarcity supply and so
41:27I think that's a really good setup
41:29>> and there was a bare case that hey okay
41:31well in in a in the open AI anth
41:35anthropic maximalist view where they're
41:37the only two companies and they're
41:38designing their own chips then like
41:41where what's the room for anyone else
41:42well like I don't think they're going to
41:44have a reusable starship and multiple
41:46spaceports anytime soon and if the
41:48economics of computer such that orbital
41:51is where it makes sense increasingly
41:53going forward because Starship should be
41:55deflationary, you know, terrestrial
41:57cooling, you know, power should be
41:59inflationary. Well, like even in in a
42:01world where
42:03they fumble the ball with their first
42:05party AI applications, like they do
42:07still have
42:07>> they're a massive infrastructure
42:08business.
42:09>> Yeah. Yeah. I I'm I'm so fired up about
42:11the uh the Starbase Louisiana. Uh
42:13>> Oh, yeah.
42:14>> I can't wait to visit, man.
42:16>> So cool. Yes.
42:17>> Uh I was reading about it last night and
42:19uh yeah, it's sort of like it's now the
42:22they now have the infrastructure for you
42:23know thousands of launches a year.
42:26>> Yeah. And eventually I think you will
42:28see like these star bases in multiple
42:31places, multiple coasts all over the
42:34world.
42:34>> Yeah.
42:35>> Like you know at some point you'll
42:36probably see one somewhere in the Middle
42:38East. You'll see
42:40>> you know whatever European country is
42:41like the least bureaucratic at the time.
42:43You'll see one there. You know, you'll
42:45for I think you'll see probably one in,
42:47you know, whether it's Japan, South
42:48Korea, who knows?
42:50>> Yeah. Yeah. Yeah. Yeah. Yeah. It's
42:51pretty exciting.
42:52>> Yeah.
42:52>> Yeah. The uh the capability to do to
42:56call it, you know, whatever 5,000
42:58launches a year, like that feels very
43:00futuristic.
43:01>> Yeah. I mean, it's wild. And I do think
43:04a distinction that um you know, SpaceX
43:07really tried to kind of hammer home
43:08during their their IPO is there's a
43:11difference between reusability and and
43:13China. They did catch kind of a rocket
43:15using this um it was actually kind of
43:17ironic. It was this kind of juryrigged
43:19system of kind of wires. Yeah.
43:21>> That had actually been suggested on the
43:23SpaceX subreddit.
43:24>> Yes.
43:25>> Like seven or eight or n or no no it was
43:28before they landed the first Falcon. So
43:29it's like more than 10 years ago
43:31>> and like China's clearly paying close
43:33attention to the SpaceX subre subreddit.
43:36But that's very different catching that
43:38thing from what they're trying to do
43:39with Starship where you know the uh the
43:42booster gets caught with the things and
43:44then it gets moved and then the Starship
43:46gets caught and then it gets stacked, it
43:48gets fueled and just sent right back.
43:51Yeah. Two a day. Two a day per pad.
43:53>> Like those numbers add up pretty fast.
43:55>> And there and I do think I think they're
43:57engineering the pads for more than two a
43:59day if I
44:00>> Yeah. I think that's a conservative I
44:01think that's a conservative assumption.
44:02Yeah.
44:03>> Yeah. Um but I mean
44:05>> Yeah. What's the Okay, so SpaceX, like
44:06again, you and I have talked a ton about
44:08SpaceX.
44:10What's like the most futuristic thing
44:11that you think about with SpaceX? Like
44:15the 10-year Okay, so you and I were at
44:17this conference together and there was
44:19this whole debate about um among a small
44:22group of public investors of like what's
44:23going to be the the first 10 trillion
44:25company. And uh I think what you said
44:29was like I have no idea, but I know
44:30which one's going to be the first 20
44:31trillion dollar company. Uh, so like
44:35what's the most futuristic like product
44:37or market or technology thing about
44:39SpaceX that that you can think of?
44:41>> Look, I mean this sounds crazy, but
44:43asteroid mining is going to be a very
44:45real thing. We're going to capture, you
44:46know, there's asteroid psyche. It has
44:48more gold, silver, platinum, you know,
44:51every precious metal in it that exists
44:53in the Earth's crust.
44:55At some point, particularly with
44:57Starship, you will be, you know, and we
45:00may need that um lunar base to make this
45:02happen. You'll be able to cap capture
45:04these asteroids. You'll bring them into
45:07a stable kind of geocynchronous orbit
45:09over some, you know, Americanowned
45:12atal in the middle of the Pacific. Um,
45:16you know, no humans within whatever 50
45:18miles. you'll, you know, you can imagine
45:21like Optimus robots, you know, um doing
45:24doing the work. Yeah.
45:25>> Yeah. Doing the work. Um and then, you
45:28know, delivery to Earth is free and for
45:30sure some of it's going to burn up,
45:32>> but I think that's going to happen. And
45:35[clears throat] I always think um
45:38Jeff Bezos said something very
45:40interesting. He said, "I think in the
45:41future Earth is going to be zoned
45:44residential." And you know, somebody
45:46asked him, this is like 15 years ago,
45:47what do you mean by that? He's like all
45:49heavy industry will take place in outer
45:51space. And then this addresses the
45:53pollution concerns. It addresses
45:54everything.
45:55>> You know, people always get like really
45:57worried about, oh, you know, we still be
45:58able to see the stars.
46:00>> And it's just like I think it's hard for
46:02like the human mind to understand how
46:05big space is, how big outer space is,
46:08>> you know, it's
46:09>> we don't have to worry so much about
46:10emissions up there. Yeah.
46:11>> Yeah. Yeah. So I think that is um
46:16that's probably the most futuristic
46:18thing.
46:18>> But in terms of an economic application,
46:20but it does um [clears throat]
46:24I mean
46:26I I do think in the next few years
46:29you're going to have a fleet of
46:30starships land on Mars. Next few years I
46:33mean I don't know let's just say at the
46:35outside this is eight years away.
46:37>> Yeah. They're going to land on Mars.
46:39Going to have like, you know, a little
46:41ramp's going to come out of the PEZ
46:43dispenser and it's going to be a
46:44modified Starship, the Mars colonial
46:46transporter, and it's going to be wild.
46:48You're going to have Optimus robots
46:50holding American flags like walk down
46:54and then, you know, they're going to
46:56pull out a bunch of solar panels and
46:58batteries and racks of compute and
47:01they're going to set all of that up.
47:03they'll be dropping Starlinks,
47:05you know, and maybe the orbital
47:08mechanics don't allow this, but I think,
47:10you know, they'll they will figure out a
47:11way to have, you know, capacity. So,
47:14just think how crazy it is to watch like
47:16the views from Pathfinder,
47:18>> you know, or, you know, whatever these
47:20different, you know, Mars um
47:22>> rovers and stuff,
47:22>> rovers are and like, you know, 4K video
47:25through Optimus robots all over Mars and
47:28then after that there will be humans
47:30>> who can inhabit it. Yeah. Yeah. Yeah.
47:32That is crazy to think about.
47:33>> And that that's going to be an amazing
47:35moment for America.
47:36>> Yeah.
47:36>> Oh, I mean, think about the moon
47:38landing. [laughter]
47:39>> This is a little bit bigger. Yeah.
47:40>> Yeah.
47:41>> Yeah.
47:41>> Um, so that seems cool. Um,
47:45[clears throat]
47:45>> that's a good one. That's That's a good
47:46That's a good one. Yeah. Not a lot of
47:48chatter about that one out there. Yeah.
47:50But I think it's highly likely to
47:52happen.
47:52>> Yeah. Yeah. Yeah. So, you mentioned
47:54Microsoft.
47:55>> Yeah. and the bet that they made which
47:57is like a little bit of you know like
48:00Apple's the extreme kind of bet against
48:02the future kind of bet they made and
48:04like Microsoft is kind of a gradient of
48:06that.
48:06>> Yeah.
48:07>> Like what's your what's your outlook for
48:10their decisions?
48:12>> Well, I do think the world has gotten a
48:13lot friendlier for their strategy. Um
48:15you know they clearly tried to make a
48:17frontier model. They failed.
48:18>> Yeah.
48:19>> You know Satia said we're going to have
48:20our own models that are very
48:21competitive. Like I think he said that
48:2318 months ago. they don't have their own
48:25models that are competitive, but what
48:28you're seeing with um I think the future
48:32is an ensemble of models. You know,
48:34there's a paro curve. No one model is
48:36going to be the best at everything. And
48:38I think the future for certainly, you
48:40know, kind of the global, you know,
48:4210,000 biggest companies is you're going
48:45to take whatever the best open source
48:46model is, I think probably in the in the
48:49very near near future that's going to be
48:51an NVIDIA model.
48:52>> Yep. The labs making AS6 create very
48:55interesting
48:56>> incentives for to get into each other's
48:58business
48:58>> incentives for Jensen and everybody's
49:01well oh in a world where open source
49:02wins who funds the training well this
49:04the chip companies could fund the
49:06training yeah
49:06>> it's trivial to do a 50 to$100 billion
49:09training run uh you know for Jensen and
49:11maybe soon I do wonder if this is kind
49:14of Google's like super long-term play
49:17like they they they seem to like maybe
49:18have opted out of the frontier race for
49:20now um We're going to monetize our
49:23compute at high rates and we're going to
49:26um sell TPUs externally, but that
49:29generates so much cash flow and open
49:32source is getting closer and closer and
49:34closer to the frontier. And it just may
49:36be the winner is ultimately just who has
49:39kind of the most cash flow to to fund
49:40these big training runs. But I do think
49:43you're going to see American open source
49:45led by led by Nvidia get really close to
49:49the frontier like they paid that
49:51poolside acquisition was made for a
49:53reason. Poolside actually had a lot of
49:54really good American open source talent.
49:57I they're they're you know they're doing
49:58a lot of smart things but that that is
50:00really good for Microsoft and at some
50:03level almost every application software
50:06company because what you can do now is
50:08you can take a base model and Neatron to
50:11date has not had a lot of post-raining.
50:13It's kind of been a good pre-trained
50:15model that you could do with what you
50:16want. So if you take a really good
50:19pre-trained base model and then instead
50:22of sharing your own kind of enterprise
50:26context that's truly your IP that's
50:28truly the value you know of your company
50:31is like you know the context embedded in
50:33all of your data and like sharing that
50:35with a frontier lab you know that may be
50:37hazardous for your financial health.
50:38Yeah, certainly with like the shift in
50:40the ZDR policy like Yes.
50:43>> Yes. And so you take a really capable
50:46open source model and you do a lot of RL
50:49and supervised fine-tuning on your own
50:51data. So you own it and it's your model.
50:53>> Yeah.
50:54>> And then if intelligence is like a super
50:56important input into your business, you
50:59want to own and control your
51:02intelligence, its capabilities, its
51:04cost. And then what we've seen from a
51:07lot of companies and you know Grockbot
51:10my understanding is you know I think
51:11it's Gemini 3.7 flash
51:15>> um Grock 4.6
51:17>> and some Opus
51:18>> y
51:19>> and what you and behind a router
51:22>> and you will um
51:25>> and I'm sure Elon is very focused on
51:27having it all grow as soon as possible.
51:29>> Yeah. Yeah. Of course. Um but I think
51:31what you'll see these companies do is
51:34they'll have their own model on their
51:36data and it will work with one or two
51:38other frontier models. Um not not you
51:41know necessarily but just you know
51:43checking each other it'll be kind of
51:45transparent to you the the most frontier
51:47for planning and then have execution run
51:49by everything else that's lower costed.
51:50Yeah, absolutely. And so I think that
51:53feels like a very likely future to me.
51:57And that's a that is a much Microsoft
51:59friendlier future than one in which
52:01there's just only two dominant frontier
52:04models. And it certainly looks like
52:06there's going to be at least three with
52:08Grock. I do think you got to give Meta a
52:10lot of credit.
52:11>> They've done a great job.
52:12>> Yeah. And I mean they were out of the
52:13game and they got back in the game. And
52:15it's just it's kind of amazing. Who
52:16could have imagined a year ago, you
52:19know, when it was like Gemini was
52:21ascendant exactly that this is the
52:23scenario
52:23>> Gemini wouldn't even be in the
52:25conversation
52:26>> and Muse and Meta would be significantly
52:28ahead of them from a capability
52:30perspective.
52:31>> Um, so it's just, you know, this is
52:33>> kind of like the highest stakes game of
52:36like corporate chess ever played.
52:38>> And, you know, people, you know, some
52:39people have made bad moves, they've made
52:41good moves. You seem some people come
52:42out of the game, others come back in.
52:45Um, but a future where that future where
52:48it's a, you know, I don't know if we're
52:50going to call it multimodel, a hybrid
52:52model, you I don't know what terminology
52:54the world is going to settle on, but I
52:56think that's the future.
52:57>> Yeah.
52:58>> And I'm actually surprised. I think the
53:01best broad instantiation of that today
53:05outside of Grockbot, outside of cursor,
53:08outside of you know like Harvey's done
53:09some cool things with that
53:11>> where they've done it is actually just
53:13the Fireworks Nexus product.
53:14>> Yeah.
53:15>> Where you can Yeah. You can
53:18>> choose your frontier model.
53:20Let us take whatever open source model
53:22you want, RL it for you, for your data
53:24for Gold Coleman Sachs, for Morgan
53:25Stanley, for JP Morgan, for Fidelity,
53:27for A16Z. You have all your own data.
53:30you control your intelligence and we
53:32make it transparent behind a router.
53:34>> Yeah,
53:34>> I think that is like a very plausible
53:37future and that's clearly what um Lynn
53:42from Fireworks, she was the first one to
53:43say it and then Alex Karp and Satia,
53:46they both kind of like
53:47>> Yeah, they've they've taken their own
53:48version of it. Yeah.
53:49>> Yeah. But you know, Satia's essay of
53:51specialized intelligence, like I think
53:53it's very plausible,
53:55>> but this stuff is really hard to do.
53:58Like that that sounds easy.
54:00>> I was it sounds easy to describe like
54:02the way I describe it to people is like
54:04who gets to be the abstraction layer to
54:07the organization and the users with
54:09intel like of of intelligence. It's like
54:11the most whatever vi after space or
54:14position that you could imagine in
54:15business like in the history of
54:17business.
54:17>> Yeah, for sure.
54:18>> Right. I think it's like the answer is
54:19and again.
54:20>> Yeah. Yes. And for sure it's Yeah. Who's
54:22the arbiter of intelligence for global
54:24enterprises and probably consumers? I
54:27was a retail analyst and um
54:32you know everybody kind of thinks
54:33running one of these big chains is easy
54:36and there's a lot into it and it's like
54:38well it's really easy to start an
54:41American retailer in any category cuz
54:44America's so big it's worth over $50
54:46billion almost any category.
54:48>> Yeah. All you have to be able to do is
54:51have a fleet of a thousand stores in 50
54:54different states that have very
54:55different climates, consumer
54:57preferences.
54:59You need to have them stocked with the
55:01right products at the right time for
55:03that region at the right prices. They
55:06need to be staffed by friendly and
55:07knowledgeable employees who don't steal
55:09from you
55:09>> who turn over at 100% a year.
55:11>> Turn over at least 100% a year. The
55:13stores need to be clean and well lit.
55:15And if you can do that,
55:18presto, $50 billion dollars. Yeah.
55:20>> And like in the history of American
55:22business, like you can I mean it's more
55:24than one hand, but you don't have to go
55:27through many.
55:28>> Yeah.
55:28>> It's really hard to do. And
55:31>> having that abstraction layer, having it
55:35work, having it seamless is, I think,
55:38way harder to do than people think. And
55:41I do think what something I think is
55:42very interesting about cursor, I'd love
55:44your opinion on this is like everybody
55:47else in the lab space, you know, had
55:50this like
55:53we're creating a digital deity, you
55:54know, and AGI and ASI like we're
55:58>> and the Curser guys were just like we
56:00want to make great product.
56:02>> Yes. Exactly. in a in a strange way o of
56:05everybody at the frontier. Um probably
56:08Kerser and it was the most product
56:11focused.
56:11>> Yes.
56:12>> Yeah. I'd say in you know now they're
56:13part of SpaceX but that suits Elon and
56:16his mindset really really well.
56:18>> Yeah.
56:19>> Let's make it an engineering problem.
56:21You know create the model factory and
56:23then we need to have a really good
56:25product.
56:26>> Yeah.
56:26>> You know the you know the the the Tesla
56:29cars they're amazing. I mean it's I
56:30don't I don't know if you drive one but
56:32it drives
56:33>> everywhere. Yeah. Yeah. But like the
56:34what cursor figured out is
56:37>> they're they had I would say a similar
56:40instate vision as what those others guys
56:42had.
56:43>> It was just a different path to get
56:44there and it's sort of like a practical
56:45meet the customer with what with where
56:47they are meet the technology where it
56:48is.
56:49>> Um and I think you know they'll sort of
56:51they have already demonstrated that they
56:52kind of led their way up into autonomy
56:54from from that starting point.
56:56um coding is unique compared to
56:59everything else in knowledge work. This
57:01this would be like in support of the
57:02point that Microsoft is in a good
57:03position
57:04>> because it is verifiable and perfectly
57:07documented and like nothing else in
57:08enterprise
57:09>> is verifiable and perfectly documented
57:11and so it will be messy like that that
57:14leads you to a good you know bullcase
57:16for something like Microsoft that
57:17abstraction layer
57:18>> if they execute but it's really really
57:21hard to make it really simple for
57:24>> oh you know click my co-pilot link to
57:26all my stuff train a model yeah
57:29>> on our data
57:31convince me that you're not going to
57:32share it with anyone else and then put
57:34it behind a router that's seamless for
57:35me and continuously upgrade that open
57:38source model.
57:39>> Yeah. It's not just some middleware like
57:40it's very hard to do. Yeah. And and by
57:42the way, they're going to compete
57:43they're going to be competing with not
57:45only the labs to be that abstraction
57:47layer
57:48>> but data bricks. So like
57:51>> Palunteer um the inference the inference
57:54providers um the application companies
57:57right so like Harvey has done an
57:58incredible job of this and you know like
58:01legal has sort of in take off and um and
58:04and I think they can see the future of
58:06how to be that abstraction layer um and
58:08do the work um but like legal is also
58:11unique because it's very documented and
58:13it's somewhat verifiable tax we'll see
58:15that we see see things like that but
58:17like the the one and a half billion the
58:18really appealing brought by is going to
58:20be very messy to go get.
58:22>> Yeah. Although I do always think and um
58:24you know I think probably in their heart
58:26of hearts Harvey and Lora think oh if we
58:29solve this
58:30>> we could be that abstraction layer for
58:31everyone.
58:32>> I think probably in their heart of
58:34hearts cognition thinks something like
58:35that too.
58:36>> I think everybody thinks and by the way
58:38there's like massive validation of the
58:39category because Kirkland Ellis said
58:42>> we're going to spend 500 million bucks
58:43to build this ourselves. Like first of
58:45all you know like good luck that's going
58:48to be very hard. Yes.
58:49>> Um, but that actually tells you that the
58:52pie is really big, right? Huge.
58:53>> Yeah, it's massive.
58:54>> And that's it's and you know, just um
58:56and I'm sure they have a very smart head
58:58of a head of AI, but it's not like a
59:00$500 million onetime build. That model
59:04has to be continuously updated,
59:05switching out the base model. Then all
59:07of that has to happen transparently. But
59:10I think you're going to have this huge
59:11collision between,
59:13you know, products like Fireworks Nexus,
59:15these legal agents, coding agents, big
59:19companies like Microsoft,
59:21>> Data Bricks,
59:21>> Data Bricks, Snowflake coming up,
59:23>> you know, for sure. Um, you know,
59:26Salesforce, I think, is going to, you
59:28know, Salesforce and Workday and all
59:30these companies. This is like
59:31everybody's going to go after it. It's
59:33just going to come down to who executes
59:34the best and
59:37>> and this is just you know who has the
59:39lowest costs.
59:40>> Yes, exactly.
59:40>> But it's going to be very hard I think
59:42over time unless you're re if you're not
59:45vert vertically integrated you have to
59:47be so good to emerge as that abstraction
59:50layer.
59:50>> Yeah. Yeah. To be the lowcost provider
59:52very hard
59:53>> because yeah we you're just simply not
59:55going to be the lowcost provider if
59:56you're not vertically integrated if you
59:58don't own your own compute over the very
1:00:00long long term. Um and you know it's
1:00:04that's another reason like I um you know
1:00:06I increasingly look at these
1:00:07hyperscalers on EV to net PP&E.
1:00:10>> Yes.
1:00:10>> Because net PP& is compute and that is
1:00:13just what the market thinks you're going
1:00:14to monetize your fleet of compute at and
1:00:17you can kind of look at them and there's
1:00:18some pretty obvious inefficiencies too.
1:00:20>> Yeah. Yeah. Yeah.
1:00:21>> Yeah. Kind of an AI version of price to
1:00:23book.
1:00:23>> Yeah. I like the price to book. Okay.
1:00:26[laughter]
1:00:27>> Um so okay you you mentioned Jensen. you
1:00:29know, I I'd share your sentiment like
1:00:30he's like carrying this industry
1:00:31forward. Like tell me your thoughts on
1:00:33Nvidia.
1:00:35>> So, um
1:00:37I think he's in a very very good
1:00:40position and his strategy of being
1:00:42vertically integrated but horizont
1:00:44horizontally open and it's like okay
1:00:47like let's just say um
1:00:50you know let let's say there's some
1:00:52accelerator that emerges that is really
1:00:53really really really good. almost
1:00:56certainly it will be better if it can
1:00:58plug into and this is why like I know
1:01:00you have an accelerator investment my
1:01:03number one thing is if you're a
1:01:05semiconductor CEO the only thing you
1:01:08should ever say is thank you Jensen
1:01:11thank you for creating this opportunity
1:01:13thank you how can we work with you we
1:01:16want to enable you sure we're going to
1:01:18compete with you on the edges
1:01:20>> but you know my rule of thumb for
1:01:21accelerators every 1% share today is
1:01:23probably worth a hundred billion Yes.
1:01:25>> So there's no need to go head on with
1:01:28Nvidia.
1:01:28>> Yeah.
1:01:29>> Um just pick a niche, get your 1%. Make
1:01:33sure that you know
1:01:34>> is very big.
1:01:35>> He has he has nine chips. Yeah.
1:01:37>> Um you know he's got he's got multiple
1:01:39flavors of accelerators. He's got CPUs.
1:01:42>> He's got you know Ethernet switches. He
1:01:44has two kinds of GPUs. You know he's got
1:01:47you know we've gone from just um scale
1:01:49out networking being a thing. We have
1:01:50scale up scale out scale across now
1:01:52scale in.
1:01:53>> Yeah. So just try to find a way to plug
1:01:55into his ecosystem.
1:01:56>> By the way, this is not foreign. Like
1:01:58his biggest customers all have competing
1:02:01products with various of those nine
1:02:03chips.
1:02:04>> Yeah. And just try to find a way to plug
1:02:06in, but just be nice to him. Be nice. Be
1:02:11nice. It's all personal. Yeah. You know,
1:02:13and it's just like sometimes like, you
1:02:15know, you hear some of these and it's
1:02:17like, have you ever seen game tape of
1:02:20the Chicago Bulls when Jordan was is,
1:02:23you know, it's game 50 of the season.
1:02:25>> Yeah.
1:02:26>> And he's a little bored.
1:02:27>> Yeah.
1:02:28>> And the Bulls are down cuz, you know,
1:02:29they're up eight games. You know,
1:02:30they're up eight games over the number
1:02:32two person in their conference.
1:02:34>> And he's a little bored. And then
1:02:36somebody
1:02:36>> somebody talks
1:02:37>> Somebody who's you who's who's kind of
1:02:39young decides, I'm going to talk to
1:02:41him because we're beating him. And then
1:02:42he just looks
1:02:43>> and it's like
1:02:44>> and it's like
1:02:44>> it's the best. Those are my favorite.
1:02:46>> It's amazing. Yeah. Yeah. You We've all
1:02:47seen, you know, whatever the last dance.
1:02:50>> Just don't do that.
1:02:51>> Yeah. Exactly.
1:02:52>> You know, just just like, "Hey, Michael.
1:02:54Man, I'm so happy to be on the court
1:02:56with you." Like that's that's to that's
1:02:59that's the move. But the reason it's
1:03:00particularly important is because
1:03:03Jensen's data centers are financable.
1:03:05>> Yes. And it goes back to that point like
1:03:08let's say it's $50 billion
1:03:12um for an Nvidia data center you need a
1:03:16$15 billion equity check.
1:03:18>> Yeah.
1:03:19>> Okay. You can finance the other 35
1:03:21billion.
1:03:22>> Yeah.
1:03:22>> And it's not circular financing. I have
1:03:24a lot of respect for the people I have
1:03:25met from Blackstone and KKR and Apollo.
1:03:28Yeah.
1:03:28>> And they're underwriting each of those.
1:03:30>> Yeah. and they finance it. And then
1:03:34there's a residual value guarantee,
1:03:36which as long as that residual val value
1:03:39guarantee is less than the gross profit
1:03:40dollars he's getting from selling the
1:03:42chips into that data center,
1:03:44>> it's like essentially it's super NPV
1:03:47positive with very little risk for him.
1:03:50>> Um, and then he, you know, he gets a
1:03:52revenue share. So if you're um and his
1:03:56data centers are the most financable.
1:04:00>> Yes.
1:04:00In I like let's just say a good case for
1:04:04probably TPUs are the second most
1:04:06financable.
1:04:07>> It probably takes I don't know double
1:04:10the equity check at least. Yeah.
1:04:11>> And then the rates on the rest of it are
1:04:14higher.
1:04:15>> Yeah. Exactly.
1:04:15>> And so cost of capital is a huge
1:04:19advantage and that's why you just want
1:04:21to be part of his ecosystem. And you can
1:04:24see he's he's he has all these chips.
1:04:27He's acquiring land power and shell
1:04:29companies now matchmaking them with
1:04:31offtake agreements. I think one reason
1:04:33he's doing these RVGs is if he doesn't
1:04:35do them, it's kind of an anthropic and
1:04:37open AI dominated world because they can
1:04:39pay the most for compute. He can
1:04:42effectively help other people
1:04:44>> compete with anthropic and open AI.
1:04:46>> Yeah. In the same way that he stood up
1:04:47the neo clouds in the first place. Yeah.
1:04:49>> It's just democratizing compute which is
1:04:51good for the world. Again, I think he's
1:04:52a patriotic American. His his interests
1:04:54are aligned with that though with with
1:04:55with the patriotic American ones, right?
1:04:58Fragmentation, right?
1:04:59>> Fragmentation, no dominant AI. Exactly.
1:05:02Which is which is really good because
1:05:03he's like a he is a ruthless competitor.
1:05:06And it's awesome that his incentives
1:05:09around fragmentation of AI,
1:05:10fragmentation of models, and you know,
1:05:13fragmentation of power um are completely
1:05:16aligned with what's good for America.
1:05:18And just going back to open source, just
1:05:20like I I just can't take it that people
1:05:24think that Jensen is like the world's
1:05:27biggest advocate for open source and
1:05:29it's somehow the a giant risk to his
1:05:32business.
1:05:33>> Yeah, exactly. No, it's great for his
1:05:34business. It's great for his business.
1:05:35>> It's amazing for his business because it
1:05:37means that instead of, you know, having
1:05:39a 90% margin on top of a token made with
1:05:42an Nvidia GPU,
1:05:44>> maybe it's a 40% margin. So more of
1:05:47those tokens are going to be consumed
1:05:48which means you need more compute.
1:05:50>> Yeah, exactly.
1:05:51>> Um
1:05:52>> in a supply constrained world
1:05:53>> in a supply constrained world and you
1:05:55know let's just what percentage of the
1:05:57world's supply has he locked up?
1:06:00>> 70 80 somewhere in there. And then um
1:06:04>> you're talking about fab capacity.
1:06:06>> All of it. All of it. You know it's just
1:06:07because he's saw this coming before
1:06:09everybody else.
1:06:10>> Yeah. And all the system supply chain.
1:06:12>> Yeah. He's got he's got the fab
1:06:14capacity. Yeah. locked up. He's got DRAM
1:06:17capacity locked up. He's got NAND
1:06:19capacity. He's got laser capacity. He
1:06:22has capacitor capacity. He has, you
1:06:24know, what you need to make the racks.
1:06:27And it's just like he, you know, he used
1:06:28to say, if I go back,
1:06:32>> you know, 15 years, he'd say, "Listen,
1:06:33I'm making a two or three billion dollar
1:06:35bet every two years, and I'm moving
1:06:38really, really fast."
1:06:39>> Yeah. Now he's making these multiundred
1:06:42billion dollar bets, bringing the supply
1:06:44chain alongside him. He's bringing the
1:06:47financing alongside him by kind of
1:06:49standardizing it, making it easy for the
1:06:51very smart people at Blackstone, KKR and
1:06:53Apollo and Goldman Sachs and Morgan
1:06:54Stanley, JP Morgan to finance
1:06:57>> and like that is hard to compete with.
1:07:00>> Yeah. And you know it is um
1:07:04we um my firm trades we have a pretty
1:07:07big portfolio private portfolio
1:07:08companies uh that are semiconductors
1:07:11and it's just um you know Elon said a
1:07:15lot of people are going to learn a hard
1:07:16lesson in hardware and like I will just
1:07:19say I've learned a lot of hard lessons
1:07:20in semiconductor investing like you can
1:07:23you can bet on the best team and you
1:07:26tape the chip out you feel great okay
1:07:28we've taped it out and it and that's
1:07:30happening happening faster than ever
1:07:30right now.
1:07:31>> Yeah, it's happening faster than ever.
1:07:32You feel great about it and we're
1:07:34getting really good with the emulation
1:07:36and the simulations and you feel great
1:07:38about it [clears throat]
1:07:40and then um you know you'll experience
1:07:42this the chip comes back from the lab
1:07:44everybody you get a facetime from the
1:07:45CEO they plug it in.
1:07:47>> Yeah. you know, and like and then
1:07:50sometimes it doesn't work, you know,
1:07:52[laughter] it's just like
1:07:54>> Yeah, this famously happened with
1:07:55Cerebrus twice, right? Like, and they've
1:07:57powered through and like they've done
1:07:58great.
1:07:59>> Well, I don't I think the chip I think
1:08:00each Cerebrris chip worked, it just
1:08:04struggled to find product market fit.
1:08:06>> Yeah. Yeah. Fair.
1:08:06>> For the first two generations, the chip
1:08:08worked. It just didn't have product. And
1:08:10they've done great with it. Yes.
1:08:12>> Yeah. But there's a different thing
1:08:13between you, you plug it in, doesn't
1:08:14work at all.
1:08:15>> And it doesn't work at all. Exactly. And
1:08:17then it's like if it doesn't work at
1:08:19all, you might be back to the drawing
1:08:22board and hey, we need another, you
1:08:24know, hundreds of millions of dollars,
1:08:27billion dollars, and we've we've learned
1:08:29our lesson. It's going to work the next
1:08:31time two years from now.
1:08:32>> Yeah. Assuming you can finance it. Yeah.
1:08:34>> As Yeah. Assuming you can get financing.
1:08:36So, it's um you know, semiconductors are
1:08:40hard. Like the real world is hard. Like
1:08:43hardware is hard. and what he is doing
1:08:47at the scale he is doing at and the
1:08:49speed and bringing all of this alongside
1:08:52him cuz you know the land and the power
1:08:54has to come.
1:08:54>> Yeah.
1:08:55>> You know the entire supply chain has to
1:08:57come the financing has to come.
1:08:59>> And so given that he's you know 70 80%
1:09:04whatever we want to say you just want to
1:09:06plug into that ecosystem.
1:09:07>> Yeah. Part of why Elon made the decision
1:09:10he made right. Yeah.
1:09:11>> Yeah. which I also think was like a very
1:09:14high elo move.
1:09:15>> Yeah, totally.
1:09:16>> So, [clears throat]
1:09:17you've had everybody else try and build
1:09:19their own ASIC.
1:09:20>> Yeah,
1:09:21>> they've gotten up on stage. Sometimes
1:09:23they say negative things about, you
1:09:25know, Jensen or Nvidia or take shots.
1:09:28>> Um you I did think it was pretty smart.
1:09:31You know, the jalapeno team last night
1:09:33and we should give credit where credit
1:09:34is due. Jalapeno is the I would say the
1:09:39first good ASIC other than TPU or
1:09:42tranium I have seen from internal
1:09:44>> in a in a what seems to be a pretty
1:09:45short amount of time.
1:09:46>> Pretty short amount of time. It's
1:09:48impressive. We should give credit where
1:09:49credit is due.
1:09:50>> They do have a good team working.
1:09:51>> They have a good team. Yeah.
1:09:53>> Um so they had a really good team. I
1:09:55think they had a lot of advantages and I
1:09:57do think
1:09:58>> if you are a lab and you have the model
1:09:59and you see the direction of research
1:10:01that's a big advantage for designing
1:10:03your own chip. But then you go back to
1:10:04Nvidia and they work with everyone.
1:10:07>> Yes.
1:10:07>> And everybody, you know, keeps thinking
1:10:09it's going to really standardize. And if
1:10:11you look at the three big, you know,
1:10:13Chinese open source models, Deepseek,
1:10:15Kimmy, Quinn, they're kind of all um
1:10:19evolving in very different ways.
1:10:22>> Yeah.
1:10:23>> And they can, you know, they can all run
1:10:24on, you know, a more general purpose
1:10:26chip, um, a GPU, but you're going to
1:10:29need, if you want to specialize,
1:10:31>> Yeah. You're going to need general
1:10:32purposes at a minimum for the types of
1:10:33evolution you see from that. Yeah.
1:10:35>> So, um like I think he's I'm very happy
1:10:40his incentives as a CEO are perfectly
1:10:43aligned with what's good for America.
1:10:45>> Yes.
1:10:45>> Um
1:10:47so I just make sure your semiconductor
1:10:51guys do [laughter] not talk trash about
1:10:53Michael Jordan.
1:10:54>> Be nice to be nice to MJ. Be nice to MJ.
1:10:57Yeah. Exactly.
1:10:57>> Yeah. And then it's like, you know,
1:10:58sometimes it's like, you know, you tug
1:11:00on Superman's cape and you get
1:11:01confident.
1:11:02>> Yeah.
1:11:02>> You know, you get confident and you
1:11:04start to talk a little bit of trash.
1:11:06Well, you know, Superman sometimes he
1:11:07just flies away like that's what
1:11:09happened to the TPU team.
1:11:10>> Yeah. You know, and you know, Jalapeno,
1:11:13they're tugging on Superman's cape a
1:11:15little bit.
1:11:15>> Yeah. We'll see.
1:11:16>> We'll see. And it is kind of amazing
1:11:18that like
1:11:19>> Jalapeno did something that none of the
1:11:22big
1:11:23>> like I this is as competitive of a chip
1:11:26as I have seen. Yeah.
1:11:27>> But again, it's just competitive with
1:11:29one of his eight or nine chips.
1:11:31>> Yeah. One of his nine. Yeah, of course.
1:11:34>> They'll continue to work closely
1:11:34together. Yes.
1:11:35>> Yeah. They'll continue to work closely
1:11:36together. So, it's like, hey, that's
1:11:38great. You did the one thing. Well, to
1:11:40actually be competitive with him at the
1:11:41system level, you need another eight
1:11:43chips.
1:11:43>> Yeah. Exactly.
1:11:44>> Yeah.
1:11:45>> Um and he is at and you know, Dylan at
1:11:49some analysis talks about how he's the
1:11:52bank of AI. He's like he's the central
1:11:53bank of AI. He's the Federal Reserve of
1:11:55AI. Yeah.
1:11:56>> And so I actually think it was really
1:11:57smart for Elon instead of like
1:12:01>> competing, you know, with somebody who
1:12:03is
1:12:04>> fully aligned.
1:12:04>> Yeah. Fully aligned.
1:12:06>> Mhm.
1:12:06>> And I think that history is going to
1:12:08judge that to be a wise decision. In a
1:12:10world that is so supply chain
1:12:11constrained, it's actually really hard
1:12:13to tell what true customer preferences
1:12:15are, right?
1:12:16>> Because like you come out,
1:12:17>> Yeah. they'll take anything. Yeah. This
1:12:18is this is how you know that like very
1:12:19old whatever the price is held up of
1:12:21H100 is very high.
1:12:23>> Yeah. Yeah. And if you have a TSM
1:12:25allocation, you're going to be sold out.
1:12:27Yes.
1:12:27>> Particularly if you can get the DRAM to
1:12:29pair with it. You're going to be sold
1:12:31out.
1:12:31>> So, it's actually kind of hard to infer
1:12:34true customer preferences. And I
1:12:37actually think one of the best ways you
1:12:38can like see true customer preferences
1:12:41is the kind of deals they cut with chip
1:12:43companies. So, broadly speaking, you
1:12:46know, the first deal is where the chip
1:12:48company invests
1:12:48>> Yep.
1:12:49>> in a customer. And you saw TPU and
1:12:52Tranium, Amazon and Google do that with
1:12:53Anthropic. Yep. And that was to their im
1:12:55immense advantage because it really
1:12:57helped their businesses, I think, helped
1:12:58those chips really level up because you
1:13:00kind of need to use a chip. There's a
1:13:01cold start problem.
1:13:03>> And [clears throat] um
1:13:05and in that scenario, as long as the
1:13:07dollars you invest are less than the
1:13:09gross profit, you can't lose money. And
1:13:11then there's a scenario where you do the
1:13:13RVG,
1:13:14Blackstone finances it or whoever,
1:13:16Blackstone, Apollo, KKR, Goldman Sachs
1:13:18finances it. Um, and as long as that RVG
1:13:21is actually less than your gross profit,
1:13:23you can't lose money and you have upside
1:13:24probably through a revenue share on top
1:13:26of it,
1:13:27>> then there are deals where you give
1:13:30warrants away, but they're tied to um
1:13:33like a fixed price per million tokens.
1:13:35And as long as the performance of your
1:13:37chip kind of outruns the performance of
1:13:39your stock,
1:13:40>> you're going to do good in that
1:13:42situation. If you just give warrants
1:13:44away, it could be negative NPV because
1:13:46the better the does the more value
1:13:49that's captured by the person. Yeah.
1:13:50>> Yeah. And so you can kind of look at
1:13:52that hierarchy of deals and like infer
1:13:55something about true customer
1:13:56preferences.
1:13:57>> Yes. That's interesting.
1:13:58>> Yeah.
1:13:58>> So Nvidia does pretty good deals.
1:14:00>> Uh like Yeah. I mean there's a reason
1:14:03that people I consider smart are
1:14:06investing in their deals.
1:14:07>> Yeah, I see it. Gavin, thank you. Fun.
1:14:09Always fun to hang with you.
1:14:10>> Thanks, David. This was great, man.

About this transcript

This page contains the full transcript of Why AI Demand Is Outrunning Compute Supply by a16z, generated from the public captions YouTube serves with the video. The transcript has 14,393 words across 2,242 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.