YT YouTube2Text

YouTube2Text → Transcripts → Why the Markets Are Pricing AI Wrong | Gavin Baker

Why the Markets Are Pricing AI Wrong | Gavin Baker — Transcript

by Invest Like The Best · 13,308 words · 2,334 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I want to be scared, you know, I don't
0:01want to feel like a lunatic watching
0:04these stocks get more cheaper thinking
0:07the expected forward returns are growing
0:09up. I may be kind of missing out here
0:12this week.
0:13>> pressure test?
0:14>> Yeah.
0:14>> Yeah.
0:15>> Find tell me something negative, but I
0:18haven't been able to find one that is
0:20like a quantitative metric. The
0:21underlying fundamentals are improving.
0:24Uh and stocks Nvidia's actually has we
0:27record this
0:28at its lowest forward PE of the last 10
0:32years. The market 100% thinks they're
0:34significantly overvalued.
0:48>> Gavin, [music] it's only been 2 months
0:49like the model release cycles, the gap
0:52between our podcast episodes are
0:54shortening.
0:55>> [laughter]
0:55>> We're basically you and I are basically
0:57on a model release cadence at this
0:58point.
0:59>> Well, I was I was sensitive to criticism
1:01that um
1:02that I think somebody pointed out that
1:04um our podcasts were coincident with
1:07like
1:08local market peaks.
1:10>> [laughter]
1:11>> And nobody can say that after this.
1:14>> What's on your mind? It's been a crazy
1:16crazy month.
1:18>> Yeah, I would describe
1:19um
1:20July as 2022
1:22in a month.
1:23>> Yeah.
1:23>> There are some fundamental negatives
1:25which you which we should talk.
1:27But like on on the whole, the balance of
1:30fundamentals I think is improving
1:33significantly. Loads of AI names are
1:36down 50 60% from their highs. We'll call
1:40it 40 to 60% in a month in a straight
1:44line.
1:45And I asked you before we started,
1:47you've you've been out here for the
1:48summer. Have you heard a single negative
1:53quantitative metric about AI?
1:55>> Yeah.
1:55>> A single instance of deceleration.
1:57>> Nothing.
1:58>> Nothing.
1:59In fact, every metric is accelerating.
2:02>> And to your point, not just blind
2:03optimism from people excited about AI.
2:06Here's some data that they can show you
2:08and from their different vantage points.
2:09>> Absolutely. I mean, however you cut it,
2:12whether you cut
2:13GPU availability, whether you cut GPU
2:15retail pricing,
2:17I mean, whether you cut like the spot
2:19price of DRAM this month,
2:22token growth,
2:24everything is actually accelerated.
2:27And I do think a big part of the problem
2:30is
2:31one, the market does not have visibility
2:35into Anthropic, OpenAI, and then I would
2:37say these open-source inference clouds
2:39that monetize inference here in America,
2:41Fireworks, Baseten, Model together.
2:44>> [snorts]
2:44>> And the picture looks very different
2:46when you see that. Because open-source
2:49is accelerated massively because of GLM
2:515.2 KiB K3.
2:53And then, you know, Neurotron continues
2:56to kind of chug along. We had a great,
2:58you know, very small American
3:00open-source model release. OpenAI has
3:02accelerated.
3:04Anthropic continues to grow
3:06really strongly
3:08and is almost certainly pumping out
3:10significant amounts of free cash flow.
3:13And I just think if, you know, there's
3:14this chart that everybody looks at of
3:17semiconductor cash flow going like this
3:20and hyperscale cash
3:22free cash flow going like that,
3:24and is you're missing these private
3:26companies. And then, I also think that
3:28that chart,
3:30um,
3:31misses something very important, which
3:34is just that you have everyone in '24
3:36and '25 thought, even if you were really
3:39bullish, you thought that GPU prices, if
3:41you were really bullish, you thought
3:42they would to price around a GPU would,
3:45you know, decline slowly. You know, if
3:47you're bearish, you thought it would
3:48decline precipitously.
3:51I I think anyone in 24 or 25
3:54thought that the prices of old GPUs
3:58would still be would be going vertical
4:01in 2026.
4:03Yeah, and so everybody thought hey,
4:05we're going to be smart. We're going to
4:07sign these long-term contracts.
4:09And to some degree like a lot of the
4:10deal clouds had to do that because they
4:12needed an off-take agreement to finance
4:14the GPUs.
4:16And so essentially you have the
4:18contracted base of installed compute
4:22trading at a massive discount
4:25to the current spot market.
4:27And
4:29has those
4:31contracts roll off and compute gets
4:33repriced higher.
4:35It's spot could decline and compute will
4:37still get repriced higher.
4:40You know, I think you're going to see a
4:41lot of acceleration that's going to
4:43answer these ROI questions.
4:45You've started to see that this quarter
4:46if we look at operating cash flow, not
4:48free cash flow. Operating cash flow
4:51from Microsoft, Meta, and Amazon has
4:53reported accelerated from 28 to 32.
4:56There were some actually pretty big
4:58unusual items now like these
4:59hyperscalers they always seem to have
5:02like
5:03billions of dollars of legal expenses
5:05that are unusual.
5:07Mostly fines to the EU.
5:10But there was an unusual amount of
5:12one-timers this
5:14this quarter. If you adjust for that, we
5:15went from 28 to 35 and that's that's a
5:18that's a material acceleration at this
5:20scale.
5:21And that's really before
5:24like they start to light up the Rubins
5:26which will come at a meaningful premium
5:28before these contracts reprice. It's
5:32been a it's been a it's been a
5:33challenging month that it's almost um
5:37you know, like is it helpful to kind of
5:38like walk through the month? How we got
5:40here? You know, so first there's Meta is
5:43going to rent out compute. And this is
5:45seen as like very bearish. They have
5:47excess capacity. They're going to cut
5:49CapEx.
5:50This is a disaster. This is not at all
5:53what it was. They just reported. They
5:55didn't cut CapEx. What it was is they
5:58saw SpaceX have a big installed base of
6:01compute
6:02and sell some big trading optimized
6:05clusters into the market at a truly
6:08massive premium to these contracted
6:10rates.
6:11And
6:13you know, at least the at least at least
6:15analysts like that, they saw an
6:16opportunity. There's a lot of
6:18speculation they're going to raise
6:19capital. So, like
6:21you know, maybe what they're thinking is
6:23like, "Hey, we will show on a small
6:25chunk of capacity that we can generate
6:28really strong IRRs. Then we'll go raise
6:31equity capital and and we'll and we'll
6:33be off to the and we'll be off to the
6:35races and probably raise CapEx."
6:37It doesn't look like what that's what
6:39they're doing. But nevertheless, the
6:41market sold off because it interpreted
6:43this very negatively.
6:45And
6:46I was really sure it wasn't negative.
6:49You know, there's a lot of telemetry
6:51into Meta's CapEx plans. None of that
6:53telemetry had shifted at all. If
6:56anything, it was, you know, continuing
6:58to or continued to get more aggressive.
7:01And then shortly after that, they
7:02released their best model in a long
7:04time, use 1.1, which is actually really
7:07a very good model. I mean, it was
7:08overshadowed by Grok 4.5,
7:11but it was a good model.
7:12Um way better than anything in two
7:15years. So, just
7:16no chance they're taking their foot off
7:18the gas. Then Kimi comes out.
7:21And then there's this huge freak out
7:23about open source. And at the same time,
7:25this silica data
7:27token index kind of dips and flattens.
7:31And the two are connected. What the
7:32silica data token index captures is mix.
7:36And they don't see all the tokens, but
7:38because of GLM 5.2 and and and and then
7:41Kimi, all of it took a while to layer
7:43in.
7:44There's kind of a
7:45a mix shift in this data from more
7:48expensive frontier tokens, which
7:50probably have an inference margin
7:52we can debate whether it's 80, 90, or
7:5495.
7:55>> Yeah.
7:55>> But super high.
7:57Towards open source tokens.
8:00And for whatever reason, the market
8:02thought this was negative, but the
8:03reality is a token is a token, and you
8:05need the exact same amount of compute
8:08to make a token all else equal. Takes
8:10the same amount of flops, the same
8:12amount of memory,
8:14the same amount of watts. Now, tokens
8:16are not equal, but broadly speaking,
8:18all open source taking
8:21share does is kind of
8:23take margin dollars
8:26out of the
8:28uh frontier model layer
8:30and effectively by
8:32thereby, you know, there there is
8:33elasticity, thereby driving token
8:36demand.
8:37You need more demand for compute. And
8:39the margins, you know, Anthropic and
8:44open source, they all run on the same
8:46underlying cloud providers
8:48who charge the same amount of compute.
8:51You know, so you're literally just um
8:53taking margin from frontier models and
8:56essentially driving more margin dollars
8:58into the AI infrastructure layer. And
9:01like I think that's
9:03>> That was the catalyst. This this
9:04combination of things.
9:05>> Well, yeah, then it kept it it's it's
9:07like Jenson is the world's largest
9:09supporter of open source.
9:11Do we really And he's like a super
9:13idealistic guy. He's a patriotic
9:14American.
9:15I think he always does what's right.
9:19But is it does it really stand to reason
9:23that Jenson would be the world's biggest
9:25supporter of open source if it was bad
9:27for his business?
9:29>> [laughter]
9:30>> He'd still support if it was the right
9:31thing for the world.
9:32>> Yeah.
9:33>> But maybe it wouldn't be a signature
9:35issue.
9:36>> Yeah.
9:36>> And by the way, I think open source is
9:37really important to world where there's
9:39just one or two dominant for tier
9:40bottles that charge like
9:4290% margins.
9:44It's not good for humans. It might not
9:46be good for society. And I think we want
9:48a lot of bottles as we've discussed
9:50before.
9:52So then it's like, okay, the market
9:54digests that comes through with it. Then
9:55China has a DUV machine and this causes,
9:59you know, everybody said these baskets
10:00has caused a huge sell off in semi-cap
10:02equipment.
10:04And then we get to what I think is
10:07in a lot of ways,
10:08um
10:10like the real concern, which is real
10:12yields have gone up, which makes sense,
10:14you know, we're investing a lot
10:16to fund this investment and for sure
10:18credit is an increasing part of it even
10:21if the majority is still funded over
10:23what the majority is still funded out of
10:24operating cash flows.
10:26And so real yields go up and spreads
10:29widen. Meta Meta priced, um
10:32a bond last week and, you know, it it it
10:36did not price where you would think a
10:37meta bond would price. And this just
10:40shows that the credit market
10:42>> And the media CDS was was blowing
10:44>> All of these CD CDS for everybody is is
10:46blowing out.
10:49And, you know, very smart private
10:51capital people just like that, hey, this
10:53is just exactly
10:55what you would expect. These are just
10:56banks, you know, kind of hedging
10:58hedging their commitments. But
10:59nonetheless, it doesn't look good and
11:01these are undeniable facts. CDS is up,
11:03spreads are widened, real yield real
11:05yields are up.
11:07And that is that would be really really
11:09scary if we needed debt
11:14to finance this build out.
11:16And that's where I think it's this
11:19differential between spot and contract
11:22pricing for the installed base of
11:24compute is so important.
11:27>> Ramp is the only platform built to make
11:28your finance team leaner, faster, and
11:30better, saving businesses 5% annually
11:33[music] on average, so you can stay
11:34focused on growth. Ramp customers grew
11:36revenue 3.2 times faster than the
11:38average American business. [music] Visa,
11:40Foursquare, Cursor, Stripe, Notion,
11:4211:FS, Shopify, and 70,000 other
11:44businesses all run on Ramp. Mine does,
11:46[music] too, and so should yours. Learn
11:48more at ramp.com/invest.
11:51>> The best AI and software companies from
11:53OpenAI to Cursor to Perplexity use
11:55WorkOS to become enterprise-ready
11:56overnight, not in months. Visit
11:58workos.com [music] to skip the
12:00unglamorous infrastructure work and
12:02focus on your product.
12:04>> Felix by Rogo is a personal finance
12:06agent that turns a single prompt into
12:07finished client-ready [music] work using
12:09your firm's own templates, contacts, and
12:11standards. Send Felix an email like,
12:13[music] "Take these comments and turn
12:14them for me." Or, "Update my tracker
12:16with the context of these emails." Or,
12:18"Run the ability to [music] pay math on
12:20this buyer." And Felix sends back
12:21finished PowerPoint decks, Excel models,
12:23and sourced [music] research. Felix
12:25works the way your team already does,
12:27delivering work quickly and accurately
12:29around the clock. Learn more at
12:30rogo.ai/felix.
12:32[music]
12:33>> It's so important to understand what the
12:35financing will be like for the next 6
12:37months or something.
12:38>> The degree to which this buildout is
12:42going to require credit.
12:44>> Right.
12:44>> Which would be the classic like capital
12:46cycle, and then we start to overextend
12:48ourselves with debt, and that's where
12:49things get dicey.
12:49>> 100% to the, you know, debt-fueled
12:52buildouts, you know, they demand
12:53immediate repayment. Yeah. So, if supply
12:55and demand get a little bit out of
12:57whack, things can unwind very, very,
12:59very quickly. That's what happened to
13:00the internet.
13:02And so,
13:04if
13:06one believes as I do, rightly or
13:08wrongly, and I'm like, after this month,
13:09I'm super open,
13:11you know, I'm I'm looking like I've been
13:13pressure testing all of these, and like,
13:16I really went deep on credit because,
13:18hey, this is real, it's undeniable.
13:21And if we need credit to fund this
13:24buildout,
13:26this is like a significant negative.
13:28>> [snorts]
13:29>> And
13:30if you model it out,
13:33has
13:35if you look at the amount of gigawatts
13:36that are supposed to come out
13:39in consensus estimates
13:41for hyperscalers, they're effectively
13:43modeled and these are gigawatts of
13:45Blackwell and Rubin.
13:47Rubin being Nvidia's next chip,
13:49Blackwell being the current chip.
13:52They are essentially modeled
13:55to monetize roughly at the rate of
13:57Ampere,
13:58which is two generations behind. Not at
14:00Hopper, but Ampere. So, there's 1.3
14:03trillion and 1.3 to 1.4 trillion in
14:05hyperscale operating cash flow.
14:08If you just assume that they
14:11they're not I I think it's very unlikely
14:12they monetize at the rate of Ampere and
14:14we can we can go into why. Some of it
14:16comes from just, you know, seeing what
14:18is happening on the ground with demand
14:20here for real quantitative metrics.
14:24But, like, let's just say they monetize
14:25at a discount to current Blackwells.
14:28Then it's more like 2 trillion of
14:31operating cash flow
14:32and that kind of takes 700 billion of
14:36credit demand out. Um
14:40and, you know, and then obviously these,
14:43you know, ironically, has, you know,
14:45that improves all the credit ratios, has
14:48these installed bases of compute
14:50reprice.
14:53We're going to continue accelerating.
14:54Consensus is modeling in a deceleration,
14:56which I think is unlikely.
14:58Um then the credit metrics look better
15:00and then all of a sudden
15:01it gets easier to finance with credit.
15:03Now,
15:04whether they whether they they choose to
15:06do that or not, we'll see, but this this
15:08is all a little bit, um you know, I
15:10think we spoke
15:13>> 2 months ago.
15:14>> No, but the time before that about kind
15:16of the risks of a Blackwell air pocket
15:18where you're spending
15:20hundreds of billions of dollars on
15:21Blackwells. they're mostly being used
15:23for trading initially. Trading does not
15:26generate, you know, a return.
15:29And that this could be a risk. And
15:34we actually really saw that kind of it,
15:36you know, in the first quarter.
15:38And
15:39I think one reason, you know, like to
15:42the podcast two months ago,
15:44I I got comfortable with that risk was
15:46just that you were seeing such
15:48incredible things out of Anthropic.
15:50And then it's like, "Okay, well, the
15:51market's kind of going to look past
15:53this."
15:54And it did look past it in April, in
15:56May, in June.
15:58And then in July, because of this kind
16:00of confluence of things, stopped looking
16:03past it.
16:04Just as the operating cash flow
16:07started to really accelerate. And it's
16:09this is just a fact. It is accelerating
16:12at big scale.
16:14Um
16:15and you know, like Microsoft, they
16:17brought out a huge slug of capacity in
16:19the month of June. That didn't even show
16:21up in the second quarter.
16:24So essentially, what this
16:26all comes down to is do you believe that
16:29the kind of
16:31quantitative demand signals
16:34seeing on the ground here in Silicon
16:36Valley from from private companies are
16:38going to continue such that
16:41the installed base of compute reprices
16:44higher as contracts roll off.
16:46>> Operating cash flows go up.
16:47>> Yeah, operating cash flows go up and you
16:49can fund this out of most of this out of
16:51operating cash flows. Maybe all of it.
16:53Like if it reprices at current rates,
16:56you could probably fund all of it for
16:58the next several years.
17:00It's so it's it's been a it has been a
17:03very un-
17:05usual episode in the market.
17:09And
17:11you know, in some ways, the fact that
17:14and we should talk about what the
17:15fundamentals are that are getting better
17:16that I'm talking about.
17:18You know, technicians would say it's
17:20actually in '22,
17:22okay, the market is worried about a
17:23recession,
17:24rates going up,
17:26you know, inflation. That's what the
17:28market was worried about in '22. You
17:30knew exactly what it was. Okay, deep
17:31seek, you know what it's worried about.
17:33Liberation day, you know what it's
17:34worried about. Uh there's something very
17:36clear and it in a weird way that's that
17:39is comforting reassuring. And here, you
17:41know, we talked about a lot of specific
17:42things, but it just feels all those
17:44specific things with the exception of
17:46credit
17:48like are are just kind of ridiculous.
17:51Um it's to the fact that it is still
17:54going down.
17:57You know, a technician would say, "Hey,
17:58that's
18:00that's a little scary. You know, it's
18:01definitionally the bullet you don't see
18:04that gets you." You know, I think we've
18:05talked before about how like I think the
18:07three most important words in investing
18:09aren't margin of safety, but I don't
18:10know.
18:13But just, you know, I've you've you've
18:15been out here for 2 months. I've been
18:16out here, you know,
18:17I literally spoke to a company this
18:19morning who rented a cluster of several
18:23and this is one of, you know, kind of
18:25sexiest startups that people want to be
18:27in business with.
18:29And they had rented a cluster of several
18:30thousand block wells.
18:32And we'll just call it, you know,
18:34somewhere in the mid $2 per GPU hour.
18:37They're renting the exact same cluster,
18:40exact same size cluster,
18:42essentially identical in every way,
18:44B200s. Said, "No no differences."
18:47And they're hoping
18:497 months later
18:51to pay just under $4.
18:54Like you know, just you hear this today.
18:56And that's
18:57like that's pretty crazy because again,
18:59you would just you would expect a really
19:02like a gentle decline in prices would be
19:05bullish.
19:06Instead, you know, we're up, you know,
19:08depending on the starting point
19:1050 to 60% in 6 or 7 months.
19:15And it just there've been so many
19:17anecdotes like that. Like I think one of
19:20the inference clouds
19:21I think it was based in I'm not sure.
19:23They went on a podcast and they
19:24essentially said
19:26we are planning to pay 100% more
19:30for Blackwell's when our contract
19:33expires. And that just means that
19:35essentially all the hyperscalers are
19:36under running.
19:38And I haven't like my main kind of
19:41mission out here this week
19:43>> It's like pressure test?
19:44>> Yeah.
19:45>> Yeah.
19:45>> Find tell me something negative, you
19:48know? Like you know, the question I
19:50asked you, have you is there one
19:52negative quantitative metric you've
19:53you've heard? Has been what I've been
19:56asking everyone.
19:58>> The main thing people are saying is the
19:59Anthropic like the third-party data
20:01suggests that the Anthropic like curve
20:03started to
20:04go off of its trajectory a little bit.
20:06That's like the only thing that I
20:08>> I think I think that's I think that may
20:10very well be true, but then you have
20:12OpenAI and open source massively
20:15accelerating.
20:17And if you look at the sub
20:19it is not accelerating. Maybe I don't
20:22know that it looks the same. I think it
20:24may have accelerated. Like I think open
20:26source is a little bit of a
20:28you know, they talk about dark matter in
20:30the universe. Like open source is kind
20:31of dark matter to the public markets.
20:33You know, it's hard for public markets
20:35to measure it.
20:37But like if you just track what these
20:39inference clouds are saying
20:41and you know, these are people saying
20:43things on podcasts or people saying
20:44things in meetings
20:46they're not you know, audited financials
20:49but like demand is clearly accelerating
20:51which makes sense cuz you had this huge
20:53capability leap which you learn 5.2 and
20:55KBK3
20:57which I think we're going to see
20:58continue. I think you're going to see
20:59Nvidia bring Neobtron steadily closer to
21:03the frontier. It has been a very like
21:05it's been a hubbly
21:07challenging month and but just
21:11it's also like wow, I've kind of
21:12pressure tested every assumption.
21:16The underlying fundamentals are
21:18improving.
21:19Uh and stocks Nvidia's actually as we
21:22record this
21:23at its lowest forward PE of the last 10
21:26years.
21:27>> Crazy.
21:28>> The only time the Sibbys have been
21:30cheaper were liberation day deep seek
21:32and that was
21:34those were kind of uh V bottoms.
21:36Um
21:37>> And that means to you just that the
21:38market thinks they're significantly over
21:40earning?
21:41>> Yeah, the market 100% thinks they're
21:43significantly over earning.
21:45And you
21:47we need to be humble.
21:48>> Maybe they are.
21:49>> Maybe they are.
21:50Um
21:52but like my kind of mission out here
21:53this week was to look
21:56for negative data points as hard as I
21:59could. I normally come to Silicon Valley
22:02and you know, there's a mixture of like
22:04okay, here's here's something negative,
22:06here's something positive, da da da. On
22:08balance, it's positive, you know, tech
22:11it creates value over time.
22:13But I haven't been able to find one that
22:15is like a quantitative metric. Other
22:17like that that Anthropic third-party
22:19data, I would say that seems to be a
22:21hotly contested by the um
22:24by the Anthropic shareholders who are
22:26like
22:26>> [laughter]
22:27>> who are bound or kind of like chopping
22:30at the bit to tell you what they know.
22:32They're also very scared they're not
22:33going to get an IPO allocation
22:35>> [laughter]
22:35>> if it gets back to the company that
22:37they're the ones who said, "Actually
22:39things are great." You know, you can
22:40just see Anthropic shareholders
22:43like they want to be like, "It's not
22:44true." You know.
22:45>> [laughter]
22:47>> I mean, it's hard for me to believe that
22:50um
22:51open source and open AI have accelerated
22:52to the extent they did and but yeah,
22:54Anthropic is clearly, you know, kind of
22:56in the
22:57in the pole position. And oh, by the
22:58way, you know, Grok and Cursor have
23:01also, you can see from third-party data,
23:04like July was a pretty transformational
23:06month with um
23:08Grock 4.5 Grock builds coming out. So,
23:12it has been a tricky month and um
23:18and I have I have a friend um
23:20I have a friend at Fidelity
23:22who just says the way to have navigated
23:25like the last 3 years
23:28is just do the dumbest, most superficial
23:32thing as quickly as possible and just
23:35cycle between them.
23:36>> What is that? What is that now?
23:39>> Well, that's just that has been to cut
23:40risk
23:41all month in response to these
23:44kind of narratives that just
23:47like factually except for credit
23:51are not true and the work we've done
23:53makes me think that credit just isn't
23:55going to matter has this reprice. Let's
23:57just say you do need credit
24:00to like build the flops we eat. Well, if
24:02credit's not there, it just means the
24:04flops that are there
24:06are going to be even more valuable cuz
24:08there is an interesting like essay that
24:10got sent sent to me.
24:12You know, I think we've talked before
24:13about Mike Mauboussin's theory that like
24:15a breakdown of diversity is kind of what
24:17leads
24:18you know, to bubbles and crashes.
24:21And essentially everyone I know in the
24:22public equity investment business,
24:24whether retail or institutional
24:27everything immediately, every piece of
24:29news gets fed into Claude.
24:32And Claude Claude code, sometimes, you
24:35know, a Claude agent
24:38and you know, Claude it's probabilistic.
24:41There's probably not that much variation
24:43in the way it's interpreting this news.
24:46It's uh it's almost like we're back to
24:49um
24:50you know, in stock market terms
24:52like the like there's never really been
24:54this way in the stock market before, but
24:55people talk about the fragmentation of
24:57media and how it used to be like Walter
24:59Cronkite was the only voice of truth and
25:02now we don't have that anymore.
25:04It's like Claude
25:06is kind of Walter Cronkite for the stock
25:08market and everybody just believes
25:11>> Whatever it says.
25:11>> Whatever it says. [laughter]
25:13And this is leading to like
25:16>> really and and by the way, it's really
25:18smart,
25:19but it's not
25:21always right. It's not
25:24um it's interpretation isn't always
25:26correct. And with the stock market, you
25:29are fundamentally dealing about, you
25:30know, a probabilistic Bayesian
25:32interpretation of the future.
25:34And so it just it feels like
25:37in the market, there is this
25:40Here's this piece of news. It gets fed
25:42through Claude. Claude interpreted this
25:44way.
25:4590 a huge chunk of people
25:48trade on Claude's view.
25:50Um
25:51And so you've seen stuff. There's this
25:52guy uh TBU, TBU. He's like part of like
25:57the autonomous semiconductor mafia, but
25:59he posted this amazing chart of Japanese
26:01capacitor stocks.
26:03And he said we've had a capacitor an
26:05entire capacitor cycle in 6 weeks. And
26:08it's true, you know, the stocks like
26:10whether they double, triple, or
26:11quadruple, I don't know, but like
26:13vertical.
26:14And then whoosh, you know what I mean?
26:17Like the actual fundamentals haven't
26:20even hit. And yet you've already had
26:23what probably would have normally been a
26:243-year cycle
26:26in like 6 weeks.
26:28>> What's your sense of being out here
26:29especially it makes me especially
26:30curious about this, the innovation that
26:33is going on here to improve the
26:36efficiency and every aspect of serving
26:39inference, of training models, et
26:40cetera, and how that will affect like
26:42public markets over time. Like have you
26:44learned anything interesting about like
26:46the long lead time innovation type stuff
26:48that has you especially excited or or
26:50curious?
26:51>> Yeah, I am very curious. It was
26:54like all there seemed to be
26:58like a lot of people seem to feel like
27:00they're very close
27:03to solving continual learning and
27:04sample-efficient learning, which we've
27:06talked about before. And it is possible
27:09that if those are solved that, you know,
27:12could that be like a temporary like kind
27:16of like discontinuity?
27:18You know, it demand if it's that of, you
27:20know, having to like I think somebody
27:22told me that the uh
27:24like I was traded on effectively 20
27:27billion tokens, and then it's like these
27:29models are traded on 300 trillion
27:30tokens. And if, you know, you could
27:33trade something on 10 trillion tokens
27:35and then let it out into the world and
27:37learn sample efficiently,
27:39you know, that that doesn't sound good
27:41for trading demand, but like trading as
27:43a percentage of
27:45semiconductor demand and compute
27:47is going to asymptote to something not
27:50approaching zero, but very small. But I
27:52would say that is the most kind of
27:54interesting, and you know, who knows if
27:56it's long horizon or short horizon.
27:58You know, SSI says that they're going to
28:00come out, you know, with their their
28:01model in in August. You know, there's
28:04this whole generation of new labs that
28:06are focused on this.
28:08>> And this would be good for the world.
28:09>> This would be amazing for the world,
28:10yeah. This would be awesome for the
28:12world.
28:12>> Yeah, we all want we want this, right?
28:13>> Yeah, we want this. It would be amazing
28:15for the world, and it's just it's hard
28:17for me to believe that that would
28:18actually be negative for AI
28:20infrastructure demand. But again, trying
28:23to be really, really open-minded. I I
28:25would say that was probably
28:28like the biggest like what what do we
28:30call it? Scientific or technical
28:32takeaway.
28:34But it's just, you know, it's also
28:35>> We still don't know.
28:36>> Well, yeah, and also like Nvidia is
28:38heavily involved with
28:40all of these startups.
28:42>> So, what would like if I was forced to
28:44if If just forced to come up with
28:47the set of circumstances that would
28:49really switch you around and get you
28:51really scared.
28:52Is it would it just be
28:54that this operating cash flow thing
28:56doesn't play out and therefore we just
28:57need to debt finance this whole thing?
28:58>> cash flow does not continue to
29:00accelerate. That that would be negative.
29:03Um it that to some degree is going to be
29:06a function of how Anthropic, Open AI,
29:10Grok Cursor, we should call it Grok, and
29:12open source
29:14you know, if like if there was a pretty
29:16dramatic like
29:19contraction in GPU prices that was kind
29:22of sustained, I mean the market would
29:24react to that instantly. That would be
29:26worrisome. If it started to get to be
29:27really easy to get GPUs,
29:30I mean have you heard anyone say they
29:32have too many GPUs?
29:35>> [laughter]
29:35>> Like not not a single person. And it's
29:37not like it's the opposite. It sounds
29:38like a drug market or something. It
29:40really does. It's just wild. But yeah, I
29:42mean I think there's a long list of
29:44pretty obvious things. You know, if like
29:45Anthropic, Open AI, if the sum of these
29:48labs
29:50plateaus or you know, starts to decline,
29:53that's really negative unless it's just
29:55because open source tokens
29:58are not growing the pie and taking
30:00share. Uh and I do really think the
30:02future is like multi multi model.
30:05Yeah, I think particularly for the AI
30:07natives, they're going to want
30:11to take an open source model. It's got,
30:13you know, all these inference clouds
30:14have got really good at um you know, at
30:18supervised fine-tuning and reinforcement
30:19learning. So you can take your data,
30:21customize an open source model, and then
30:24get something that you can
30:26put behind a router, and the router
30:28routes it to often first your model, and
30:31then Claude, a frontier model, whatever,
30:34Claude, Grok, um checks it, and you can
30:38in a lot of cases get slightly better
30:41outcomes
30:43at half the cost.
30:45But again, that half the cost, I think a
30:47lot of people hear that, they're like,
30:48"That's bad for AI demand." It's
30:49actually not at all because the cost the
30:53user pays has
30:54you know, it's just a function of the
30:56margin on the tokens, and you're
30:58literally just shifting
31:01tokens from really expensive tokens with
31:04like 90% gross margins to tokens with
31:07maybe, let's call it a 30% gross margin.
31:10And that's where the savings are coming
31:11from, but the tokens cost the same
31:13amount of compute
31:15to produce.
31:16And then also all these things are kind
31:18of happening
31:21at kind of um at different cycle times.
31:24You know, all these, you know, big
31:26public companies are like, "Oh my god,
31:27my AI spend is 20x. I've burned my
31:29budget in 3 months." So, they set up a
31:32router,
31:33and that actually cuts their AI spend,
31:37but it doesn't really impact. It may
31:39actually increase the amount of tokens
31:43that they are generating just by
31:44shifting them to these cheaper
31:46open-source tokens, and that's just more
31:48compute.
31:49So, you know, a company getting smarter
31:51about which model to use for which task,
31:56that you know, that may lead to a a a
31:59stabilization of their spend or even a
32:00decline, but it actually has nothing to
32:03do with the amount of
32:05you know, GPU compute hours
32:08they are effectively consuming
32:11behind
32:13you know, these the these model layers
32:15of this router. The GPU compute hours
32:17probably are going up as you, you know,
32:20shift to these cheaper tokens you can
32:21use more of.
32:24So, and then, you know, that's happening
32:26to like a cutting-edge of public
32:28companies,
32:29and then you have this whole wave of AI
32:31natives, and
32:34like they're leading into this so hard,
32:37and they're not hiring
32:39humans. They're just really putting it
32:41mostly into tokens. And so, they're not
32:44slowing down. And then you have
32:45companies on the East Coast of America
32:48who have like barely adopted AI,
32:50companies, you know, broadly speaking,
32:52on other, you know, not on the coast who
32:55maybe aren't as cutting and then Europe
32:57who's like just trying to figure out how
32:59to regulate AI,
33:01>> [laughter]
33:01>> before using it.
33:02>> Yeah, so just like there's kind of these
33:04differential differential kind of waves
33:06of adoption all happening at the same
33:08time.
33:10But the thought I can't get out of my
33:11mind is like I think I said it maybe
33:13last time, but just Yoc's estimate like
33:15I don't know,
33:16500,000 people in the world, 250,000
33:20maybe are using a genetic AI.
33:23And we're in a cute compute shortage.
33:27That's
33:28Do you know there's seven or eight
33:29billion people on the planet?
33:31What happens when we go from 500,000
33:34>> to 1%
33:34>> 100 billion, you know, to 500 billion?
33:39And then I do think it's it it it is
33:41interesting, you know, a lot of people
33:42are just like, okay, well, you know, I I
33:44do think it's like helpful to post on X
33:46to see the pushback.
33:48And a lot of people are saying, well,
33:50you know, where
33:52fundamentally is the Okay, we accept
33:55your argument that hyperscalers are
33:58under earning in this compute re-prices.
34:00Are their operating cash flows going to
34:02accelerate and maybe we can fund this,
34:03but like
34:05who Where is that operating going to
34:06come from? Where is the customer?
34:09And kind of definitionally it has to
34:11either come from, you know, faster
34:13economic growth through productivity
34:15kind of Satya's comments like either
34:17we're going to start growing 10% or
34:19we're not.
34:21Or labor substitution.
34:23And for sure, I think in a lot of these
34:26AI natives,
34:28you're seeing labor substitution, but
34:29not because they're
34:31firing people, they're just not hiring
34:33nearly as many humans. You know, the
34:35gross profit dollars per FTE and you
34:37know, A16Z iconic, a bunch of companies
34:40that have done this work,
34:41you know, they're you know, they're
34:43they're vertical
34:44uh particularly relative to
34:46past generations of startups.
34:48And then it is interesting, you know,
34:50like are you kind of doing any surveys
34:52of
34:53your companies that their tokens bid
34:55relative to labor spend?
34:56>> Oh, yeah. I mean, it's always reported
34:58as a percent of percent tokens as a
35:00percent of like total comp spend or
35:02something like that.
35:03>> what are the ranges you've seen?
35:05>> I mean, like in the really pill
35:06companies, like it gets really high.
35:0820%, 25%, something like that.
35:11>> Well,
35:11our
35:12our your our friend Dylan Patel at
35:14Jasper
35:15>> [laughter]
35:16>> So, he's an ASI maxi, but he's at 30%.
35:19>> Yeah.
35:20>> Uh
35:20>> He probably that's probably the highest
35:21one I've heard.
35:22>> Uh I've actually heard a 50.
35:24And there's 25 trillion dollars in
35:26knowledge work. And so, let's you know,
35:29let's say that that's you know, let's
35:31take your 20% number.
35:33That's 5 trillion and that either comes
35:36out of labor substitution or faster
35:38economic growth.
35:40And we really, really, really want to
35:42have, you know, humans it to come from
35:44faster economic growth.
35:46>> One interesting thing I heard this
35:47morning from one of the great like
35:49leading technology CEOs has founded
35:51several companies that if you look at
35:52the founder letting controlled companies
35:54and adjust for some of the like COVID
35:56era, you know, over hiring, like
35:57nobody's really laying people off. Like
36:00these are the people that would probably
36:02be most quick to adopt AI to you know,
36:05become more efficient or whatever. Like
36:06they're not really doing jack aside like
36:09huge scale layoffs, which probably tells
36:11you something about where they think
36:13there will be lots of opportunity to
36:14still have people plus
36:15>> 100%
36:17well, the bull case
36:18>> So, growth not labor not labor growth.
36:20>> the bull case and you know, you've seen
36:21charts from Cognition, Ramp and Stripe
36:24that the companies that are spending the
36:26most on AI are growing growing
36:27meaningfully faster.
36:28>> Yeah, I love that cognition index.
36:30>> Yeah, the cognition index is wild. Now,
36:32all the skeptics will point out
36:33rightfully, it's not really controlling
36:35for industry, but then if like you dig
36:37down into it,
36:38you know, I think one of them gave an
36:39example of I forget if it was a plumber
36:41or an HVAC contractor, but like, you
36:44know, and everybody who's a blue-collar
36:46workers doing great cuz of AI.
36:49By the way, something that I think we
36:50should touch on and we we could do it
36:51now or later
36:53is just everybody is citing these LTAs.
36:56So, that we're everything is at a
36:58shortage. Everything is at a shortage
37:00right now. You know, if if there's
37:01weakness, it's just cuz we can't
37:02energize the gigawatts fast enough. The
37:05gigawatts are going to get energized
37:06like it, you know, regulatory policy is
37:08moving in a in a good way. You the
37:10turbine manufacturers, the diesel gen
37:12set manufacturers, you know, they're
37:14ramping up.
37:15>> You're you're you're ripping
37:17turbines off old airplanes and, you
37:19know, reconditioning them and then
37:21repurposing them. There's crazy things
37:23happening. Capitalism is very, very good
37:24at this.
37:25But I do think one of the most important
37:27questions in the market
37:29and like a transition of the market that
37:31like I got wrong
37:34is we are shifting, particularly for
37:38particularly for memory more than
37:39anything else,
37:41from, you know, crushing numbers
37:45in in the short term
37:47to they are trading short-term upside
37:49for these, you know, what do they call
37:51them? Supply chain agreements, long-term
37:53agreements, LTAs,
37:55where they essentially, you know, agree
37:56there's there's many flavors, but the
37:57customer prepays
37:59and it's, you know, there's a floor and
38:01a ceiling.
38:03And this comes back to the
38:04point about labor because, you know, a
38:06lot of people after
38:08um
38:10you know, after kind of like firing,
38:13you know, too many people,
38:15were, you know, you during during COVID,
38:17were really reluctant to lay people off.
38:20And that, you know, they talked about
38:21labor hoarding if you remember a few
38:22years ago. You remember this?
38:25I'm just
38:27Let's just think about the game theory
38:28of breaking an LTA.
38:31So, there's four companies that like
38:34matter at scale. There's Amazon with
38:36their trade ups.
38:37There's Google with their TPUs.
38:39There's [snorts] AMD. And then there's
38:41Nvidia, who's like much bigger than
38:43everybody else combined. You know, let's
38:45just say it's 2027.
38:48It's very important to realize memory is
38:51The more memory you put
38:53with flop for a given unit of compute,
38:56the more tokens you get out. It's the
38:58single most important thing you could do
39:01to increase kind of token output per
39:03unit of compute. And then that obviously
39:05definitionally actually lowers costs,
39:08which is why the demand hasn't responded
39:10at all negatively. There's been no
39:12elasticity just because it's like kind
39:15of the only It's the axis that is
39:17dominating all others.
39:19Um And this is like at some level like a
39:22giant Game of Thrones or Imposters
39:24between these companies.
39:26And okay, it's 2027.
39:30You're like or 2028. You're vaguely
39:33tempted to break one of these LTAs. Try
39:36and get a lower price.
39:38But to a large degree, market shares are
39:41I think for the next several years are
39:43going to be determined by supply by
39:45supply chain allocations and kind of
39:48what you have
39:49kind of pre-purchased.
39:51So, if you break the LTA
39:53and you And this is This is assuming
39:56we're not in a severe oversupply
39:58situation.
40:00And but all the logic almost the game
40:02theory even holds in a severe oversupply
40:04situation. If you break your LTA
40:07and then in the next
40:10two or three years for any reason
40:13leverage shifts back to the memory guys,
40:16you're out of business. It's over.
40:19You know, like let's let's just say
40:20Google breaks an LTA. You know, there's
40:23there's an over there's an over supply
40:24of making this up in 28, 29.
40:27They break their LTAs. Well, if they're
40:29breaking their LTAs, it probably means,
40:30you know, you're over supply, prices are
40:32coming down. And then, you know,
40:33capacity that naturally contracts.
40:36Well,
40:37like what do you think's going to happen
40:39to Google's allocations? And then, you
40:41know, this is a cyclical industry and
40:43over supply is followed by under supply.
40:45What do you think they think is going to
40:47happen to their allocations next time?
40:49So, I just think given that this is like
40:52the access around which kind of
40:54everything is revolving,
40:57man, like you might blow up your entire
41:00business
41:01and your franchise by breaking an LTA.
41:04And that was never the case before, you
41:05know, Apple, who cares, you know,
41:07they're buying they don't have a
41:09competitor.
41:10They're the over they're overwhelmingly
41:12the largest purchaser. They know they
41:15can do whatever this is, you know, going
41:16back three, four, five years. They know
41:18they can do whatever they want with no
41:19consequences cuz their volume is so big,
41:22you know, that even if they like super
41:24screw Hynix, Micron will of course take
41:26them.
41:27This is this is just different, you
41:30know, you have at least four players.
41:33Then you have all the startups. You're
41:34an investor in Etched.
41:36And
41:37if you break an LTA,
41:40and that they just say, "Okay, fine. You
41:42know what? Great. You broke the price
41:44agreement.
41:46We're going to break the volume
41:47agreement. And, you know, screw you.
41:50We're going to give the volume to your
41:51competitor."
41:52You just you just lost share, you know?
41:55That's so I think the
41:56the you know, it like I think, you know,
42:00Nvidia's dominance I think is uh
42:05like I think
42:08the current environment they state to
42:09which it favors Nvidia,
42:12like it is a hard for me to understand
42:14why it's trading at such a low multiple.
42:16You know, in other words, like if you
42:17need to be able to finance the chips and
42:19you do, nothing's more financeable than
42:21an Nvidia GPU. Nothing.
42:24If you need to get, you know, land and
42:26power,
42:27well, they're doing a very good job of
42:30playing that chess game and and
42:31matchmaking.
42:33And then they've kind of rolled out this
42:34really clever, you know, new business
42:36model, which I would describe as kind of
42:37like a credit wrapper
42:39um with a revenue share if GPU prices
42:43are above a floor. Yeah. Um and this
42:46could lead to them like having a really
42:49giant cloud business effectively through
42:51royalties really quickly.
42:54And it is another way of kind of
42:56alleviating this um
42:58you know, cash flow mismatch. Like, hey,
43:00we're making all the cash.
43:03Yeah, and and like
43:04this isn't this isn't really vendor
43:06financing cuz they're not loading them
43:08the money. Somebody else is loading
43:11the GPU buyer the money. So, it's not
43:14quite vendor it's not vendor financing.
43:16It's um
43:18it's, you know, they're still making
43:19equity investments, but it's not it's
43:21not like you're just putting money into
43:22someone
43:24in return for them, you know, you know,
43:25and then some of that money, you know,
43:27is used to buy your chips, even though,
43:28you know, Nvidia said that they write
43:29into all their
43:31you know, equity investments that um you
43:33know, the money can't be used to buy
43:34Nvidia chips, but obviously money is
43:36fungible.
43:37And um
43:37>> Funny thing.
43:38>> What's that?
43:39>> like a funny little thing.
43:40>> Yes. [laughter]
43:41Um
43:42Yeah.
43:43But, you know, I think at some level it
43:45probably makes everybody feel better.
43:47Um
43:47>> What would you do if you were the memory
43:49like if you were the CEO of Hynix?
43:50>> I'd do the exact same thing Nvidia's
43:52doing right now.
43:54>> Which is?
43:55>> I I would be going
43:56>> I say, all right, I'm going to say
43:57>> be going to the buyers of GPUs, Radeon,
44:00and whoever
44:01and say, I'll participate in the Nvidia
44:05credit wrapper. Now, their business is
44:07just inherently less stable and
44:09predictable,
44:10but in some way, and maybe they just put
44:14up some cash up front, so it's like
44:16they're not on the hook. You know what I
44:17mean? I'm just making this up.
44:19But like,
44:20do something like you can because you
44:23have money now,
44:25and credit markets are revolting.
44:29There many, you know, like, you know,
44:31the the
44:32people, I'm sure the, you know, our
44:34friends at, you know, Blackstone and
44:36Apollo are suggesting some variant of
44:39this to the memory companies. But hey,
44:41we will like
44:43put up some amount of money from our
44:45cash flow today, and then it's gone.
44:48It's, you know, surety
44:50uh that they, you know, makes the the
44:52the person who's extending the debt feel
44:54better,
44:55but we want
44:56a some sort of a cut
44:59of the ongoing revenues as well. Right.
45:01Like that is like 100% what I would do.
45:05And it's almost like a logical extension
45:07of, you know, the LTAs where they're
45:09kind of trading upside for durability.
45:12Here, you know, you can
45:15you know, you can effectively get a
45:16royalty on recurring revenues. And that
45:18is that is what Nvidia is doing. And I
45:20do think that is very misunderstood.
45:24And I think it would serve Nvidia well
45:27to really
45:29explain this. One, they're really
45:32bullish on AI.
45:33Um
45:35Essentially, every time they haven't
45:36taken an equity stake in something, it's
45:38been a mistake.
45:39You know, I mean,
45:41they've taken equity stakes in
45:42everything essentially except the memory
45:44companies that for a long while
45:45Anthropic that they took an equity stake
45:47in Anthropic. But like, why not if you
45:50have cash flow and you're bullish on AI?
45:52And Jensen because he sees every lab. He
45:55knows all the advances, you know, like
45:57all these continual learning labs, you
45:59know, safe superintelligence is now
46:00working with them.
46:02You know, he he sees everything and like
46:04what he sees makes him bullish. Um
46:07So, what have some equity upside
46:09and then two, have a revenue share
46:12and you're generating hundreds of
46:14billions of dollars of
46:16um
46:18of free cash flow um and a helping
46:21to kind of bridge, you know, what what
46:24is clearly kind of a gap at least, you
46:26know, given everybody's gone free cash
46:28flow negative
46:29until the operating cash flow
46:30accelerates enough that you can
46:31internally fund this.
46:33It's almost like I mean it's um
46:36very opportunistic and it like
46:37significant and in a good way
46:40and it significantly increases their
46:42revenue per gigawatt. And then it also
46:44strengthens their competitive position.
46:46You know, that's you know, you and I, we
46:48both have startups, but okay, that's
46:49that's that's great. Use that startups
46:52chip. Um
46:54well, what prices are they paying big
46:56atomic to me? Higher than Nvidia and all
46:58these guys. What prices are they paying
47:00for um HBM D-Ram? Higher. Um
47:05Can you finance those chips easily at
47:06the same rate as Nvidia? No. And so it's
47:09always like, you know, there's there's a
47:11real burden, particularly
47:14if you use HBM D-Ram, like you're just
47:16you're in the crosshairs of this. Um
47:19unless like actually you you know, maybe
47:20actually like they made really different
47:24architectural choices. Everything that's
47:26happening
47:27is actually pretty good for hip. Just
47:30are going back to game theory.
47:33Entropic
47:34if they had been as aggressive on
47:36compute as OpenAI had been, they would
47:38have run away with it.
47:40And so now OpenAI is back in the game. I
47:42think Grok is in the game. Those are the
47:44companies on the Pareto frontier.
47:46>> And they have the compute.
47:47>> And do you think
47:49after watching that
47:51anyone is going to let off the gas?
47:54Cuz you just you know, it was
47:57I think 4 months ago that Dario was
47:59talking about how
48:01you know, it was a real it was a really
48:03thoughtful commentary, but he's like,
48:04it's really, really hard because, you
48:07know, if you buy too much compute, you
48:10could go bankrupt at the scale of these
48:11things.
48:12But if you don't buy enough, you could
48:13lose.
48:15Well,
48:15>> We saw it happen.
48:16>> OpenAI just got back into the game, and
48:18now SpaceX is in the game in a big way
48:20of Grok 4.5 and Cursor.
48:23And like, after watching that, from a
48:25game theory perspective, is anybody
48:28going to back off anytime soon,
48:30especially if it can be funded out of
48:32operating cash flow?
48:33>> Vanta automates security and compliance
48:35for over 16,000 fast-moving companies
48:37like Ramp, Cursor, and Harvey, keeping
48:39them audit ready around the clock. It's
48:41the number one agentic trust platform,
48:43[music]
48:44and it now helps companies like yours
48:45watch for the risks that show up between
48:47audits across your [music] vendors, your
48:49AI tools, and your whole environment.
48:51Every new tool your team signs up for,
48:53every vendor that turns on AI features
48:55is an opportunity for something to go
48:57wrong, [music]
48:57and most security programs weren't built
48:59for AI's pace of growth. The Vanta agent
49:01works like a 24/7 [music]
49:03GRC engineer in the background finding
49:05issues, drafting fixes for you, and
49:07cutting vendor assessment time by up to
49:0950%.
49:10Whether you're [music] a fast-growing
49:11startup or a global enterprise, Vanta
49:13helps you earn and prove trust.
49:15Invest Like the Best listeners get a
49:17special offer for $1,000 off at
49:19vanta.com/invest.
49:23Ridgeline is the first end-to-end system
49:25of record with embedded AI for
49:27investment management [music] firms,
49:28running portfolio accounting,
49:29reconciliation, reporting, trading, and
49:32compliance
49:33>> [music]
49:33>> on one unified platform.
49:35Firms are moving off legacy technology
49:37and onto Ridgeline because of how far
49:38ahead Ridgeline's AI features are
49:40compared [music] to anything else in
49:41investment management software.
49:43I've been hearing from a lot of
49:44investment managers about AI, and they
49:46fall roughly into two camps, with some
49:48unsure of where to even start, and
49:50others convinced they can build their
49:51own order management system over a
49:53weekend. [music]
49:54The reality is that running an
49:55investment firm will always require
49:56governance controls and a single source
49:58of truth for your data, and no amount of
50:00AI enthusiasm changes [music] that
50:02requirement. Ridgeline is built on
50:04exactly that foundation, which is why I
50:06believe that firms that come out ahead
50:08in the AI era will be the ones running
50:10on Ridgeline's unified [music] platform.
50:11If you're serious about your firm's AI
50:13strategy, Ridgeline should be part of
50:15that conversation. You can request a
50:17demo at ridgeline.ai. [music]
50:19>> Have you met anyone in your travels out
50:20here that you would say is like way more
50:22bullish than you, and if so, what do
50:24they believe that you don't?
50:25>> I mean, essentially everyone out here is
50:27more bullish than me, man.
50:29>> [laughter]
50:29>> I like
50:31You know, I read this thing that
50:32Dworkesh wrote, and I was like
50:34>> The 3x compute price thing or whatever?
50:36>> Yeah, well, he was I forget what it was.
50:37>> No, no, it was like 15x or something.
50:39>> Yeah, but no, but just basically that um
50:42you know, renting an H100 for a year
50:44would cost $250,000.
50:47You know, um the salary
50:48>> that's 15x the current spot or
50:50something. Yeah.
50:51>> Exactly. Like, wow, you know, that was
50:54just like that
50:55>> in my book.
50:55>> That wasn't in my
50:58you know, forget my like Bayesian
51:00probability space of expected outcomes.
51:03That wasn't even in my
51:06>> [laughter]
51:06>> considered but dismissed as totally
51:09unlikely outcomes.
51:10You know, and then that guy is, you
51:12know, he's very Dworkesh, he's a very
51:13smart guy, he's very plugged in.
51:15And um
51:17and you know, and then he pointed out
51:18that like, hey, the you know, something
51:20like I think he just said margins on
51:23compute are going up, the amount of
51:24compute is going up,
51:26and inference margin's going up, and if
51:29you multiply those three, that's how
51:31you're getting this crazy acceleration
51:33in the sum of the labs plus open source,
51:35although obviously open source the
51:37margins on open source are not
51:40really going up. But I mean
51:41>> So, everyone's more [laughter] bullish
51:43>> Yeah, yeah, I like you know, I just
51:46I look at what's happening in the stock
51:47market and I feel like a foolish
51:49optimist.
51:51And then when I talk to people
51:55whether it's people at the labs, whether
51:58anyone in this ecosystem
52:00like I'm like bearish relative to
52:03essentially everyone.
52:04>> [laughter]
52:05>> Which is just a strange state of
52:06affairs. What do you make of the DUV
52:08news out of China where I've seen
52:10reactions really along a spectrum of
52:12like this is the equivalent of like what
52:14ASML had in 2001 or something.
52:17Or like no, this is actually the first
52:19bit of news in a a new story for how we
52:22should think about the global supply of
52:24cutting-edge compute.
52:25>> I think both could be true.
52:28You know, it's just like um
52:30like let's just make an analogy. Like
52:32let's just say
52:34a DUV machine was a jet turbine and now
52:36like an EUV machine is like warp drive.
52:39Um you know, or what whatever it's going
52:41to be, you know, a
52:43DUV machine is like a propeller plane
52:45EUV is like a jet turbine.
52:47Um
52:49but like they didn't have it before
52:52and
52:53now they allegedly do.
52:55And that is like a phase transition, you
52:57know, you it's like
52:59you've gone from like liquid to solid.
53:01Now that solid that you know, jet
53:04engine, prop plane, whatever is 25 years
53:07behind but still it's important and I
53:10don't think should be dismissed, but I
53:12also
53:14you know, it's kind of
53:15>> funny you just see this in the stock
53:16market, you know, it's like the stock
53:17market massively overreacts and then
53:20like if this ever hits ASML's orders,
53:23maybe it hits it in 5 years and like the
53:26market has forgotten about it, got
53:27worried about it, forgotten about it,
53:29got worried about it, forgotten about it
53:31multiple times um along the way. Um so I
53:36do think that was probably an
53:37overreaction, but we shouldn't dismiss
53:39that either.
53:41And if you're China, like this is like
53:44really important to you.
53:46Um and they're
53:48you know, there are some reports that
53:50like an EV machine had been smuggled
53:52into China.
53:53Um
53:55and I mean, what a feat of espionage cuz
53:57those things are like giant machines.
53:59They're
54:00they're [laughter] huge. Um I don't know
54:02if that's true. You know, there's some
54:03noise about it, but um
54:06you know, China, they're really really
54:08good. They're really really smart. They
54:10work brutally hard.
54:12And you know, they see this is super
54:13important for them as a country.
54:16Um
54:17but are they going to go from the year
54:202001
54:22to 2026
54:24or even 2000 and you know, 30?
54:27Are they going to it it cuz it really is
54:29it is
54:29>> It's a learning by doing.
54:30>> It's it's a learning by doing.
54:32And you kind of have to Yeah, if like
54:35like you can't you can't accelerate the
54:37doing. You can't you can't teleport into
54:39the future. You actually have to go
54:40through those learning cycles.
54:43So,
54:44is it significant? Yes. Did the market
54:47overreact? Probably. But like I think a
54:49lot of
54:51like I think it's it's very hard as an
54:53American
54:54to really understand what is happening
54:56in China and like have
54:59like total conviction and clarity, you
55:01know, like for
55:03for better or worse, like we are
55:05decoupling. And
55:07um
55:10just that is a process that is been set
55:13in motion.
55:14And at this point it almost feels like
55:17it's kind of self-reinforcing on each
55:19side.
55:20And you know, that's that's unfortunate.
55:23Um
55:25but
55:27we are where we are. And they're not
55:29they're not going to stop, neither are
55:30we.
55:30>> Any commentary on like every other
55:32company in America? Like I feel like
55:34right now it is 10 companies, couple
55:36private.
55:37>> Well, that that last month, I mean,
55:40everything but AI was vertical.
55:43And I do think open you know, open
55:46source getting closer to the frontier
55:49and companies like Fireworks making it
55:51really easy to customize a model such
55:54that you can get
55:56in some cases better than frontier for
55:58performance for meaningfully lower cost.
56:01That is a godsend for the software
56:03industry.
56:04And it's also a godsend for all these
56:06like, you know, there's there's a lot of
56:07AI natives and like all these AI
56:10natives, you know, it's like our friend
56:11Vashria, I think he said 2 years ago,
56:14I've never seen more companies go from
56:16like being founded
56:18to like $50 million a year in revenue
56:20and generating cash flow in like
56:22whatever it is, 9 months.
56:24And it's hard to know if any of them are
56:25durable because like back then, like
56:29it's like, hey, you know, these are a
56:30lot of people would dismiss them as chat
56:32GPT wrappers.
56:34Well, now with open source, you've
56:35actually you you've generated some data
56:38that's unique to your your use case,
56:40whatever your vertical you're going
56:42after has a wrapper is. Fireworks, they
56:44did come out with a really cool product
56:46called Nexus.
56:48And if you're using cloud code, open AI
56:50codex, grok build,
56:52it is literally three lines of code,
56:54like 20 words.
56:56And um Fireworks adjust your data
57:01kind of, you know, they can RL a model
57:03and then there's a router that sends the
57:06query and they've had amazing results.
57:10Um and this is kind of the solution
57:12for every AI native and that's why you
57:14saw,
57:15you know, Harvey
57:17uh before it was acquired, um Cursor
57:21leads so heavily into this. Harvey,
57:23Legora, all of them. Because if you can
57:27go from just using one, two, two three
57:30frontier models to use a
57:32>> Whatever's optimal.
57:33>> those frontier bottles for whatever it
57:35is, 30%
57:3660%
57:39of your token consumption and then use
57:41your own RL bottle, all of a sudden
57:43you're not a rapper. You're way more
57:45defensible.
57:46Um
57:47>> I was so interested by that cursor thing
57:49that came out. I think it was cursor
57:51where it's sort of like a AI speed
57:53running like what we've learned amongst
57:54humans, which is you could use the
57:55frontier model to plan and then farm out
57:58tasks to the dumber models.
58:00>> 100%
58:00>> And and it's 15 times more efficient or
58:02whatever the metric was.
58:03>> It it it may be that like
58:06if this is like super ironic,
58:09um
58:10but it may be that like lower margin
58:13open source tokens that are just
58:16a little bit behind the frontier and you
58:18know, we have a we have friends who
58:20believe that you know, frontier
58:21[clears throat]
58:22once a frontier model hits RSI
58:25it will actually have a dramatically
58:27lower cost
58:28>> to serve the local
58:30>> at at every
58:32at every level of intelligence by kind
58:34of distilling this and then there's no
58:36place for open source. I would say
58:38that's like a
58:40you know, a um
58:42Anthropic
58:43OpenAI Grok maximalist view. And you
58:48know, we should we shouldn't dismiss
58:49anything. I don't know or really
58:50important. Anything is possible. Like
58:53you know, we we we would have like be be
58:55very humble. I particularly want to be
58:57humble after the month I've had.
59:00But that doesn't seem that likely to be
59:04and
59:05>> Why?
59:06>> Well,
59:07um
59:08one because there are so many of these
59:10AI natives that have actually
59:14generated a decent amount of domain
59:16specific proprietary data.
59:18>> Yeah.
59:19>> And kind of before like
59:22open source had this moment to these
59:24inference clouds and these routers
59:26really developed like you kind of didn't
59:28have a choice. Like whatever the terms
59:30of service were, you accepted them. But
59:33if you can now kind of get off that
59:35treadmill,
59:36um that gives you a degree of
59:38independence,
59:39maybe durability, safety. But kind of
59:43going back to your point, it may be that
59:45these cheaper tokens
59:48just
59:50massively inflate the value of the most
59:52cutting-edge frontier tokens.
59:55Because if like today if you have I you
59:56know I'm going to make this up, you
59:58know, 120 IQ open-source models,
1:00:01um and they're really cheap to run,
1:00:05well, doesn't that make a 160 IQ model
1:00:08that can orchestrate them
1:00:11more valuable? And so just we talked
1:00:13last time about how
1:00:14I've been really surprised that you know
1:00:16so much of the economic returns have
1:00:17accrued to the frontier.
1:00:19Now that that is changing with what
1:00:21we're seeing with these kind of
1:00:22inference clouds. Um
1:00:25together, Modal, um
1:00:28uh Base 10 in a very cash-efficient way.
1:00:30What's shocking about those business
1:00:32models
1:00:34is they're growing almost as fast as the
1:00:36frontier labs in the early days,
1:00:39but burning very little cash.
1:00:42Like it's it's pretty extraordinary, you
1:00:44know, from like you know look to go back
1:00:46to silly SaaS metrics like the you know
1:00:49the rule of 40 perspective. Like
1:00:51these are crazy numbers.
1:00:53>> Do you think there's a lot of
1:00:54instruction in just like the
1:00:55distribution of pay inside of an
1:00:56organization? Like the CEO makes X times
1:00:59more than the median person at a company
1:01:01and maybe that's frontier tokens versus,
1:01:03you know, open-source tokens.
1:01:04>> Absolutely. Yeah.
1:01:05>> Something simple like we can
1:01:06>> Yeah, it may be that what we discussed
1:01:07last time where you know frontier tokens
1:01:10I think they may lose a like the pie is
1:01:13growing really really fast.
1:01:16They may continue to capture the
1:01:18overwhelming majority of economic value,
1:01:20but kind of not all of it the way they
1:01:21have been, and open source tokens might
1:01:24be the majority of token source tokens
1:01:25processed.
1:01:27It just again, going back, that's great
1:01:29for infrastructure demand because a
1:01:31token is a token and it
1:01:33takes the same amount of flops,
1:01:35watts,
1:01:37space, cooling to make.
1:01:40>> What's the worst thing that could happen
1:01:42in AI? Is it regulatory? Is it some sort
1:01:44of like
1:01:44>> I think regulatory has to be the biggest
1:01:46risk.
1:01:47Um
1:01:48I mean, it's the most obvious risk.
1:01:51And so that was kind of one reason I was
1:01:52excited to be here this week
1:01:54was to just like
1:01:56I
1:01:57I want to be scared, you know? I like I
1:01:59don't I don't want to feel like a
1:02:01lunatic, you know, watching these stocks
1:02:05relative to um you know, get more
1:02:08cheaper thinking the expected forward
1:02:10returns are going up.
1:02:12You know, while you know,
1:02:14it feels like the on the ground
1:02:15fundamentals have like pretty materially
1:02:18improved.
1:02:19Um
1:02:20in July relative to even June, but I
1:02:23still came come away thinking like
1:02:26you know, regulation, it just has to be
1:02:29the biggest risk. Like you just can't
1:02:30ignore New York
1:02:33making a data center moratorium.
1:02:36It just like we are we're living in this
1:02:38weird
1:02:40post-factual, post-logical political
1:02:43world.
1:02:44It
1:02:45you know,
1:02:46and I mean, I think the AI industry it
1:02:48has done a
1:02:49terrible job
1:02:52of PR, and I do think they're
1:02:54>> it at least realizes that now.
1:02:56>> Yeah.
1:02:56>> Maybe if not fixed it, it realizes it.
1:02:58>> Yeah, but like kind of the narrative in
1:03:00Washington, you know, that that that the
1:03:02political narrative, you know, I think
1:03:04amongst a lot of ordinary Americans is
1:03:05like data centers,
1:03:07they're going to raise your electricity
1:03:08prices, they're going to take all your
1:03:09water, and they're going to take your
1:03:11job.
1:03:11>> [laughter]
1:03:12>> And the reality is like given the deals
1:03:15that are being cut now, when a data
1:03:16center goes in, electricity prices
1:03:18actually generally go down for everyone
1:03:20around there because of behind the meter
1:03:22deals.
1:03:23This is that like data center pledge
1:03:25that kind of Trump asked people to to
1:03:27sign. Generally, the data center
1:03:29developer, you know, it used to be they
1:03:30just had to build a like, you know,
1:03:32whatever. They had to get the police
1:03:33department or the fire departments like,
1:03:36you know, new trucks and new cars and,
1:03:38you know, new body armor or whatever.
1:03:40Now, it's like, well, we're going to
1:03:42build you a hospital, a school, a new
1:03:44police station, and a fire station. And
1:03:46we're going to lower your power bills.
1:03:48How does that sound? And by the way, the
1:03:51jobs are ongoing because it turns out
1:03:53that you kind of need these plumbers,
1:03:55electricians, you know,
1:03:59HVAC contractors. And this is like data
1:04:02centers are like
1:04:03are in a lot of ways the best thing to
1:04:05happen
1:04:06for blue-collar wages in my lifetime.
1:04:10And yet, you have the Democrats who
1:04:11ostensibly
1:04:13represent the, you know, the blue, you
1:04:15know, these blue-collar workers
1:04:18taking their jobs away.
1:04:20Um
1:04:21And so, um
1:04:24it it also like
1:04:25it's it's just kind of wild how like,
1:04:28what is the phrase? Like a lie can go
1:04:30around the world
1:04:31>> Faster than the truth gets out of bed,
1:04:32yeah.
1:04:33>> Yeah, faster than truth gets out of bed.
1:04:35But an author made a mistake in a book
1:04:38and overestimated the amount of water
1:04:40usage in data centers by 10,000 X. Not a
1:04:43little bit. Like not one order of
1:04:45magnitude. Not two orders of magnitude.
1:04:47Not three, you know?
1:04:48Um
1:04:50and um
1:04:52she's admitted that mistake many times.
1:04:55I was completely wrong.
1:04:57It's like been super debugged.
1:04:59>> uh Popeye effect. Did you Did you hear
1:05:01that example?
1:05:02>> No.
1:05:02>> The the, you know, Popeye eats spinach.
1:05:05The reason was same deal in an academic
1:05:07in an academic book. They placed the
1:05:08decimal two things wrong. So, spinach
1:05:10does not have more iron than everything
1:05:12else. It was just this one source and
1:05:14then that propagated through people
1:05:16still say it has more iron.
1:05:17>> I literally had I thought it had more
1:05:19iron. [laughter] I mean that's wild.
1:05:21That's wild. I literally thought spinach
1:05:24had more iron.
1:05:24>> [laughter]
1:05:25>> That's amazing.
1:05:26>> Yeah, you learn something new every day.
1:05:27>> Same thing though.
1:05:28>> Uh yeah, it's the same thing and it's
1:05:30just
1:05:31So somebody just needs to tell the
1:05:32truth. Like like I I feel like the
1:05:35industry and I thought like jeez, maybe
1:05:37if nobody else is going to do it like
1:05:38I'll do it. Like there needs to be some
1:05:40sort of foundation. Maybe it's a pack
1:05:43that runs ads during the final four,
1:05:45during the NFL games, during college
1:05:47football games.
1:05:48>> Here's the virtues.
1:05:49>> World Series.
1:05:51Here's what a data center does. Your
1:05:52power a data center that signed this
1:05:55pledge in your community.
1:05:56>> Yeah.
1:05:57>> Your power prices are going to go down.
1:05:59They're almost certainly going to
1:06:02um
1:06:02you know
1:06:03like contribute to the community in a
1:06:05material way. You're going to see a
1:06:07massive influx of super high paying blue
1:06:11collar jobs
1:06:12that are going to persist and I think a
1:06:13lot of people thought that they were one
1:06:15time and they're just not. Like there's
1:06:17for sure a spike and then that moves to
1:06:18the next data center but there is an
1:06:20ongoing kind of
1:06:22you know, need for kind of RMA and then
1:06:25upgrades at these data centers and and
1:06:26technology is changing. So you're going
1:06:28to have more jobs, you're going to have
1:06:29cheaper power.
1:06:31You're going to have a wealthier
1:06:32community. Um there's going to be no
1:06:34impact on on water, no impact on the
1:06:38environment. You know, and it's easy to
1:06:39build the data center 10 miles out of
1:06:41town, you know.
1:06:42And so like that story needs to be told
1:06:45along with, you know, like there are you
1:06:47know, we we we we we heard a we we heard
1:06:49a story I think we talked about it last
1:06:50time about how AI is increasingly really
1:06:53saving lives, curing rare diseases.
1:06:55Like we um
1:06:57you know, I think I can't remember if it
1:06:58was I I think it was at ASCO this year.
1:07:01You know, the kind of vibe, you know,
1:07:03the the vibe was like hey, we've this is
1:07:06the most scientific breakthroughs we've
1:07:08ever seen
1:07:10at a single conference.
1:07:12And for sure some of that is due to AI.
1:07:14And so we need to like tell those
1:07:17stories. Like, you know, if you have a
1:07:18sick child, you know, a sick parent, uh
1:07:21a sick loved one, like
1:07:23AI meaningfully increases the odds
1:07:28of them recovering.
1:07:30Like if we just we we need it's
1:07:32everybody needs to tell this. And I
1:07:34think people out here
1:07:37it's all of this is so blindingly
1:07:40obvious to them
1:07:43that they they
1:07:44>> seems everyone else already knows.
1:07:45>> They they can't Yeah, they can't process
1:07:48that this is a
1:07:49true but wildly divergent view
1:07:53from most Americans.
1:07:57And so like I think the industry really
1:07:59needs to tell its story better.
1:08:01Cuz this is
1:08:04like
1:08:05New York, it just feels like it's the
1:08:06first of many and even in some of these
1:08:09deep red states, they're super
1:08:11pro-growth.
1:08:13They're just like, "Hey, you guys are
1:08:14not doing a good job telling your story.
1:08:17Then we can't we can't tell your story.
1:08:19If you tell your story though, we can
1:08:21retell it, but like you're the experts."
1:08:25Um you know, if you like
1:08:27like something I've if you do not speak
1:08:30your own truth, no one else will.
1:08:32>> [clears throat]
1:08:33>> Yeah. What have we missed with
1:08:34>> I think something that is missing from
1:08:36all of this conversation about compute
1:08:39is what is going to happen when you put
1:08:40these SRAM-based accelerators that are
1:08:43not constrained
1:08:44by HBM DRAM and are often made on older
1:08:47nodes that are not competing
1:08:50with like the latest and greatest GPUs.
1:08:53You can whether you there's when you
1:08:55disaggregate Fritz, there's people talk
1:08:58about prefill and decode, but decode has
1:09:00two parts, attention and feed forward
1:09:02network, and like the ultimate holy
1:09:04grail is if you could do pre-fill
1:09:07on one chip.
1:09:08Um it probably doesn't have HBM DRAM. Do
1:09:12the attention on a super high-powered
1:09:15chip with HBM DRAM, and then do the feed
1:09:18forward network on one of these SRAM
1:09:20chips.
1:09:21But like
1:09:23the ROI on adding these SRAM
1:09:27accelerators,
1:09:29uh
1:09:30to the existing install base of compute
1:09:31and and new compute.
1:09:33But like what we're seeing is like
1:09:36you do better. You just can't beat SRAM
1:09:40in particular for that feed forward
1:09:42network. And And you just almost you
1:09:44can't, no matter how much you try to get
1:09:46the ratio of compute to HBM DRAM
1:09:49to SRAM on the chip correct. Like the
1:09:52workloads are always changing, and
1:09:53there's different workloads.
1:09:55And like
1:09:56being able to disaggregate it to these
1:09:58three parts,
1:09:59uh
1:10:01like I think this is
1:10:03this is going to be really really
1:10:05positive for the ROI of AI.
1:10:07>> For some reason I just thought of a
1:10:08funny question, which I love their
1:10:09framing of Game of Thrones versus all
1:10:11these people. Can you imagine a player
1:10:13that is not currently on everyone's mind
1:10:15becoming relevant at like the major Game
1:10:17of Thrones scale? Like that could be
1:10:18like Micron all of a sudden, you know,
1:10:20it'd be like a sample answer to the
1:10:22question of someone that becomes as
1:10:24important as Anthropic, OpenAI,
1:10:26Microsoft, Amazon, you know, Nvidia,
1:10:29SpaceX.
1:10:30>> Yeah.
1:10:31So like a dark horse Game of Thrones
1:10:32player?
1:10:33>> that come to mind.
1:10:36Um
1:10:38like Li Bu is probably a dark horse.
1:10:41Um
1:10:44I do think um Lin at Fireworks, she is
1:10:49like a
1:10:50just an absolute killer.
1:10:54Um I think
1:10:56uh you know, our friend Scott Wool.
1:10:58>> Mhm.
1:10:58>> You know, cognition is kind of like uh
1:11:01>> Here, here to that one.
1:11:02>> Yes.
1:11:03Uh uh
1:11:04I think those are
1:11:08uh the most obvious names.
1:11:11>> What about SpaceX? What's it been like
1:11:12watching that be digested by public
1:11:15markets at least initially?
1:11:16>> Uh uh
1:11:17>> Do you think the market understands it
1:11:19as a company? The most important new
1:11:20company to be public?
1:11:23>> It doesn't really feel like it
1:11:26it does because it's kind of like such a
1:11:30it's such a
1:11:32like everything to me is
1:11:35the fundamentals have gotten better
1:11:36since it IPO'd. Like Rock 4.5, the
1:11:39Cursor acquisition.
1:11:40You know, Cursor
1:11:42uh has clearly accelerated meaningfully.
1:11:46And then they have showed that they
1:11:47could, you know, they've they've showed
1:11:49over the last 3 years they could bring
1:11:50out more compute faster than anyone
1:11:52at lower prices. And now we know that
1:11:54they could even adjusting for the spot
1:11:56first contract gap, like their big
1:11:58advantage was they came into the market
1:12:02you know, and just hit those spot highs.
1:12:05Uh uh
1:12:06And in a strange way, like one of the
1:12:07more bullish things for compute
1:12:09is like, you know, they put a vast
1:12:11amount of compute into the market
1:12:13overnight.
1:12:14And it wasn't even really a blip.
1:12:16It was like the market just
1:12:19utterly absorbed it, you know? Like just
1:12:22the free trade didn't sold out at all.
1:12:25Uh uh
1:12:26But you know, a you know, a Substack
1:12:28writer Wolfund Fund AI,
1:12:31they think that SpaceX is going to try
1:12:33and bring out 8 gigawatts of compute.
1:12:36I will never bet against Elon.
1:12:41But I mean
1:12:42that would be a truly incredible feat.
1:12:46And they are
1:12:48rates have gone up since they signed
1:12:49those last contracts, not down.
1:12:52And they're monetizing at something like
1:12:5450 billion a gig.
1:12:56And ConsenSys estimates for next year
1:12:58are 73 billion. So, forget Starlink V3.
1:13:02Forget Starlink direct to cell. Grok 4.5
1:13:06and Cursor, the sum of that probably
1:13:08hits a $10 billion ARR pretty quickly.
1:13:12For
1:13:13Forget all of that. Um
1:13:16you know, forget like the core base
1:13:18Starlink business.
1:13:20If they bring out anywhere near that,
1:13:23the ConsenSys estimate is 73 billion.
1:13:26And that's 8 gigs at 50 billion a gig.
1:13:28And obviously, that would not all be lit
1:13:30up at the beginning of '27.
1:13:33And it seems very implausible to me.
1:13:36Like, I almost don't believe the Funder
1:13:38report.
1:13:39Um
1:13:42but
1:13:44you it To this day, the only companies
1:13:46that have brought out more than 500
1:13:47megawatts
1:13:48of power
1:13:51in a year are the hyperscalers,
1:13:54Coreweave, Crusoe, and SpaceX.
1:13:57And SpaceX has kind of brought out the
1:13:58most the fastest at the lowest cost.
1:14:01And then
1:14:02people do actually really like their
1:14:04clusters.
1:14:05Um
1:14:08but again, it's kind of like the market
1:14:10is going to need to see that.
1:14:13>> That would not be the market's
1:14:14interpretation of SpaceX today.
1:14:16>> No, no.
1:14:17Um
1:14:18and it does feel like, you know, there's
1:14:19this There's There's a big New York
1:14:21hedge fund short case on it.
1:14:23And I think they think, you know, oh,
1:14:24the spot price for compute's going to go
1:14:26down 90% and, you know,
1:14:28you're going to bring out all this
1:14:30You're going to bring all this on all
1:14:31this compute. It's not going to
1:14:32generate, you know, nearly as much
1:14:33revenue as you think. Maybe, but also
1:14:35want to be really clear like
1:14:38Like, I have seen those I have seen
1:14:40Yolt's companies, you know, do really
1:14:41impressive things over the year. Pretty
1:14:44the Funder AI report of 8 gigawatts at
1:14:4618 months.
1:14:48I'm I'm just quoting that cuz it's
1:14:49public. It's available to everyone.
1:14:52Ooh, like that.
1:14:53>> yeah.
1:14:53>> That Yes.
1:14:56Um
1:14:57you know, I think one of Elon's phrases
1:14:59is we specialize in making the
1:15:01impossible late.
1:15:03>> [laughter]
1:15:04>> I've never heard that. That's great.
1:15:05>> Yeah. Um it you know, there's like kind
1:15:08of a lot of truth to that. Yeah, yeah.
1:15:10Um
1:15:12but I just think
1:15:14very little is built in
1:15:18from my perspective to that stock
1:15:22for the amount of compute
1:15:24that they might be able to bring on. And
1:15:27again, I don't think it's anywhere near
1:15:29eight.
1:15:30Um and it's going to be really hard and
1:15:32energizing these GPUs is really hard.
1:15:35But they've been good at it and it
1:15:37doesn't feel like that's in estimates or
1:15:39really in people's thinking.
1:15:40>> I'm thinking about that funny meme that
1:15:41says SpaceX, the data center company?
1:15:44>> [laughter]
1:15:45>> You said yes,
1:15:46absolutely.
1:15:47Um and then I would also just say like
1:15:51from
1:15:52Yeah, I did spend a lot of time at
1:15:53Starbase and
1:15:55um
1:15:56orbital compute feels more real every
1:15:59day.
1:16:00>> Pretty cool to see that Starship landing
1:16:01the other day.
1:16:02>> Pretty cool to see the Starship landing
1:16:04and then it's you know, it is funny.
1:16:05There's our friends at Benchmark. They
1:16:07funded Star Cloud and
1:16:09I don't know last time Star Cloud is an
1:16:10orbital compute company that like SpaceX
1:16:12is kind of partnering with.
1:16:14Um they're going to I think let them use
1:16:16the Starlink laser technology, which is
1:16:17really important for orbital compute.
1:16:19And like but I do think that's like kind
1:16:21of a good sanity check.
1:16:23Last time I checked, you know, the
1:16:24Benchmark guys were pretty smart.
1:16:27And they're not
1:16:29coming from the Elon ecosystem at all.
1:16:33And they chose to fund an orbital
1:16:35compute company
1:16:37at like you know, a decent valuation
1:16:41without the internal launch that SpaceX
1:16:43gets.
1:16:44And that's just to me that's a good
1:16:46like, "Hey, am am I
1:16:48>> crazy?"
1:16:49>> Am I crazy? And it's like, well, maybe
1:16:51I'm crazy and maybe Elon's crazy and
1:16:54maybe Benchmark is also crazy and maybe
1:16:57the SpaceX engineers are also crazy.
1:17:01But man, that just doesn't seem that
1:17:02probable to me.
1:17:05And I mean, we we should say should we
1:17:06say whose offices we're in?
1:17:08>> Yeah, we're sitting in the middle of the
1:17:09famous table.
1:17:10>> Yes, this is their famous table.
1:17:10[laughter] Yes, this is their famous
1:17:12table for their famous dinners. Um so
1:17:15thank you Benchmark. Thank you Benchmark
1:17:17for this episode. Yes, thanks Eric. Um
1:17:19and and she we should thank them all.
1:17:21>> Um Eric Eric coordinated for me so he
1:17:24gets a special shout out.
1:17:25>> Thank you, Eric.
1:17:25>> Thank you all of the partners.
1:17:26>> Thank you, Eric. Well, you know, just
1:17:28you know, we will see where all of these
1:17:29stocks are in a year.
1:17:32And the great thing is, you know, time
1:17:34will tell.
1:17:35You know, people are going to be right
1:17:37or wrong.
1:17:38You know, the future's probabilistic,
1:17:39but we are at like it's an exciting
1:17:41moment.
1:17:42>> Well, if we keep doing this on the the
1:17:44model release cycle, I'll see you in a
1:17:45couple weeks.
1:17:46>> [laughter]
1:17:47>> Yeah, it's crazy.
1:17:49>> As always a blast to do with you.
1:17:54>> You know how small advantages compound
1:17:55over time? That's true [music] in
1:17:56investing and just as true in how you
1:17:58run your company. Your spending system
1:18:00is your capital allocation strategy.
1:18:02Ramp makes it smarter by default. Better
1:18:04data, better decisions, better economics
1:18:06over time. See how [music] at
1:18:07ramp.com/invest.
1:18:09As your business grows, Vanta scales
1:18:10with you, automating compliance and
1:18:12giving you a single source of truth for
1:18:14security and risk. Learn more [music] at
1:18:16vanta.com/invest.
1:18:18Ridgeline is redefining asset management
1:18:19technology as a true partner, [music]
1:18:21not just a software vendor. They've
1:18:23helped firms 5x in scale, enabling
1:18:25faster growth, smarter operations, and a
1:18:27competitive edge. Visit
1:18:28ridgelineapps.com
1:18:30to see what they can unlock for your
1:18:31firm.
1:18:32>> [music]
1:18:32>> The best AI and software companies from
1:18:34OpenAI to Cursor to Perplexity use Work
1:18:36OS to become enterprise ready overnight,
1:18:38not in months.
1:18:39>> [music]
1:18:39>> Visit workos.com to skip the unglamorous
1:18:42infrastructure work and focus on your
1:18:43product.

About this transcript

This page contains the full transcript of Why the Markets Are Pricing AI Wrong | Gavin Baker by Invest Like The Best, generated from the public captions YouTube serves with the video. The transcript has 13,308 words across 2,334 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.