Milk Road Stocks

Silicon Data's Steve Ho argues bond markets reject AI productivity boom while GPU futures launch amid falling token prices.

46 min Listen to the original Episode page
Artificial IntelligenceGPU HardwareBond MarketsCME FuturesAI ValuationCompute Economics

Summary

Steve Ho, Head of Research at Silicon Data, argues that the bond market's skepticism—evidenced by 10-year yields near 5%—suggests investors do not believe AI will significantly accelerate GDP growth. He contends that for AI CapEx to be justified, the technology must generate new economic output rather than merely displacing labor. Ho details Silicon Data's new GPU rental indices, now licensed to the CME for futures contracts, which track soaring hardware costs. Conversely, he highlights a divergence where LLM token prices are collapsing due to efficiency gains and competition, creating potential margin risks for AI providers.

Episode sections

  1. 0:00 – 2:46

    Introduction and Guest Background

    Host LG Du introduces Steve Ho, Head of Research at Silicon Data, who holds a PhD in macroeconomics from the University of Michigan (2018) focusing on US federal debt maturity structures.

  2. 2:46 – 7:09

    Macroeconomic Analysis: Bond Yields and AI

    Ho argues that 10-year yields hovering around 5% indicate the bond market does not believe in the AI productivity story, as transformative AI would likely push yields higher via accelerated GDP growth.

  3. 7:09 – 11:24

    AI ROI and GDP Growth Requirements

    Ho asserts that AI investments require GDP growth of 3-4% by 2028 to be justified, emphasizing that AI must generate new aggregate output (elasticity) rather than just displacing labor (erosion).

  4. 11:24 – 20:16

    Anthropic IPO and Valuation

    Discussion of Anthropic's $2 trillion valuation target, with Ho citing Aswath Damodaran’s conservative $1.4 trillion estimate as a more realistic benchmark.

  5. 20:16 – 24:52

    Consumer Adoption and AI Agents

    Ho predicts mass consumer adoption will be driven by autonomous 'agents' that perform tasks in the background, requiring AI to become 'brainlessly simple' for viral growth.

  6. 24:52 – 28:04

    Silicon Data and GPU Rental Indices

    Ho explains Silicon Data’s tracking of real-world GPU rental prices to create transparent benchmarks, aiming to develop financial derivatives for AI compute.

  7. 28:04 – 29:20

    CME GPU Futures and ETFs

    Confirmation that GPU rental indices for H100 and B200 are licensed to the CME, with futures contracts pending CFTC approval around November 9, 2026, followed by ETFs.

  8. 29:20 – 32:02

    GPU Rental Price Trends by Chip Generation

    On-demand GPU rental prices have risen sharply since November 2025, with newer chips seeing parabolic increases and older chips like the A100 seeing increased demand for inference.

  9. 32:02 – 34:17

    Lack of Substitution Between Old and New Chips

    New AI chip launches do not cause immediate depreciation of older chips; instead, demand is additive as new supply is absorbed by different use cases.

  10. 34:17 – 37:24

    The 'Most Important Chart': Token Prices vs. Compute Costs

    Presentation of a chart showing rising GPU rental indices against falling LLM Token Index prices, highlighting a divergence between soaring compute costs and dropping AI output prices.

  11. 37:24 – 40:23

    Market Risks and AI Adoption Debate

    Ho notes a risk if token prices continue to fall without a surge in volume, as AI equity indices previously peaked when the token index peaked.

  12. 40:23 – 45:46

    Open vs. Proprietary Models and Future Outlook

    The decline in token prices is attributed to substitution toward cheaper open models; Ho suggests prices could rise if users return to high-quality proprietary models, while long-term GPU rental forward curves trend upward.

Assets discussed

AssetKindStanceContext
NVIDIANVDAequityneutralBuy more NVIDIA, buy more SpaceX, buy more Micron, or light it on fire by buying some crypto.
Meta PlatformsMETAequityneutralMuse is obviously growing in incredibly fast right now, uh, which is really good news for Meta.
AlphabetGOOGLequityneutraluntil I guess Google earnings last night or whatever.

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

Steve Ho holds a PhD in econometric and macroeconomics from the University of Michigan, obtained in 2018.

Unidentified speaker ·1:44

How to verify: Check Steve Ho’s LinkedIn profile or University of Michigan alumni records for his degree and graduation year.

In 2018, US 10-year bond yields were substantially lower than 2.5% or 2.8%.

Unidentified speaker ·4:26

How to verify: Check historical US Treasury yield data from the Federal Reserve Economic Data (FRED) database for 2018.

In 2021, the 10-year US Treasury yield hovered around 2.0% despite significant money printing and deficits.

Unidentified speaker ·4:50

How to verify: Check historical US Treasury yield data from FRED for the year 2021.

As of the recording (Oct 29, 2026), the 10-year US Treasury yield was approximately 5%.

Unidentified speaker ·6:25

How to verify: Check current US Treasury yield data from FRED or financial news sources for late October 2026.

Aswath Damodaran, a professor at NYU, valued Anthropic at $1.4 trillion.

Unidentified speaker ·18:09

How to verify: Search for Aswath Damodaran’s online threads or valuation models regarding Anthropic’s pre-IPO valuation.

Anthropic’s IPO valuation target was reported as $2 trillion.

Unidentified speaker ·11:33

How to verify: Check financial news reports and SEC filings (S-1) related to Anthropic’s IPO announcement in late 2026.

Silicon Data tracks NVIDIA GPU rental indices.

Unidentified speaker ·27:36

How to verify: Visit the Silicon Data website or check their published reports for GPU rental price indices.

Silicon Data also tracks AMD GPU rental indices.

Unidentified speaker ·27:50

How to verify: Visit the Silicon Data website or check their published reports for AMD GPU rental price indices.

The AI agent 'Poke' was acquired by Cognition.

Unidentified speaker ·16:26

How to verify: Check press releases or news articles regarding Cognition’s acquisition of the AI agent 'Poke'.

Steve Ho uses the AI model 'Opus 5.5' from Anthropic.

Unidentified speaker ·15:23

How to verify: Check Steve Ho’s social media posts or public statements regarding his preferred AI tools and model versions.

NVIDIA H100 and B200 GPU rental price indices have been licensed to the CME for futures contracts.

Unidentified speaker ·28:29

How to verify: Check CME Group press releases or filings for 'NVIDIA GPU Rental Index' or 'Silicon Data' partnership announcements.

The CFTC review period for these futures contracts was scheduled to conclude on November 9, 2026.

Unidentified speaker ·28:48

How to verify: Check CFTC public notices or CME regulatory filings for the specific approval date of these futures contracts.

GPU rental prices for A100, B200, B300, H100, and H200 have increased sharply since November 2025.

Unidentified speaker ·29:59

How to verify: Access Silicon Data's published GPU rental index data or third-party cloud pricing aggregators (e.g., Vast.ai, RunPod) for historical spot rates.

Older chips like the A100 are seeing increased demand for inference workloads, causing their rental prices to rise despite the launch of newer chips.

Unidentified speaker ·30:22

How to verify: Review Silicon Data's breakdown of GPU utilization by workload type (training vs. inference) and historical rental rate trends for A100s.

The LLM Token Index (price per million tokens) has declined due to price cuts by frontier labs and substitution toward open models.

Unidentified speaker ·36:34

How to verify: Compare published token pricing from major AI labs (OpenAI, Anthropic, Google) over time and analyze usage share data between proprietary and open-source models.

AI equity indices peaked in alignment with the peak of the token index, suggesting a correlation between token pricing power and hardware stock valuations.

Unidentified speaker ·38:29

How to verify: Plot the performance of AI-focused ETFs (e.g., SMH, AIJ) against the Silicon Data Token Index over the relevant period.

Anthropic is profitable and has become more profitable over time.

Unidentified speaker ·40:16

How to verify: Check Anthropic's public financial disclosures, press releases, or reputable financial news reports regarding their profitability status.

Forward curves for long-term GPU rentals are trending upward, indicating market expectations of future price increases.

Unidentified speaker ·42:18

How to verify: Review Silicon Data's forward curve data or long-term contract pricing reports from major cloud providers (AWS, Azure, GCP) for AI accelerators.

Do your own research

  • Verify the current 10-year US Treasury yield to assess if it remains near 5% as claimed, indicating market skepticism on AI-driven growth.
  • Check the CFTC website for the status of the CME GPU futures regulatory review, specifically looking for approval dates around November 9, 2026.
  • Review Silicon Data’s published GPU rental indices to confirm the price trends for NVIDIA H100, B200, and A100 chips.
  • Analyze LLM token pricing data from major providers (e.g., OpenAI, Anthropic, Google) to verify the decline in price per million tokens.
  • Examine Anthropic’s S-1 filing or IPO prospectus for actual valuation metrics and compare them to Aswath Damodaran’s $1.4 trillion estimate.
  • Assess enterprise AI adoption rates and CapEx spending reports to determine if GDP growth forecasts of 3-4% by 2028 are supported by corporate investment.
  • Monitor the correlation between AI equity indices and the LLM token index to evaluate the stated risk of revenue model squeezing.
Full transcriptexpand
0:00
Unidentified speaker: If AI, after all these trillions of dollars of investments, and we are still growing just about the same rate as we were before, that means that the whole thing has been a bust.
0:09
Unidentified speaker: What's up everybody?
0:10
Unidentified speaker: It's LG Du Set here and welcome to Milk Road Stocks, a daily market show that can't decide what to do with my last thousand dollars.
0:15
Unidentified speaker: Buy more NVIDIA, buy more SpaceX, buy more Micron, or light it on fire by buying some crypto.
0:20
Unidentified speaker: Today is November 2nd, 2026, recording on October 29th.
0:24
Unidentified speaker: The AI trade has taken a beating recently, or a slowdown, depending on how you look at it, and investors are starting to wonder whether we have built too much of it, or and if the ROI will ever actually be positive on all this money that's been spent.
0:37
Unidentified speaker: But what if the actual demand for AI compute is telling us a completely different story?
0:42
Unidentified speaker: Today we're joined by Steve Ho, head of research at Silicon Data, a company tracking the real prices people are paying for AI compute.
0:48
Unidentified speaker: We're gonna find out whether the AI boom is still accelerating, which parts of the market are getting squeezed, and what investors should be watching next.
0:55
Unidentified speaker: Steve, welcome to Milk Road, man.
0:57
Unidentified speaker: Thank you so much for having me, Algie.
0:59
Unidentified speaker: Okay, so there's there's a lot, there's a lot to cover today.
1:02
Unidentified speaker: I want to get your thoughts just on the general build out because I feel like you're closer to the data than really than anybody we have on the show, including our own analysts.
1:08
Unidentified speaker: Um, but first, first, I think it's important just take a couple couple minutes to understand who you are, because despite being the head of research at uh Silicon Data, you're you also have a PhD in macro, which actually uh is gonna make your views on that probably more advanced than even our own macro experts.
1:24
Unidentified speaker: So tell us a little bit about yourself first.
1:26
Unidentified speaker: Yeah, so uh thank you very much for that quick uh intro.
1:29
Unidentified speaker: And uh indeed uh I am currently the head of research at Silicon Data, which tracks AI data.
1:34
Unidentified speaker: But uh in my prior life, uh uh I did get a PhD in uh econometric uh and uh macroeconomics in uh 2018 from the University of Michigan.
1:44
Unidentified speaker: And I studied a topic which at the time I think was a little bit ahead of its time.
1:48
Unidentified speaker: I was trying to understand the optimal maturity structure of uh US federal debt in an inflationary environment.
1:55
Unidentified speaker: Uh and uh That's insane.
1:58
Unidentified speaker: Right before COVID, like a year or two before COVID, that's right.
2:01
Unidentified speaker: That's right in 2018.
2:02
Unidentified speaker: Yeah, it was a little bit difficult to get people to care back then about inflation or uh whether or not the US government would run into trouble of uh issuing long-term debt.
2:11
Unidentified speaker: Um that's so crazy.
2:13
Unidentified speaker: And how much how much how much of your views on that uh changed?
2:17
Unidentified speaker: Like if you had to do your PhD in 2022, I'm assuming it would have been quite.
2:22
Unidentified speaker: People are welcome to check on my PhD.
2:24
Unidentified speaker: I I would say uh the current lingo is that you have I think it's aged pretty well.
2:28
Unidentified speaker: Uh at the time, I I the analysis was that uh I was trying to understand to what extent you know sort of bonds, you know, are going to the supply of government bonds, the issuance government bonds is going to have an impact on bond yields or interest rates and the environment in which it will matter more or less.
2:46
Unidentified speaker: And I found this curious pattern where the correlation of returns between stocks and bonds, which actually has had this apocalypse, like regime shifts, before 2000, it was basically positively correlated.
2:59
Unidentified speaker: Stocks and bonds will move up and together, uh down down together.
3:03
Unidentified speaker: After 2000, there was a sharp drop in correlation where bonds became negatively correlated with stocks in returns and became a hedging asset.
3:12
Unidentified speaker: And this is also during the period when you have had popular strategies like uh you know risk parity and so on and so forth, that um people have come to view bonds as a hedging asset for downside uh recession risks.
3:25
Unidentified speaker: And post-COVID, uh um, you know, whatever it was that basically shocked people's uh view of inflation saliency, uh made stop-bond correlation positive again.
3:36
Unidentified speaker: And this is an environment in which uh, you know, basically uh to incur to entice investors to want to hold bonds, you've got to pay them higher bond yields.
3:45
Unidentified speaker: In other words, the demand, this the demand curve for bonds basically steepens.
3:50
Unidentified speaker: And uh I made a prediction back in 2018 that uh if inflation ever returned, stock bond correlation will likely flip very positive.
3:59
Unidentified speaker: During this period, uh the the sensitivity of bond yields to additional supply of long-term uh long maturity debt will become very high, uh, and uh you would discourage the government from uh issuing very uh long-term uh you know high quantities of long-term debt and uh persuade them to essentially shift towards shorter maturity and funding itself uh with uh shorter maturity um debt instruments.
4:23
Unidentified speaker: So and what were what were bond yields then in 2018?
4:26
Unidentified speaker: Like two and a half, 2.8 or something?
4:28
Unidentified speaker: Yeah, something substantially lower.
4:29
Unidentified speaker: Yeah, that's right.
4:30
Unidentified speaker: Yeah.
4:31
Unidentified speaker: Are you surprised are you surprised that post-COVID it took this long for bond yields to go up then?
4:34
Unidentified speaker: Is that is that kind of what you're saying?
4:36
Unidentified speaker: That you thought maybe when COVID happened and they started printing money, you're like, well, these are the bond yields will skyrocket soon, but it's it's taking a bit of a lot of time.
4:42
Unidentified speaker: You certainly felt a little surreal in 2021 after all the you know new money printing and uh you know additional deficits uh and and checks being mailed out, that bond yields were not responding very much, like 10-year yield was hovering around 2.0 for a long time, right?
5:00
Unidentified speaker: And almost as though the bond market didn't want to believe that the that prior 10 years had already become a bygone era.
5:06
Unidentified speaker: Uh, and when that was happening, it made you question whether or not your fundamental model uh was actually incorrect, that you was you were missing some structural forces that uh would and and it it turns out you just took a little while for uh you know uh people to actually uh uh I think realize that what was happening.
5:24
Unidentified speaker: Oh man, what a wild uh eight years, really, when you think about it in finance.
5:27
Unidentified speaker: Yeah, it was opening up while ride, yeah.
5:29
Unidentified speaker: Uh I do you want to talk about AI, but what let while we're just still talking about macro, Steve.
5:33
Unidentified speaker: Maybe one thing I can ask you is in this current environment, what is something on the macro side that people don't pay enough attention to, right?
5:40
Unidentified speaker: Because obviously, more so than any time I think ever, everybody who follows the stock market and finance is hyper-aware of like rates and like the Russell, you know what I mean?
5:51
Unidentified speaker: Like everybody knows all these things now, but what's something that people don't quite grasp yet?
5:56
Unidentified speaker: That they just that is not common knowledge enough of its effect on the market, or that uh has something that's coming in the future that people haven't thought about yet.
6:04
Unidentified speaker: I think this these days people have all become very, very sensitive.
6:08
Unidentified speaker: You know, uh, you know, a couple years before, I think maybe most people weren't paying super close attention.
6:12
Unidentified speaker: I think what is interesting today is that people the way people are looking at rates is um that the bond market has actually never, even for once, in my opinion, believed in the AI story, right?
6:25
Unidentified speaker: The fact that 10 years only at 5%, just about today, right?
6:30
Unidentified speaker: Tells you that uh if you were to uh believe that the next 10 years was to say have uh about 2.5% inflation and uh 2.5% or 2%, let's say real HGDB growth, which is the long-term historical average, that gets you to about 5%, right?
6:47
Unidentified speaker: In other words, you're not too far off if you believed that the world was just gonna continue the way it was, right?
6:54
Unidentified speaker: Uh going forward the next 10 years.
6:56
Unidentified speaker: And that is clearly not what the AI story is.
6:59
Unidentified speaker: If AI, after all these trillions of dollars of investments, and we are still growing just about the same rate as we were before, that means that the whole thing has been a bust, right?
7:09
Unidentified speaker: That at some point we're gonna come to the ROI story later, because at this point, the amount of investment has gone in.
7:15
Unidentified speaker: Uh, the story, the the the exercise of determining whether there will be an ROI to AI investment has almost become a bit easier because the sum of investment is so large that you've got to make a difference, you know, to essentially the rates of GDP, right?
7:30
Unidentified speaker: If inflation, if if if GDP does not get faster growing at 3-4% in 2028, you know that uh basically this does not did not actually create the type of productivity improvements that people had been anticipating.
7:44
Unidentified speaker: So if that were the case, and if you do some simple math of say, you know, 2% put 2% additional GDP and in this environment, at least for the first few years, it seems implausible that we will have like massive deflation while we have got all this build out.
8:00
Unidentified speaker: Let's just say inflation is another, say, three, four, uh, three percent, let's say, right?
8:04
Unidentified speaker: It's not unreasonable for the next 10 years.
8:07
Unidentified speaker: You're already looking at uh substantially meaningfully higher 10-year yield, right?
8:12
Unidentified speaker: You know, seven, six, seven, eight percent, you know, uh uh uh six, seven percent.
8:16
Unidentified speaker: So so the fact that we are where we are is actually showing an interesting kind of a uh um, I think I don't know who's right, right?
8:24
Unidentified speaker: But clearly the bond market is just like, yeah, whatever AI, you know, you can if you guys want me to pay pay me a higher yield, you know, we're happy to take it, but uh there's no sense in which the bond market seems to actually believe uh that uh you know AI was actually uh a real you know thing, uh beyond the investment itself.
8:42
Unidentified speaker: Yeah.
8:42
Unidentified speaker: We're gonna pause for a second.
8:44
Unidentified speaker: The AI trade is changing.
8:46
Unidentified speaker: The bottleneck stocks are either running out of steam or getting ready for another leg higher while the application layer is finally starting to show some real profits.
8:53
Unidentified speaker: Our analysts were early to the last wave and now they are buying for the next one.
8:58
Unidentified speaker: Catch your next call inside Milk Road Pro for just a dollar at the link below.
9:01
Unidentified speaker: Wouldn't that be Steve, though, because the because AI also eats parts of the market, right?
9:05
Unidentified speaker: Because that's been the story this year, right?
9:07
Unidentified speaker: I think when a lot, a lot of people listening have probably started to tune into the AI trade has been this year, is that you know, we started off the year where every week there's a new announcement from Anthropic saying, like, hey, accountants don't matter anymore, this job doesn't matter anymore, you have all of software falling off a cliff.
9:22
Unidentified speaker: So isn't I I I what you're saying makes a lot of sense that it's like for this CapEx, there needs to be an enormous growth in GDP considering how much has been spent, but doesn't that first require like the funeral of stuff that was previously propping up GDP in as in software?
9:39
Unidentified speaker: I don't know.
9:40
Unidentified speaker: I mean, obvious, obviously, part of it will be that um the RRI um of AI is going to come through the erosion or the displacement displacement of existing modes of production, right?
9:53
Unidentified speaker: Certain occupations will be displaced and uh more efficient ways of doing the same thing will come.
9:59
Unidentified speaker: But uh the point is also that uh AI will not be able to pay for itself if it entirely did it through erosion of what it already existed by displacing, you know, like you can't really fire exit your fire, you can't really fire people your way out of this whole thing, right?
10:16
Unidentified speaker: If uh you know uh the tech firms got rid of half of their employees and uh just you know replaced them with you know AI instead.
10:24
Unidentified speaker: Uh and if that's all that happened, uh, I don't know if you actually be able to find enough you know uh um return to actually justify the investments because you've gotta make the pie bigger, right?
10:35
Unidentified speaker: You can't really just you know do it entirely by uh replacing parts of the pie.
10:39
Unidentified speaker: Yeah, it can't just be supplementing parts of the pie.
10:41
Unidentified speaker: It has the pie has to grow.
10:43
Unidentified speaker: Yeah, yeah.
10:44
Unidentified speaker: You can't spend all that money just to replace software companies with AI agents, it has to at some point become bigger, right?
10:50
Unidentified speaker: That's right.
10:51
Unidentified speaker: At some point, that the aggregate productivity has got to come.
10:54
Unidentified speaker: At some point, you know, there's this economic jargon of uh elasticity, that at some point the aggregate elasticity of demand has got to come such that uh you're not just substituting one in production input with another, you're actually collectively, we are all getting more done, get make generating more uh um output, right?
11:13
Unidentified speaker: Generating more things of value uh such that the overall pie is bigger.
11:18
Unidentified speaker: Right.
11:18
Unidentified speaker: Yeah, okay, yeah.
11:19
Unidentified speaker: That's that's that's I think that's a good way of looking at it.
11:21
Unidentified speaker: What what let's let's jump into the ROI part though, right?
11:24
Unidentified speaker: Because that is that's the main story that we cover on our show.
11:26
Unidentified speaker: Uh, whenever there are new numbers, right?
11:28
Unidentified speaker: Anthropic's filing is coming very soon, or they filed and and they're that that IPO is coming very soon.
11:33
Unidentified speaker: And that's very uh contentious, obviously, as to whether uh people are gonna want to buy it at $2 trillion, especially and on the retail side after it told after it told everybody it was gonna kill them, anyways.
11:43
Unidentified speaker: Um, so obviously there's there's a bit of a funny sentiment story.
11:46
Unidentified speaker: But what I mean from your point of view, you're deep in the data, man.
11:49
Unidentified speaker: Like, is is this where where is the ROI right now?
11:53
Unidentified speaker: Like, is this can it grow?
11:54
Unidentified speaker: Can it can it ever achieve these levels that it needs to to justify all this spend?
11:59
Unidentified speaker: Well, I mean, therein lies the debate, right?
12:02
Unidentified speaker: You know, and at this point, I think, you know, if I knew the answer, you know, I'd be a very rich man.
12:09
Unidentified speaker: You know, I'd be betting with certainty one way or another.
12:12
Unidentified speaker: Uh, compared to two, three years ago when this whole thing was still viewed as like I remember back in middle of 2023 when I first became aware of the whole AI trade, the picks and shovel trade, and uh people called it the baby bubble.
12:24
Unidentified speaker: That got everybody very excited, at least got me very excited because you know, babies inevitably grow up and become bigger.
12:30
Unidentified speaker: And uh, you know, the the uncertainty you know pointed strictly to the to the upside because you can only be positively surprised when you have not very little, very uh that you know, sort of the the the stakes were smaller.
12:42
Unidentified speaker: Today, I think clearly the uncertainty has become more asymmetric, right?
12:47
Unidentified speaker: Sorry, become more symmetric uh in that uh there is a big uh bet right on both sides, whether you are um someone who's bullish or someone who's bearish, on how AI demand for AI, demand for intelligence for you know um will actually uh grow.
13:05
Unidentified speaker: Um is it going to grow rapidly and uh you know uh and diffuse across society uh such that uh uh it's going to outstrip supply as much as you know supply we're building?
13:18
Unidentified speaker: Uh or will it uh you know not find enough use case and not diffuse enough uh that you end up with uh you know too much supply?
13:26
Unidentified speaker: Or could you have a case where supply demand eventually does get there, but it may take a while, right?
13:33
Unidentified speaker: Uh that uh you know you sort of you get your most obvious use cases like coding and and other things that are easy to diffuse.
13:42
Unidentified speaker: But then the other types of adoption that especially the ones that involve a lot of human uh uh agency uh takes longer, at which point you can have a potential air pocket in which uh uh during which period supply shows up, but demand doesn't get there in time.
13:59
Unidentified speaker: And then eventually it gets there because you requires people to actually be proactive, be curious.
14:05
Unidentified speaker: And as we know that some people, if not many people, are maybe not quite as proactive and curious and you know, full of agency as we would like them to be.
14:14
Unidentified speaker: So I think therein lies the lot of the big uncertainty.
14:17
Unidentified speaker: And uh every day I see evidence, you know, pointing to, you know, like some days I see it pointing one way, some days I see it pointing very much to the other way.
14:24
Unidentified speaker: Uh, and and we can talk more about the you know anthropic S3 and uh what my recent experience with anthropic and cloud model has been like, and it has made me sort of a bit more bullish, but um uh but anyway, I mean, yeah, so that's what's your what's your what's your AI of choice actually?
14:42
Unidentified speaker: Let's let's get it, let's get personal then.
14:43
Unidentified speaker: Like what do you what do you what do you like the most?
14:45
Unidentified speaker: Because I I can tell you that for me, like Rockbot has been pretty awesome.
14:50
Unidentified speaker: Uh, I know a lot of people are kind of flip-flopping between them.
14:53
Unidentified speaker: Muse is obviously growing in incredibly fast right now, uh, which is really good news for Meta.
14:58
Unidentified speaker: But what do you what do you personally prefer?
15:00
Unidentified speaker: So I am a uh quant, I code and I generate, you know, sort of uh quantitative analysis.
15:07
Unidentified speaker: So I use Claude a lot.
15:08
Unidentified speaker: I think Claude is incredibly well built uh for that purpose.
15:12
Unidentified speaker: Even when you know during the period uh earlier this year, you became quite annoying in the way it was nagging at you in in the way the pros was coming out.
15:20
Unidentified speaker: Like I don't know, with the post-training was a little bit weird.
15:23
Unidentified speaker: Uh more recently with Opus 5.5, it's actually absolutely improved that I find myself using you know Cloud almost exclusively for work, right?
15:31
Unidentified speaker: Whether it's coding, whether it's generating, writing, and and and whatnot.
15:36
Unidentified speaker: Uh and then on the on the other side of it, uh, I use grok a lot because I'm quite active on Twitter.
15:41
Unidentified speaker: Uh so you know grok has that data integration with X, and uh I think Grok generally is just quite good at reasoning.
15:48
Unidentified speaker: Uh, it doesn't really think BS you, it doesn't really uh give you like long-winded, you know, yes, if this but no, like that sort of thing, right?
15:56
Unidentified speaker: It gives you gives it gives it to you straight, a bit like you know how you would expect a product from Elon would do.
16:02
Unidentified speaker: Uh and that's really about like the the combination of AI tools I use, and I use GrogBot a lot.
16:08
Unidentified speaker: And I also have been using uh Instinct, which is another uh uh AI agent, you know, um that's operating off of uh uh iMessage, right?
16:18
Unidentified speaker: Um so it's similar to like I think before instinct, there was also Poke, which was which was acquired by Cognition.
16:26
Unidentified speaker: Generally speaking, I I like Muse as well.
16:28
Unidentified speaker: I think these three three things I've I've switched between them.
16:31
Unidentified speaker: Over time, I think there is an idea that you want to stick with one so that you learn about you and you you sort of uh that that long contacts and memory improves it as a product.
16:40
Unidentified speaker: Uh, but I do think uh you know agents are a very good like product market fit in terms of uh uh a new step towards finding that you know viral product.
16:52
Unidentified speaker: Absolutely.
16:53
Unidentified speaker: Given your usage of it, or and and and and we will get to the data in a minute.
16:57
Unidentified speaker: So for people listening who want some some technical analysis and stuff like that, it's all coming in a minute.
17:01
Unidentified speaker: Let's just talk about this this um Claude IPO though, or anthropic IPO really quick.
17:07
Unidentified speaker: Do you you use it a lot?
17:08
Unidentified speaker: We use it a lot of Milk Road, a lot of people use it.
17:11
Unidentified speaker: Is it is two trillion dollars justified though for uh not only how good it is, which isn't shouldn't really be the marker, but how what the what the revenue is actually showing?
17:20
Unidentified speaker: Well, uh, you know, I mean 2 trillion, I I don't I could I couldn't tell you because um I'm not a uh uh valuation person, but uh I do know that uh the guru of uh valuation uh was his name professor at uh NYU.
17:36
Unidentified speaker: Um what I'm I'm I'm blanking on his name.
17:39
Unidentified speaker: Uh um uh uh Aswath, uh the uh Demotoran.
17:44
Unidentified speaker: Um do you know him?
17:45
Unidentified speaker: No, no, tell us more.
17:47
Unidentified speaker: So uh if you Google Guru evaluation, there's this professor uh Aswath uh um the Motoran and uh pronunciation of his name, but um he usually is pretty you know lever-headed and uh you know takes a more uh you know, I think rigorous and conservative approach for equity evaluation, where he search actually does the careful exercise of equity evaluation.
18:09
Unidentified speaker: I think I saw a thread from him where he was placing his own valuation of Anthropic at 1.4 trillion.
18:16
Unidentified speaker: Okay.
18:16
Unidentified speaker: Uh which you know, sort of uh, you know, I mean, I don't know how real the 2 trillion you know price tag really was.
18:24
Unidentified speaker: I don't know if they were necessarily targeting 2 trillion um for their IPO price.
18:29
Unidentified speaker: You will see, you know.
18:30
Unidentified speaker: Uh I suspect the 2 trillion is more just like internally, that's the upper range of what some people might have discussed.
18:37
Unidentified speaker: I you know, I suspect um, you know, I think a while ago, it wasn't all that long ago, I heard the target was more like 1.5.
18:43
Unidentified speaker: And uh if Aspot was coming 1.4, you are not far off, right?
18:46
Unidentified speaker: And that's really um if you think about that S1 filing uh ahead of IPO, which made quite a lot of people you know uh excited, you know, uh, and the market has clearly been responding positively over the last few days, um, until I guess Google earnings last night or whatever.
19:05
Unidentified speaker: Um that uh this was actually a very much of a pleasant surprise, right?
19:10
Unidentified speaker: I think people knew that uh Anthropic was doing well.
19:13
Unidentified speaker: They've clearly found a great use case and they committed almost entirely to it, which is to hydro home you know, uh coding and make sure that uh they can you know achieve uh RSI or this recursive self-improvement through coding, and that was clearly the right bet.
19:31
Unidentified speaker: Uh and it gave them, I think, the right uh and they have apparently the great uh great internal culture and research taste, whatever it is, you maintained a persistent edge and uh and difference with their product, such that they've managed to acquire a lot of enterprises users.
19:48
Unidentified speaker: And uh, you know, onboarding a new vendor, as you know, with large enterprises always takes a long time.
19:54
Unidentified speaker: But once you're onboarded, it tends to be fairly sticky.
19:57
Unidentified speaker: It is you know high quality revenue compared to uh the much flightier consumers that also don't probably have as much uh high value usage, right?
20:07
Unidentified speaker: Like the average consumer is likely still using AI as an essentially a glorified search engine or at most some sort of uh you know, do you think that'll ever change though?
20:16
Unidentified speaker: Is that will that ever change, right?
20:18
Unidentified speaker: Because we already've already discussed how even you and I use a cocktail of these these these applications.
20:23
Unidentified speaker: Yeah.
20:23
Unidentified speaker: Do you think ever do where and when and if will people and and we're talking consumer, right?
20:30
Unidentified speaker: Will we get to the point where it's like, well, okay, everybody's decided they're using one agent and they're just you know, and they're using it properly, not just to search for you know, uh Christmas presents or or recipes for tonight, you know what I mean?
20:42
Unidentified speaker: Like what what do you think that that'll happen?
20:44
Unidentified speaker: I think so.
20:44
Unidentified speaker: I mean, the biggest difference maker was agents, right?
20:48
Unidentified speaker: The rise of uh Grogbot and the rise of uh um Amuse and so on.
20:54
Unidentified speaker: I mean, I've been saying for the longest time that for all the people who were expecting this rapid viral adoption and productivity gains, like the reality was that AI just wasn't didn't get good enough until early this year when agents you know became much more, you know, sort of uh uh uh you know, and with cloud and coding agents and other other form of agents, and it made the adoption barrier a lot lower.
21:19
Unidentified speaker: Like imagine people needed to be persuaded, like to be taught to use an iPhone.
21:23
Unidentified speaker: Like iPhones never needed four deployed engineers to tell you how to use an iPhone.
21:27
Unidentified speaker: You pick it up and you use it.
21:29
Unidentified speaker: It doesn't even come with a user manual.
21:30
Unidentified speaker: You don't need to read any manual, you just pick it up and you use it.
21:34
Unidentified speaker: Uh, and and if you if Apple had to depend on people going by the horse into Apple Store, be trained in like a seminar, how to, you know, sort of like you know how they have these bars where they have like older people come in with the iPad and say, How do you turn it on?
21:47
Unidentified speaker: How do you flip a button and so on?
21:48
Unidentified speaker: If that's what they required, the product would have never gained popularity, it would have never gone viral, and you would have never been able to ship the tens of millions of units per quarter in the way they do.
21:59
Unidentified speaker: Uh so when it comes to consumers, you've got to make the product almost like sort of uh brainlessly simple, right?
22:06
Unidentified speaker: You know, you don't need to turn on your brain to sort of intuitively know how to use it.
22:09
Unidentified speaker: And I think that is what agents have become, right?
22:12
Unidentified speaker: With a chatbot type of agent that is actually able to use a computer, use tools on the background on your behalf, check for errors, check for all sorts of things that you care about and solve problems for you.
22:25
Unidentified speaker: Uh I think that is the way in which you actually do get the type of uh uh meaningful adoption of AI that goes beyond, much beyond, right?
22:34
Unidentified speaker: You know, uh a glorified search engine.
22:37
Unidentified speaker: Because people do have complicated, complex problems in life they want to solve, right?
22:42
Unidentified speaker: They just didn't want to necessarily figure it out how to solve it by sort of lining up the process themselves, right?
22:49
Unidentified speaker: If you just could simply ask that agent, say, hey, I've got this problem, maybe I can't even phrase it very well myself yet.
22:55
Unidentified speaker: Can you understand what I'm trying to say?
22:57
Unidentified speaker: And can you find me a solution and can you just maybe take care of it for me?
23:00
Unidentified speaker: And along the way, I think that is actually gonna involve a lot of AI usage uh that is happening on the background that you may or may not be fully aware of.
23:09
Unidentified speaker: And that is actually how you actually go beyond that one-dimensional AI usage.
23:15
Unidentified speaker: Guys, I'm just gonna pause there to tell you about one of my biggest problems in the market, which is seeing what some of the big names in the world actually hold in their portfolios.
23:21
Unidentified speaker: I'm talking about like Gavin Baker, how much SpaceX does he hold, how much NVIDIA does he hold, Kathy Wood, Peter Teal, Ken Griffin.
23:29
Unidentified speaker: We always hear about these big names, but nobody actually shows you individually what stuff they hold, which is why we went and built a website that does exactly that.
23:37
Unidentified speaker: So we have relaunched milkroad.com to basically aggregate all of these insider portfolios we're calling them, basically puts it all into one spot and lets you set alerts to tell you when they make moves.
23:50
Unidentified speaker: I'm talking like Warren Buffett, freaking Donald Trump is on there, Ray Dalio, all these big names you hear about.
23:55
Unidentified speaker: You can see their full holdings updated in real time on milkroad.com.
23:59
Unidentified speaker: So make sure you check that out and sign up today.
24:01
Unidentified speaker: It's gonna be that simplicity, man.
24:03
Unidentified speaker: Like you're saying, right?
24:03
Unidentified speaker: Even even something like the internet, we all had computers and sure, we had Windows and OS and stuff.
24:09
Unidentified speaker: But then, like you're saying, you know, until iPhones came along and literally just onboarded you.
24:13
Unidentified speaker: It was so simple.
24:14
Unidentified speaker: That's when you know usage really took off.
24:16
Unidentified speaker: And and that's what I'm seeing personally.
24:18
Unidentified speaker: When I when I downloaded news, I was like, okay, I could see my parents using this or people that aren't as motivated as you and I to learn this stuff.
24:26
Unidentified speaker: I could see one of these being the one that they they end up sticking with uh as their their assistant or their agent or whatever it is.
24:32
Unidentified speaker: And and I could definitely see them end up connecting all their other stuff to it to simplify their lifestyles, right?
24:37
Unidentified speaker: So um that kind of seems to be where it's going.
24:40
Unidentified speaker: Um let's let's flip the switch.
24:41
Unidentified speaker: Let's go over to uh what you guys do at Silicon Data, because I think that this is what really kind of what we wanted to talk to you about and also understand what you guys um track.
24:52
Unidentified speaker: Can you explain to us really quick, Steve, like what what the company does?
24:55
Unidentified speaker: Uh so we are a data company, uh uh and uh Silicon Data tracks AI data.
25:02
Unidentified speaker: Uh the founder is CEO, uh Carmen Lee, uh, and I both used to work at Bloomberg.
25:09
Unidentified speaker: I don't know if you are familiar with Bloomberg, but basically it is a financial data you know uh platform that's a software company that's very popular on Wall Street, you know, pretty much everyone uses it for uh traditional equity fundamentals data, bonds, and whatnot, right?
25:21
Unidentified speaker: I I don't know why I just explained to you what Bloomberg is, but some of your audience may actually you know benefit from that.
25:26
Unidentified speaker: Uh anyway.
25:27
Unidentified speaker: It's true.
25:28
Unidentified speaker: You'd be surprised.
25:29
Unidentified speaker: We we we have an ETF analyst, uh James Seafert from from Bloomberg.
25:33
Unidentified speaker: Come on our show come on our uh with some of our shows pretty often.
25:35
Unidentified speaker: Yeah, yeah, yeah.
25:36
Unidentified speaker: You know James, yeah.
25:37
Unidentified speaker: Yeah, of course you know James, yeah.
25:38
Unidentified speaker: But uh he comes on our shows pretty often.
25:40
Unidentified speaker: But yeah, a lot of people don't and also just to explain the Bloomberg terminal costs what, like 30 grand a month or something like that.
25:45
Unidentified speaker: Like it is it is like the top-of-line industry terminal that people use on the street, really.
25:51
Unidentified speaker: So so we are trying to uh bring data to the otherwise like large, fast growing but opaque uh physical AI economy, right?
26:01
Unidentified speaker: There's a lot of build-out, and there's a lot of uh you know of different layers of the uh AI economy, and uh, you know, people are renting GPUs, people are producing tokens, and people are buying models and so on and so forth.
26:13
Unidentified speaker: How does it all work and what what prices should be in charge?
26:17
Unidentified speaker: And the Carmen had this vision back in 2024 that this is going to be a very large and capital-intensive build out, and it was going to benefit from uh a lot of data transparency and price transparency.
26:32
Unidentified speaker: Uh, so she had this vision of collecting a lot of prices and a lot of data, uh, starting in 2024 and building uh publicly trusted, you know, uh transparent benchmarks, um, on top of which you can actually uh also have financial derivatives like futures contracts and uh options, the same way you have futures contracts for uh oil and uh crude and crops, right?
26:56
Unidentified speaker: And that is sort of involved with every single industrial build out that uh you've required such financial instruments, the type that uh you know America you know uh is very much uh you know uh good at doing, uniquely good at doing, right?
27:12
Unidentified speaker: New York is good at finance and uh you know Silica Valley is good at tech, right?
27:16
Unidentified speaker: You need both to be able to finance uh this activity in a capital efficient way.
27:22
Unidentified speaker: And I think our vision very much has been proven out.
27:25
Unidentified speaker: So we track uh GPU rental indices, uh, and what I'm showing you here are the NVIDIA uh uh GP rental indices.
27:36
Unidentified speaker: So we do them for uh you know the major uh NVIDIA GP rental indices, but we also by the way you have AMD uh GPU rental indices, those aren't as you know widely available, but as they become more widely available, we'll be tracking those as well.
27:50
Unidentified speaker: So, so you know, so this is funny you're bringing this up because um one of our co-owners, our co-owner Kyle, uh, who's also an analyst on our show, he's been saying for a while that like compute is basically the new gold.
28:00
Unidentified speaker: So it's funny that you guys are basically tracking it like a commodity, right?
28:04
Unidentified speaker: And that you guys are kind of putting forth that data.
28:06
Unidentified speaker: Are there are they you guys are tracking it like an index?
28:09
Unidentified speaker: Are there are no ETFs or indexes that we can actually buy though, right?
28:12
Unidentified speaker: Like that's not at the moment.
28:13
Unidentified speaker: Uh so we have uh filed, well, we have licensed our indices H100, which is the uh NVIDIA Hopper uh um yeah chip uh uh rental price index, uh, as well as the B200, Blackbow 200.
28:29
Unidentified speaker: Those two we have licensed to the CME and we'll be launching uh futures contracts uh to be traded on the CME, subject to uh regulatory approval.
28:38
Unidentified speaker: So the CFTC is currently in a uh uh review uh um period, and uh there's a public comments period that will be that has just concluded.
28:48
Unidentified speaker: And uh November 9th uh is supposed to be the period at which the review will conclude and we'll find out um and hopefully uh uh the futures would you know trade and uh at that point other financial instruments um uh that would use the futures on the line, such as ETFs, will also uh be gradually made available to the investment public.
29:07
Unidentified speaker: Dude, that's exactly what people want right now.
29:10
Unidentified speaker: It's like an ETF that they could just plow into.
29:13
Unidentified speaker: Uh, let's talk about what we're looking at right now, because I think I think naturally you guys you're way more than just the index, you're tracking like the actual data.
29:20
Unidentified speaker: What can you tell us about like GPU rental prices?
29:24
Unidentified speaker: Uh so well, the simplest thing is that GPU rental prices are going up, right?
29:29
Unidentified speaker: And uh this is kind of a so so the GPU rental uh rates is this is what I'm showing you are on demand, you know, these are short-term rentals.
29:38
Unidentified speaker: You can think of it as essentially, right?
29:39
Unidentified speaker: You know, you you rent the GPU uh hour by hour, and the rental rate could in theory change on the dime and you'd be paying as as it fluctuates, right?
29:48
Unidentified speaker: That's the on-demand uh index.
29:50
Unidentified speaker: And as you can see, the rental rate uh is actually been increasing, you know, more very sharply through the beginning the entire year, right?
29:59
Unidentified speaker: So this is showing you since November of 2025 through today.
30:04
Unidentified speaker: And I'm showing you here five chips.
30:06
Unidentified speaker: Uh the A100, which is MPR 100, that was a chip that was made available in uh mid of 2020, right?
30:14
Unidentified speaker: And uh is still going strong.
30:15
Unidentified speaker: And uh, if anything, it's actually inflecting somewhat higher.
30:18
Unidentified speaker: Uh and I love tracking this this line.
30:22
Unidentified speaker: This is the old chip.
30:23
Unidentified speaker: The the when when this one is going up, that tells you the inference demand is really strong because uh this is almost exclusively the use case for this very older uh chips, right?
30:33
Unidentified speaker: For the A100.
30:35
Unidentified speaker: And then it's gone up quite a bit this year.
30:39
Unidentified speaker: It's gone up quite a bit this year.
30:40
Unidentified speaker: Yeah, and then you soften a little bit over the last you know uh month or two, and then you started inflecting higher a little bit if you look at that sort of last little bit, right?
30:48
Unidentified speaker: And then what's really remarkable are the newer chips, the B200, the orange line, and the uh uh the B300, the uh uh the yellow line, those are have just gone absolutely parabolic, uh especially over the last you know month and change, right?
31:04
Unidentified speaker: Uh and then you've also got the A1 H100, which has also hit an all-time high, that's the blue line.
31:08
Unidentified speaker: And finally, finally, you've got the H200, which is at the very end of H100 with a different memory, uh uh and has also gone to almost close to all-time high.
31:18
Unidentified speaker: Now, these are on-demand you know, prices.
31:20
Unidentified speaker: In other words, people who cannot find long-term reserve rental contracts are essentially forced to buy on demand, and uh, you know, that's you know that demand is pushing up the prices uh because of limited supply.
31:32
Unidentified speaker: But I think the most recent pattern, this was really interesting here, is that the the newer chips are seeing the strongest amount of uh demand, likely because the newer generation of chips are most efficient, right, at producing tokens for per kilowatt of power.
31:49
Unidentified speaker: Um, you know, they produce just way more tokens, and you're able to, I think, on the other hand side of it, obviously, uh make a generate more revenue with it.
31:58
Unidentified speaker: But then, but then despite that, there's still demand for AE100s, right?
32:02
Unidentified speaker: Like that's that's the thing, right?
32:03
Unidentified speaker: Is that there's still there's still so much demand take the above.
32:06
Unidentified speaker: Exactly.
32:07
Unidentified speaker: And this is what is remarkable about this current market.
32:09
Unidentified speaker: I think earlier you were asking me about the various things I found surprising.
32:12
Unidentified speaker: One thing is just like people keep asking me whether or not we have seen a pattern where a new chip launches, the older chip then right away depreciates or becomes less valuable.
32:23
Unidentified speaker: And the thing is, people have this casual intuition almost like the iPhone, right?
32:27
Unidentified speaker: iPhone 17 comes out, iPhone 16 right away sees a big drop, or maybe even in anticipation of launch, you see a drop in value in the resale market.
32:35
Unidentified speaker: It doesn't seem to have worked out that way.
32:37
Unidentified speaker: You can see over the last year, right, as newer chips like the P200, B300 became more widely scaled out.
32:45
Unidentified speaker: But it also like just the moment when Nvidia launches a chip and when it becomes widely scaled out to be productionized, right?
32:53
Unidentified speaker: When it is actually installed onto data centers to start generating and working, there's a bit of a lag, right?
32:59
Unidentified speaker: And uh most of that, you know, uh uh scale scaling out happened this year and throughout this year.
33:05
Unidentified speaker: Uh the older chips are still going up in rental prices.
33:08
Unidentified speaker: In other words, the additional supply is being absorbed, but uh it is in doing so in a creative way.
33:15
Unidentified speaker: It is not doing it in a substitutive way, right?
33:17
Unidentified speaker: You're not substituting demand, at least not in a very obvious kind of like, you know, one thing comes and the other thing has to go down, right?
33:24
Unidentified speaker: Maybe the other thing would have gone on more, right, if you didn't have supply of uh the newer type of chips.
33:30
Unidentified speaker: Right, right.
33:31
Unidentified speaker: That makes sense.
33:32
Unidentified speaker: So how do we so how do you how do you how should people interpret this, right?
33:35
Unidentified speaker: Because we also spent the first half, like despite you know, a lot of things are bullish.
33:39
Unidentified speaker: We all use AI, we think retail's gonna have a huge adoption moment at some point soon.
33:43
Unidentified speaker: That's good.
33:44
Unidentified speaker: The biggest companies in the world that are launching, maybe you're gonna launch uh are gonna IPO, maybe slightly below, right?
33:50
Unidentified speaker: Like anthropic, like you're saying with that that um master evaluation or whatever, whatever he was, slightly below.
33:56
Unidentified speaker: But also the revenue, revenue generation is still the biggest story.
34:00
Unidentified speaker: So clearly money's being spent, there's an incredible premium being put on any new compute that's available, right?
34:08
Unidentified speaker: But there are much bigger risks than this.
34:11
Unidentified speaker: So, how do how do we approach this?
34:13
Unidentified speaker: Uh, like how do investors even uh in in interpret this data?
34:17
Unidentified speaker: So uh I want to show you this chart, right?
34:19
Unidentified speaker: It has been uh referred to by some as the most important chart in all of the you know uh uh the market at the moment, right?
34:27
Unidentified speaker: Uh you know, of all charts.
34:28
Unidentified speaker: And this is a chart where we juxtapose um uh three lines, those are the GPU rental indices for three chips we've already shown you so far, right?
34:37
Unidentified speaker: The H100, the B200, V300, those are going up fine.
34:41
Unidentified speaker: But the other one is what we call the LLM token index, right?
34:45
Unidentified speaker: Then this index tells you on a usage-weighted basis the uh price of a million tokens spent, right?
34:53
Unidentified speaker: Where are people spending the tokens and uh what is uh you know the average price that's being charged?
35:00
Unidentified speaker: Uh so the analogy I like to give to people is almost a bit like imagine these are not AI because AIs are new things, hard to sort of have a you know intuition for imagine these were cars and uh we were at the dawn of uh the invention of cars, right?
35:13
Unidentified speaker: The first year, you know, somebody invented the Mercedes and uh you've got 10 Mercedes sold, right?
35:19
Unidentified speaker: $50,000 each, you know, 30% gross margin or 35% gross margin.
35:23
Unidentified speaker: Come next year, right?
35:25
Unidentified speaker: Let's say 10,000 cars are sold, and thousand of them are Mercedes, and the rest of 9,000, let's say, are Toyota Corollas, right?
35:33
Unidentified speaker: Which are $10,000 a pop and uh say 20% margin.
35:38
Unidentified speaker: So you basically have a 10% or 10 times increase in the revenue of the Mercedes.
35:44
Unidentified speaker: So if you are a seller of Mercedes, you're pretty happy, right?
35:47
Unidentified speaker: You are seeing this crazy revenue growth.
35:50
Unidentified speaker: At the same time, if you look at the overall market of the cars that are sold, the average car or the median car is actually a much cheaper car now because you've got this proliferation of the cheaper car, which is which which was inevitably how it was always going to be, right?
36:05
Unidentified speaker: In order to make a new invention, new technology proliferate through society and diffuse the society, you've got to make it cheaper, you've got to make it easier to use, and you've got to sort of take advantage of the fact that demand curve slopes downward.
36:16
Unidentified speaker: So the fact that we have seen a uh proliferation of cheap uh uh flash, fast AI models from both the proprietary labs' own versions of flash cheap models as well as um open models from China and so on.
36:34
Unidentified speaker: Uh, the the the upshot is that uh AI has actually gotten cheaper, right?
36:39
Unidentified speaker: Tokens have gotten uh cheaper, and uh the Frontier labs themselves have been cutting prices too this year, uh, which explains uh you know a big chunk of the down.
36:49
Unidentified speaker: So if you look at this white line, we mentioned how in the middle of the year that this thing was sort of plateauing a little bit, that it could be on the on the way down.
36:57
Unidentified speaker: So now for the market, it's trying to figure out how the overall market moves, right?
37:03
Unidentified speaker: So does if if price comes down, does quantity go up way more, right?
37:08
Unidentified speaker: Way more than offset setting, you know.
37:10
Unidentified speaker: So uh if you say price comes down 20% or 30%, you know, on average, but then you know quantity goes up a hundred times, it doesn't matter, right?
37:20
Unidentified speaker: Everybody wins.
37:21
Unidentified speaker: You know, there is enough uh to go around.
37:24
Unidentified speaker: Uh so therein lies the big debate.
37:27
Unidentified speaker: Uh so clearly, right now, uh the fact that uh there is uh such sharp surge in uh compute prices, rental rates means that there is actually a lot of demand for AI and for a lot of demand for generating AI, which on this other side of it, someone is consuming it.
37:45
Unidentified speaker: Uh, the question and the debate lies in how diffused is that you know consumption of AI, how much of it is just simply consumed by AI companies and AI adjacent companies that are developing AI, because you know the labs themselves use a lot of tokens themselves, right?
38:00
Unidentified speaker: Or how much is it being consumed by the average consumer and how much is it going to be consumed by the average non-AI enterprises that are adopting AI and uh you know seeing uh gains from it.
38:11
Unidentified speaker: So does this concern you?
38:14
Unidentified speaker: I mean, I don't know if it concerns me.
38:17
Unidentified speaker: I mean, I I I certainly is obviously has been a risk factor in that uh if you look at any typical AI equity index consisting of the sort of typical picks and shovel stock, right?
38:29
Unidentified speaker: AI semiconducts and hardware stocks, uh has aligned up well with the peak of this token index in that the market more or less put in at least a local top, right?
38:42
Unidentified speaker: Uh since that time, we have had quite a bit of uh, I think, tumbling in some of the most popular names and uh you know, have different ways of different new bottlenecks emerge, but generally that's been the trend.
38:52
Unidentified speaker: Clearly, this is a uh aggregate risk, this is a risk factor that everyone is concerned about.
38:58
Unidentified speaker: And you know, together with this question about RRI, you know, the RRI is just the flip side of the same thing, right?
39:03
Unidentified speaker: If there is an RRI, then everyone's gonna rush to adopt, right?
39:07
Unidentified speaker: And if everyone rushes to adopt and that's gonna generate revenue for the AI, you know, sort of suppliers, that's obviously just going to mean you're gonna have uh strong demand.
39:17
Unidentified speaker: So uh so I guess I mean it was concerning to me in the middle of the year when this thing was uh you know putting in kind of a top, then people were not paying very much attention to it.
39:27
Unidentified speaker: When we put out this chart and mentioned this being a potential risk factor, and many people pushed back on this index as being not representative of the entire user base, which it is not, right?
39:37
Unidentified speaker: It is only capturing uh a part of the market that uh uh on which people can you know switch between models, right?
39:44
Unidentified speaker: The public-facing inference platforms.
39:46
Unidentified speaker: Yeah.
39:46
Unidentified speaker: But nevertheless, um, you know, it's come, I think at this point being accepted that the fact that people have become more differentiating and become uh better at making the substitution and quality cost trade-off between AI models has been you know uh impactful, right?
40:02
Unidentified speaker: Uh and uh people have come to question uh the pricing power and sustaining competitive mode of the Frontier Labs.
40:10
Unidentified speaker: So uh, but nevertheless, right, you know, Anthropic just came out with their S1 and uh they seem to be doing extremely well.
40:16
Unidentified speaker: They're profitable, they've become more profitable over time as well.
40:20
Unidentified speaker: So time to time will tell.
40:23
Unidentified speaker: What is this chart?
40:24
Unidentified speaker: Last question for you about this.
40:26
Unidentified speaker: What does this look like in a year?
40:28
Unidentified speaker: Um that's uh that's a multi-trillion dollar question.
40:32
Unidentified speaker: Uh that is literally what do you think, right?
40:35
Unidentified speaker: Like, does does token does the token price stay about here?
40:38
Unidentified speaker: Does it get way cheaper?
40:40
Unidentified speaker: Does it come together?
40:42
Unidentified speaker: If you decomposed the token index into the frontier and the open models, they've actually shown this interesting pattern where uh they've actually individually both put in kind of a bottom.
40:55
Unidentified speaker: Um, and um isn't that to the extent that the aggregate token index continues to go down or at least staying flat-ish uh over the last few months, it's not been because of uh you know continued price cutting, you know, or or within the frontier labs or proprietary labs.
41:12
Unidentified speaker: Rather, it's been because of continued substitution, substitution away, at least within the sample we will observe, away from proprietary to what's open.
41:20
Unidentified speaker: But if this pattern continues, you see you can see the open models index has also been flat and going up slightly.
41:27
Unidentified speaker: So if uh we see some kind of a reversion, let's say usage mix goes back towards the frontier labs because let's say enthropic models are super good user experience.
41:37
Unidentified speaker: Everyone loves Opus 5.5 or whatever in OpenAI comes up with next, you could see an inflection of the token index.
41:44
Unidentified speaker: And um, you know, that with it, you can also see a continued surge of uh uh compute, you know, rental prices, of course, uh, which is seems to be the expectation that the market is currently pricing in this because by the way, we also have uh didn't get a chance to show you uh forward curves, right?
42:01
Unidentified speaker: We can actually use the long-term rental rates of how much it costs to rent a GPU for a long number on a long period of time, let's say you know, a couple of years, and to infer what the market is currently uh implying about uh the rental cost in the future, right?
42:18
Unidentified speaker: And what's interesting is that if you look at this blue line, yeah, compared to a year ago, which is the yellow line, the yellow line uh it was much lower on the $2 and uh in exclusive, entirely downward sloping.
42:31
Unidentified speaker: In other words, you are given a pretty significant discount for long tenancy and locking in, you know, a long-term contract.
42:38
Unidentified speaker: But today, uh that longer end of the blue curve, albeit a thinner sample, seems to be pointing in the direction of travel is like higher, right?
42:47
Unidentified speaker: In other words, providers of computers are unwilling to be locked into super long contracts, but there's a certain expectation that uh rental uh prices will go up, compute prices will go up even more, right?
42:59
Unidentified speaker: So take that as as you may, but uh but that's certainly an interesting uh data point again to suggest that uh at least the market at the moment is expecting um compute prices to keep going up.
43:11
Unidentified speaker: So you think that they'll try and like time the top, though, that it's like right now they don't want long-term contracts, and then at some point they'll be like, okay, yeah, yeah, we'll take long-term contracts as and then and then block people into these prices as prices come down.
43:23
Unidentified speaker: You know what I mean?
43:24
Unidentified speaker: Like, is that what they're gonna do?
43:25
Unidentified speaker: I mean, that's almost a bit like uh, you know how you always hear stories about how like uh some airlines, right, they will hedge Jeff Fuel and then Jeff Fuel like will plunge, and then the ones that don't hedge, like Jeff Fuel, of course, go through the sky because some war happened in the Middle East.
43:39
Unidentified speaker: So um, I I don't know.
43:40
Unidentified speaker: I mean, that but I I'm just telling you what the data shows you, and uh I think it's for people to draw their own conclusions.
43:46
Unidentified speaker: Yeah, I just think that yeah, that's interesting.
43:48
Unidentified speaker: It it gives you, I think the data, this kind of data gives you a little bit more insight into almost like the day-to-day viewpoint of the people buying and selling the stuff, right?
43:58
Unidentified speaker: Or renting, renting, and renting it, renting and renting out this stuff.
44:01
Unidentified speaker: Excellent.
44:02
Unidentified speaker: That's right.
44:03
Unidentified speaker: Uh Steve, this has been awesome, man.
44:05
Unidentified speaker: Honestly, this has been it's been great to dive into the data.
44:07
Unidentified speaker: I wish we had more time.
44:07
Unidentified speaker: Um, but we usually cap our shows here.
44:09
Unidentified speaker: Um, thank you for coming on, man.
44:10
Unidentified speaker: I know it's been fantastic.
44:11
Unidentified speaker: Thank you very much for having me.
44:12
Unidentified speaker: Great to hear from the PhD in macro who did his thesis right before COVID.
44:17
Unidentified speaker: I can't believe that.
44:20
Unidentified speaker: Great time to to to to come into this market, uh, you know, as somebody who who analyzes it and and great to see what you guys are working on, man.
44:28
Unidentified speaker: I think that getting those um eventually, if we can we can trade indexes and ETFs, that's gonna be just awesome.
44:33
Unidentified speaker: And I think people are gonna love that, um, especially especially with a couple years ahead.
44:36
Unidentified speaker: Uh people who want to learn more about our data, you know, can come on to uh our website, silicondata.com, and also we share our research and insights on our uh social media handle so you can find us on you know LinkedIn on X, X on X, which is also Silicon Data.
44:51
Unidentified speaker: That's one word.
44:52
Unidentified speaker: Uh people can also you know come to me, but I don't always talk about you know data.
44:56
Unidentified speaker: Some people talk about some other some other nonsense.
44:58
Unidentified speaker: Wait, what do you talk about on X then?
44:59
Unidentified speaker: What do you what if you're not talking about data, what are you talking about?
45:02
Unidentified speaker: Uh macro and uh you know uh other things in the news.
45:07
Unidentified speaker: Uh so uh yeah, whatever is the latest.
45:10
Unidentified speaker: Wonderful.
45:11
Unidentified speaker: Well, we'll check you out for that, man.
45:12
Unidentified speaker: Thank you, Steve.
45:13
Unidentified speaker: Thank you very much for having me.
45:14
Unidentified speaker: Thanks for listening to Milk Road.
45:16
Unidentified speaker: If you enjoyed the show, make sure you like and subscribe.
45:18
Unidentified speaker: And if you're struggling to find winners in the market, that's exactly what Milk Road Pro is built for.
45:23
Unidentified speaker: Our analysts have called some of the biggest winners early, and Pro lets you see what they're buying next, every trade they make, and the research behind every position.
45:30
Unidentified speaker: Check out Milk Road Pro at the link below.
45:33
Unidentified speaker: Everything you hear on Milk Road is for informational purposes only.
45:36
Unidentified speaker: These are our personal opinions, not financial advice.
45:39
Unidentified speaker: And we may own some of the investments we talk about.
45:41
Unidentified speaker: Always do your own research and make the decisions that are right for you.
45:44
Unidentified speaker: See you next time.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3 and interpreted by Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.

More from Milk Road Stocks

Milk Road Stocks

Nvidia Is Down 30%. How Much Was Jensen Huang Worth?

Episode discussion: Nvidia Stock Drops 30% After Jensen Huang's Departure Amid Political Pressure

Speakers’ claims: John Gillen and Vincent discuss Nvidia's 30% stock decline and $1 trillion market cap loss following Jensen Huang's departure as CEO after 33 years. Vincent attributes the exit to political pressure from both Republican and Democratic officials regarding Huang's ties to Alice Huang, framing the leadership change as collateral damage in pre-midterm politics. The speakers analyze the operational risks of losing a founder-led CEO, comparing Nvidia's structure to Tesla's dependency on Elon Musk, particularly regarding innovation speed and pricing power. While acknowledging the leadership vacuum, Vincent cites analyst Tom Lee's argument that Nvidia's deep engineering talent and asset-light model support a 'buy the dip' thesis. The episode also covers geopolitical risks, including the RACE Act's impact on Chinese chip innovation and the potential for a rival AI stack.

NvidiaJensen HuangAI Infrastructure
34 min1 assets8 claims to verify
Milk Road Stocks

One Earnings Report Just Flipped the Entire AI Stock Market

Episode discussion: CoreWeave earnings flip AI market sentiment from survival to growth

Speakers’ claims: Speaker 0 reports that CoreWeave's Q3 results, featuring a 3% revenue beat and significant margin improvements, shifted market narrative from cash burn concerns to growth potential. The speaker highlights CoreWeave's $132.5 billion backlog and doubled contract pricing, noting that switched-on power, not demand, is the primary bottleneck. Bearish concerns regarding $46.9 billion in debt and $7.5 billion quarterly cash burn are acknowledged but offset by management's 2028 profitability target and customer prepayments. The earnings triggered rallies in AI infrastructure stocks like Nebius, IREN, and Oracle, with the speaker projecting strong results for Nebius. The episode concludes with anticipation for Anthropic's IPO and a disclaimer that the content is not financial advice.

AI InfrastructureCloud ComputingSemiconductors
38 min2 assets14 claims to verify
All insights