Steve Hou argues AI economics hinge on token-GPU spreads and Jevons Paradox, dismissing bubble fears as GPU demand remains robust.
Summary
Steve Hou of Silicon Data explains that AI business economics are driven by the spread between token selling prices and GPU rental costs. He argues that falling token prices reflect a shift to cheaper 'flash' models rather than collapsing demand, invoking Jevons Paradox to suggest volume growth offsets margin compression. Hou dismisses the 'bubble' narrative, noting that on-demand GPU rental rates are rising and hardware is fully utilized, unlike the dot-com era. He highlights that while spending is currently concentrated in the top 1% of companies, the investment thesis depends on broader adoption by non-AI firms. Finally, he notes that value accrual may shift from hardware to proprietary software layers as open-weight models gain traction.
Key takeaways
- 1 Token price indices are usage-weighted; a drop often reflects a shift to cheaper models rather than a collapse in total demand.
Steve Hou · 7:43
- 2 VC-backed startups are reinvesting proceeds from early AI investments into new ventures that purchase compute, creating a circular financing dynamic.
Steve Hou · 32:03
Episode sections
- 0:00 – 2:17
AI Cycle and GPU Economics
Steve Hou explains that Silicon Data tracks GPU rental prices and token usage as fundamental indicators for the AI economy. He emphasizes that the key metric is the spread between token selling prices and GPU rental costs, analogous to CPI/PPI spreads in macroeconomics.
- 2:17 – 7:18
GPU Rental Prices and Model Mix
Hou discusses how rising on-demand GPU rental prices impact investors in neo-clouds and labs. He introduces the concept of model mix, comparing frontier models to luxury cars and cheaper 'flash' models to economy cars, noting that volume growth in cheaper models can offset lower margins.
- 7:18 – 10:36
Jevons Paradox and Token Prices
Hou addresses falling token prices despite rising GPU costs, attributing it to a shift toward cheaper flash models. He invokes Jevons Paradox, suggesting that lower prices drive such a massive increase in consumption that total revenue may still grow, though margin compression is a concern.
- 10:36 – 16:04
Open Weight Models and Enterprise Adoption
Hou argues that while open-weight model usage is growing in developer communities, enterprise adoption remains unproven. He suggests that major corporations still prefer proprietary models from labs like Anthropic and OpenAI for direct, intentional use cases.
- 16:04 – 21:05
Market Valuations and Positioning
Hou notes that AI stock valuations have corrected significantly, pricing in fears of margin compression and slowing growth acceleration. He believes excess leverage has been flushed out and that recent filings show improving fundamentals for some labs.
- 21:05 – 25:20
Value Accrual in the AI Chain
Hou suggests value accrual is shifting. While open models pose a threat, proprietary labs like Anthropic show strong retention and profitability. Hardware suppliers face competition from customer-built ASICs and AI-assisted code migration, potentially shifting rents back to software layers.
- 25:20 – 30:48
Nvidia and GPU Depreciation
Hou challenges the view that rising GPU rental rates are purely bullish for Nvidia by highlighting competition and the potential for easier hardware switching via AI coding tools. He argues that GPU depreciation is slower than expected, with older chips retaining significant value.
- 30:48 – 32:03
Investor Worries and Demand Quality
Hou identifies demand gaps and the quality of demand as key worries. He notes that much current demand is internal to the AI industry (labs using their own tokens), raising questions about whether external enterprise demand will grow fast enough to justify supply build-outs.
- 32:03 – 33:41
AI Spending Concentration and Proliferation
Steve Hou discusses whether AI spending is a circular financing loop among VC-backed startups or if it will proliferate to non-AI companies. He notes that current spending is concentrated in the top 1% of companies, but questions whether broader adoption will drive meaningful productivity gains and top-line growth.
- 33:41 – 35:05
The Bubble Narrative and Compute Demand
Hou dismisses the 'bubble' label as an oversimplification, arguing that unlike the dot-com era's dark fiber, current GPU demand is robust with on-demand rental rates rising vertically. He suggests the real risk is not excess hardware, but potential algorithmic breakthroughs that could drastically reduce compute needs.
- 35:05 – 36:06
Jevons Paradox in AI Efficiency
Hou explains that while algorithmic improvements (e.g., mixture of experts) increase memory efficiency, the Jevons paradox applies: as efficiency rises, total memory usage increases because users engage in longer, more complex interactions. This dynamic prevents a surplus of unused compute despite efficiency gains.
- 36:06 – 36:53
Contact and Resources
Phil Rosen asks where to find Steve Hou's work. Hou directs listeners to silicondata.com, noting an increased frequency of blog posts, and provides his Twitter handles for both Silicon Data and his personal account.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| Gold | commodity | neutral | If you own a gold or silver ETF, the vast majority of its metals are likely. stored in foreign vaults under foreign jurisdictions. |
| Silver | commodity | neutral | If you own a gold or silver ETF, the vast majority of its metals are likely. stored in foreign vaults under foreign jurisdictions. |
| NVIDIANVDA | equity | neutral | And then there is also uh the issue of competition. Uh, you know, uh, with NVIDIA being the dominant uh player in this space of CUDA uh uh for the longest time, this market was always NVIDIA's to dominate and theirs to also potentially, you know, uh uh uh see competition, right? |
| Anthropic | private company | neutral | You know, Anthropic is one of the biggest users of their own tokens for developing their models. |
| OpenAI | private company | neutral | And these VCs, you know, probably made a lot of their money from investing in anthropic OpenAI and early on. |
| DeepSeek | private company | neutral | Mixture of experts and all these sort of uh you know uh uh uh uh multi-head attention and you know types of algorithmic improvements from me from China. DeepSeek and elsewhere, even within the US labs, um, that uh is actually making these efficiency gains and we are a lot more memory efficient. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
Silicon Data tracks GPU rental pricing and token usage/pricing as fundamental indicators for AI businesses.
Unidentified speaker (UNKNOWN) ·0:26
How to verify: Review Silicon Data blog posts or research reports on AI fundamentals and GPU economics. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
On-demand GPU rental indices capture floating rates, which are substantially higher than reserved compute prices from previous years.
Steve Hou ·2:46
How to verify: Compare on-demand vs. reserved pricing data from cloud providers (AWS, Azure, GCP) or third-party GPU rental aggregators. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Anthropic's S1 filing revealed good margins, low burn rate, and cash flow positivity for a quarter.
Steve Hou ·19:29
How to verify: Review Anthropic's S-1 filing with the SEC EDGAR database. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Primary sourceAnthropic has a 90% retention rate.
Steve Hou ·22:07
How to verify: Review Anthropic's S-1 filing with the SEC EDGAR database for customer retention metrics. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Primary sourceAnthropic mentioned using AI methods to reduce CUDA migration time from a year to a few days.
Steve Hou ·24:51
How to verify: Review Anthropic's S-1 filing or public statements regarding infrastructure and AI-assisted development. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Primary sourceThe A100 GPU, released in 2020, retains approximately 22% of its original sale price in rental value.
Steve Hou ·28:23
How to verify: Check Silicon Data's GPU rental indices or cloud provider spot pricing for A100 instances. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
The H100 GPU retains approximately 50-60% of its original sale price after 4-5 years.
Steve Hou ·28:42
How to verify: Check Silicon Data's GPU rental indices or cloud provider spot pricing for H100 instances. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
The B200 GPU has a fair market value of 180% of its original sale price.
Steve Hou ·29:02
How to verify: Check Silicon Data's GPU rental indices or cloud provider spot pricing for B200 instances. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Memory and storage stocks (e.g., SanDisk) have sold off to single-digit forward multiples.
Steve Hou ·17:38
How to verify: Review stock performance and forward P/E ratios for memory/storage companies like Micron or SanDisk (now part of Western Digital/Kioxia) via financial data providers. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Anthropic is one of the biggest users of its own tokens for developing models.
Steve Hou ·31:55
How to verify: Review Anthropic's S-1 filing or public reports on internal compute usage and token consumption. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Primary sourceVC-backed startups are able to raise funds and buy compute using money made from early investments in Anthropic and OpenAI.
Steve Hou ·32:03
How to verify: Review SEC Form D filings for new VC fund formations and public disclosures of startup funding rounds linked to AI compute procurement; cross-reference with VC portfolio disclosures. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Primary sourceCurrent AI spending is concentrated in the top 1% of companies.
Steve Hou ·33:20
How to verify: Consult enterprise AI spending surveys from firms like Gartner, IDC, or McKinsey to analyze spending distribution across company size/sector. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
On-demand GPU rental rates have risen vertically in the last few weeks prior to the recording.
Steve Hou ·34:26
How to verify: Check GPU rental price indices from providers like Vast.ai, RunPod, or Lambda Labs for price trends in the 30 days prior to 2026-11-12. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Algorithmic improvements such as mixture of experts and multi-head attention are increasing memory efficiency.
Steve Hou ·35:05
How to verify: Review technical papers from DeepSeek and US labs (e.g., Meta, Google) detailing model architectures and FLOPs/memory usage metrics. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Users are engaging in longer conversations and using more memory as models become more efficient, illustrating the Jevons paradox.
Steve Hou ·35:25
How to verify: Analyze API usage logs or public benchmarks from AI model providers showing average conversation length and token consumption over time. For recent events, verify 2026-08-12 through 2026-11-12. Use another period only when the speaker explicitly names it.
Do your own research
- Verify Silicon Data's GPU rental price indices and token usage metrics at silicondata.com.
- Check recent SEC filings from major AI labs (e.g., Anthropic, OpenAI) for profitability and retention data.
- Analyze on-demand vs. reserved GPU rental rate trends to assess supply/demand dynamics.
- Review enterprise AI adoption reports to determine if spending is diffusing beyond the top 1% of companies.
- Assess Nvidia's depreciation schedules and competition from customer-built ASICs.
- Investigate the market share of open-weight models in enterprise settings versus public developer platforms.
- Monitor algorithmic efficiency breakthroughs (e.g., mixture of experts) and their impact on compute requirements.
Full transcriptexpand
- 0:00
- Phil Rosen: Steve, most investors rely on stock prices to look at the AI trade, right? But you are tracking compute GPU rental prices and other variables that most people don't have access to. Where do you think we are today in this AI cycle?
- 0:13
- Steve Hou: Well, uh, thank you very much for first of all for having me back. And uh uh indeed, I mean, so Silicon Data, That we track uh GPU rental pricing and uh um token you know usage and pricing.
- 0:26
- Unidentified speaker (UNKNOWN): Uh
- 0:27
- Steve Hou: we think of it as almost a bit like AI fundamentals, right? Um and fun and and and and when we say fundamentally, what market really you know cares about and recently in the last few months been nervous about is the economics of AI businesses. In particular, if you think about uh there's the AI
- 0:46
- Steve Hou: labs and inference platforms that serve post-trained open model. tokens. Uh, and then you've got the neo clouds, uh, so to speak, and uh the cloud providers uh that are providing the compute from which uh, you know, people and the labs and the inference platforms can rent uh um, you know, GPU. And then of course
- 1:06
- Steve Hou: some of them are that this these days feels like the dentists are shifting a little bit. The neo clouds are trying to get into serving tokens and the uh the the the inference platforms and uh the clouds labs are also trying to own compute. Regardless, I think the thing that you should pay attention to, and this is a chart that uh Silicon Data has been putting out a lot, and we actually have a new blog coming out about exactly this
- 1:26
- Steve Hou: is the basically the spread, right, between how much uh tokens sell for, right? And how much it costs you to uh rent the a GPU, right?
- 1:37
- Unidentified speaker (UNKNOWN): And
- 1:37
- Steve Hou: you can almost think of it as a bit like you. in traditional macroeconomics, does the the set the squeeze between CPI and PPI, right? You know, like fundamentally AI is a very complicated technological problem. But from a business point of view, it's actually maybe a lot less you know complicated. It's just, you know, price minus cost, right? Can is it a good business? Fundamentally On rentre it
- 1:57
- Steve Hou: comes down to do you have good sustainable margin, right? Price minus cost. while right able to generate that additional uh demand so that when you multiply price minus cost uh with with quantity, you get a bigger number
- 2:11
- Unidentified speaker (UNKNOWN): at
- 2:12
- Steve Hou: the end of the day.
- 2:13
- Phil Rosen: Okay, so I'm gonna follow up with one of the things you just mentioned. GPU rental prices are right near a record high. They've been going up all year.
- 2:21
- Steve Hou: Yeah.
- 2:22
- Phil Rosen: Walk me through what this means from the investor perspective.
- 2:26
- Steve Hou: So again, depends on what investor you are, right? Whether you are investor in a neo cloud, whether you invest or private investor for that matter in one of the labs or or neo labs, as they call them, or inference platforms. So these uh on on-demand rental uh indices which we publish, uh on which we're going to also launch
- 2:46
- Steve Hou: futures contracts uh subject to CCFCC approval, these actually capture the on-demand rental costs. In other words, how much does it cost you to run GPU when if you just go out there and pay the floating rate? This does not, by the way, map to if you had reserved compute for say three, five years in the last couple of years, at which you might have
- 3:05
- Steve Hou: reserved a price that's substantially lower than the on-demand price. Nevertheless, right, just because you have reserved a price, or even maybe even on computer. compute, you still want to think about the opportunity cost, right? So if you are someone who actually owns a lot of compute, if you are one of the clouds or even a hyperscaler, uh, you also are thinking about, okay,
- 3:25
- Steve Hou: if I can do this one thing and make a thousand dollars, but you can also if I can run it out for $5,000 by doing this one thing for $1,000 and effectively incurring a $4,000 four thousand dollar loss, it's got gotta to be real worthy worth it to me in the long run strategically to do this thing. Maybe one of want to grow my frontier AI model capabilities eventually I'm gonna have a different, you know, business. But uh so from the
- 3:45
- Steve Hou: investor point of view, right, if you look at the surging compute costs, that's always gonna be weighed against the broader economics of uh what tokens are being sold for and also what is the mixture, crucially, right, of the models that are actually being used, right? So
- 4:01
- Unidentified speaker (UNKNOWN): what
- 4:01
- Phil Rosen: what does that mean? The the mixture mixture of the models being used?
- 4:04
- Steve Hou: So let me actually sort of give you, you know, like what what I mean is that are people use using mostly the sort of the frontier model, the fables and and the you know opus of the world, or are they mostly using the flash models that are substantially cheaper and much, much faster? Uh and of course, there this could be a big trade-off, right? So I'd like to give an analogy. Is like um
- 4:24
- Steve Hou: so so it's harder to have an intuition. for most people when they're thinking about AI models, but imagine we're thinking about cars, right? Imagine pretending we're at the dawn of cars being invented, 10 cars that are Mercedes, 50. Let's say each one is sold at $50,000 with 35% margin, right? The first year you sell 10 cars, right? The average car price is $50,000.
- 4:45
- Steve Hou: Come next year, right? The cars absolutely blow up. Everyone wants a car now to get around, and you sell 10,000. cars, where you go from 10 to 10,000 and you sell a thousand Mercedes, right? So the sales growth for Mercedes is actually exploding. You went from 10 to 1,000, it's a hundred times explosion. Meanwhile, you've got 9,000 uh
- 5:05
- Steve Hou: Toyota Corolla cars, right, that are cheaper, $10,000 apiece, maybe 15% gross margin, right? So come next year, the average car sold is such. actually sold at a lower price and a lower average margin. So the business has on some metrics look like it's actually gotten a little bit worse. But in any reasonable
- 5:25
- Steve Hou: absolute terms, the business has gone fantastically much, much bigger. And I think that's an analogy that I think one should carry forward with a lot of these uh, you know, say frontier AI labs, right? You know,
- 5:36
- Unidentified speaker (UNKNOWN): and
- 5:36
- Phil Rosen: the analogies for tokens?
- 5:39
- Steve Hou: The sales of tokens, right? So if you are one of the frontier labs, let's say anthropic or open OpenAI, you sell not just one model, right? You sell many different types of models, right? You've got your most frontier model, and then you've got your less than frontier, non-frontier models that are much cheaper. Yeah, you've got even these models that are flash models, right? And uh what which models do you want to
- 5:59
- Steve Hou: sell to people? Because on the other side of it, you're still running GPUs to put produce all of these tokens, right?
- 6:05
- Steve Hou: So there's a trade off, and uh people are always evaluating this sandwich, right, between how much you're able to sell what kind of tokens are you able to sell at what margin, and uh you know what is actually your cost of input.
- 6:19
- Unidentified speaker (UNKNOWN): This
- 6:19
- Phil Rosen: episode is sponsored by Y'all Street Asset Management. If you own a gold or silver ETF, the vast majority of its metals are likely. stored in foreign vaults under foreign jurisdictions. That's why Yall Street built YSAU, the Yall Street Physical Gold ETF, and YSAG, the Yall Street Physical Silver ETF.
- 6:39
- Phil Rosen: The physical gold and silver backing these funds are not only fully allocated bars, but 100% of the metal is held in the United States at Texas Precious Metals Depository. You get the simplicity of an ETF while knowing the physical gold and silver itself is held here in America in fully allocated bars. The accessibility of Wall Street with the
- 6:59
- Phil Rosen: independence, common sense, and entrepreneurial spirit of Texas behind it. If where and how your gold and silver allocation is vaulted matters to you, check out YSAU and YSAG. Y S A G. To learn more about the funds, visit Yalstreetfs.com. YallstreetETFs.com.
- 7:18
- Unidentified speaker (UNKNOWN): To
- 7:18
- Phil Rosen: your point here, I have a chart from you, Silicon Data, that shows token prices are down dramatically. Looks like they've fallen off a cliff since about May this year. This has happened even as GPU rental rates are going up and to the right.
- 7:34
- Unidentified speaker (UNKNOWN): Yeah.
- 7:34
- Phil Rosen: Talk me through this dynamic between falling token costs, rising rental. rates, and what that means for the broader AI economy.
- 7:43
- Steve Hou: So indeed, this is a chart that's actually drawn a lot of attention, you know, uh in the market. And uh um there's a quite a bit of misinterpretation of it as well, in that this is not a simple price chart. As I was telling you before, um, this is a effectively a usage usage-weighted weighted price index, right? Normalized to
- 8:03
- Steve Hou: a million. So if people do the following thought experiment, if people went from uh using in exclusively the frontier model, like you know, whatever is the frontier of the day was, right? Back in May, it was a different model. Let's just say it was, you know, myth, mythos or fable or whatever, to now using exclusively Basically only uh the flash model, the
- 8:23
- Steve Hou: cheapest po possible model, haiku, right? You're gonna see a sharp depreciation or deflation of the effective price of a million tokens. Which
- 8:31
- Phil Rosen: is good. Which
- 8:31
- Unidentified speaker (UNKNOWN): is
- 8:32
- Phil Rosen: consumers.
- 8:33
- Steve Hou: Which I mean, I don't know about good or bad. I mean that's that's a preference thing, right? You know, it's a bit like is it good or bad to drive a cheaper car? I mean, some people like just the fact that they are driving a cheaper car that's more fuel efficient and gets gets the thing done and they basically drive everywhere, or some people uh prefer driving, you know, a a Gosler Gussler and and uh you know they like so there's a preference built
- 8:53
- Steve Hou: in. But from a business point of view, if you are a seller of tokens, right, if you're at the AI lab, uh you obviously have a preference about uh what type of models get sold. You wouldn't mind selling a lot of the cheaper models, the tokens, if they actually get sold at a much, much higher volume, right? So suppose, you know, I give you two
- 9:13
- Steve Hou: models, one is half the price, but you know, people are going to consume a hundred times more, right, of the cheap, then you still end up making a lot more money, right? So um so part of what is being captured here in this picture, what is missing from this picture rather, is the uh the quantity compensation or oftentimes, I think at this point I've gone a little bit uh uh cliche, the
- 9:33
- Steve Hou: Jevans paradox, which is effectively a one uh a point of view about the response of quantity demanded in response to price change, right? How much more does
- 9:43
- Unidentified speaker (UNKNOWN): it?
- 9:43
- Phil Rosen: Explain that in plain language.
- 9:45
- Steve Hou: Plain language is like you know, demand curve is downward sloping, right? If you charge a lower price, people are gonna consume more of it, right? That's you know uh understandable. Uh the question is how much more of it, right? It uh if price goes down by half and demand only goes up 50%, I I end up losing some revenue. But if price goes down by half and pri demand goes
- 10:05
- Steve Hou: up a hundred times, I make way more money. Right. So and and between those two scenarios, essentially there is a debate about this elasticity, right? Of if price goes up by one percent, how much how many percents does the quantity demand go up? Right? Uh and that is fundamentally Given's Givan's paradox, is that Given's Givan's paradox is that that
- 10:25
- Steve Hou: number, that the elasticity is much, much bigger than one, right? You are more than compensated, right, for the price decline or the margin decline. Um and therein lies the entire debate.
- 10:36
- Phil Rosen: So if you had to sum up how you think the business of AI labs is going right now, based on the data you're sharing here, what would you say?
- 10:47
- Steve Hou: I would say if you did it a subtraction. right, of uh the price of tokens produced per GPU hour my uh minus right uh the cost of marginal cost of renting GPU, that number has gotten smaller, right? What what uh Don Wilson has called the cr the cr the crunch thread, right?
- 11:07
- Steve Hou: You're crunching numbers, you're crunching out models, right? In other words, you have got to work harder to make back the money, the break-even throughput. right? The number of tokens you have to produce to make back the money you're renting the GPU for is higher. In other words, you've got to basically meet a higher bar. Now, what is interesting is that what we are seeing, and we are seeing
- 11:27
- Steve Hou: another evidence as well, and we are trying to measure this ourselves, is uh it seems to be the case that with new, especially newer chips, right? The Black Bell Bowl 300 and so on, part of the reason why rental rate has actually surged is because you've had algorithmic improvements and of various kinds that the labs and inference platforms are
- 11:47
- Steve Hou: able to produce a lot more tokens, right? So that they are more than offsetting, right, whatever compression on the margin. So they're actually able to basically boost that Givon's Jivon's paradox, make it even more paradoxical, so to speak.
- 12:01
- Phil Rosen: Okay. So I have another chart here, which I think you can comment on as well. Almost three-quarters or so of token volume has moved to open weight models.
- 12:13
- Phil Rosen: This is up from almost nothing about a year ago. What do you make of that data? And also, again, what does this mean for something like an anthropic or an OpenAI as far as their business models? This can't be good, is my read on it.
- 12:27
- Steve Hou: Well, again, so first of all, uh what is publicly observable, which is where we draw a lot of the data, we have a number of you know uh uh inference platforms from which we observe usage it is not a full view of the entire market right it is a sh a a portion of the market you think like somewhere like open route or base tens of the world
- 12:47
- Steve Hou: where you can actually switch between models.
- 12:49
- Unidentified speaker (UNKNOWN): Uh
- 12:50
- Steve Hou: those tend to be, you know, either independent developers, smaller firms, small medium firms. Increasingly, I guess they're picking up like base 10 works with uh Harvey, which you know in turn works with law firms. But by and large, right, if you think your average large company, a brand name, uh blue chip company, they are likely renting and buying uh
- 13:10
- Steve Hou: tokens from uh either of Enthropic Anthropic OpenAI or both, right? Uh and we don't observe, right? And let's say they get it from uh one of the hyperscalers as well. That's also likely a lot of the times they end up using, you don't observe their usage. So this is likely an overstatement in terms of the adoption of open models. By and large, I would say uh American
- 13:30
- Steve Hou: consumers and enterprises are not really using that much open models yet, right? It is a thesis to be proven. You know, you see a lot of them being mentioned. in terms of capability and benchmarks and whatever. But in terms of practical usage, I mean I bet you have never used a Chinese open model before.
- 13:47
- Phil Rosen: Maybe not.
- 13:48
- Unidentified speaker (UNKNOWN): I
- 13:49
- Steve Hou: I know not. Like you probably have
- 13:50
- Unidentified speaker (UNKNOWN): have
- 13:50
- Steve Hou: not, right? And I I doubt you know anyone who has.
- 13:54
- Unidentified speaker (UNKNOWN): I
- 13:54
- Phil Rosen: guess w would I have to go out of my way to find it, or companies would put them in themselves.
- 13:58
- Steve Hou: You have to. I mean, so so part of we don't know exactly, for example, you you've had now have these uh you know uh agents, right? Use consumer agents, which will get you later, like you know, instinct and whatever. You really don't know what models actually power these, you know, uh agents on the background. So it could could be, but generally speaking, I would say just in terms of direct
- 14:18
- Steve Hou: contact where you actually intentionally choose the models you use, which is most of what we do these days, right? You go on cloud or you go on open OpenAI AI or whatever, you know, I don't think the usage rate is actually all that high. Nevertheless, what you are seeing there, it is a propensity. It is a, you know, it is a leading indicator in that uh since, for example, at the end of
- 14:38
- Steve Hou: May and early June, I wrote a sort of uh ex post uh quoting our uh token index data to mention the fact that uh you know this idea of token maxing was likely coming to an end and people becoming more, I think, sophisticated or at least savvy, right, about making the quality cost trade-off was coming. And that's the reason why
- 14:58
- Steve Hou: we've seen this sharp collapse in this index. Part of what of course in here, there's a lot of switch towards open models, but uh you don't even need open models to actually see that effect, right? A lot most of the recent collapse of that index actually comes from the fact that open uh proprietary labs, right, uh Anthropic and OpenAI especially, have been cutting
- 15:18
- Steve Hou: prices themselves when they issue when they come up with a new generation of a model, right? And uh there is competition between them and there is cannibalization within them, uh, such that um sometimes a lot of the day-to-day tasks may be saturated. Like I find myself rarely using Fable, right? Not because I don't like Fable,
- 15:38
- Steve Hou: it's just because whatever I need, Opus 5.5 more than suffice.
- 15:42
- Unidentified speaker (UNKNOWN): The
- 15:42
- Phil Rosen: frontier
- 15:43
- Unidentified speaker (UNKNOWN): model
- 15:43
- Phil Rosen: takes too many tokens. It
- 15:44
- Steve Hou: takes too many tokens, slower, you know, and you know, and so on. And and for the vast majority of my day-to-day task, I just don't really find myself needing it, right? And uh and that's a form of cannibalization, right?
- 15:58
- Phil Rosen: Hmm. Okay, so I want to take us
- 16:01
- Unidentified speaker (UNKNOWN): out
- 16:01
- Phil Rosen: of the data a little bit. bit here. When you think about rental rates and token costs, how is that translating right now to equity prices, stock market valuations? Like what are you seeing as far as uh what investors are looking
- 16:17
- Unidentified speaker (UNKNOWN): at?
- 16:18
- Steve Hou: Um, so part of what the diamond we were just describing, by the way, of uh this margin compression and competition and the worries over. the sustainable economics, right, of the AI labs, uh the Frontier 2 labs, right, is actually what has been, I think, causing a lot of anxiety in the stock market. Like we basically saw like a local peak in
- 16:38
- Steve Hou: AI related, you know, stocks, um, even though the overall index has actually hovered around the highs, we're just off like a couple of percentage points high from the highs uh on the SP. But so if you look at sort of the AI names, especially the semiconductors, we are well off the highs, you know, like a median stock is maybe off 50% or something before the most recent rebound.
- 16:58
- Steve Hou: And that's exactly reflecting the fact that that tells you that the stock is actually pricing off of the second derivative, right? The rate of growth. We're not saying that AI spending or the AI build-out build out is not growing anymore, right? They're still absolutely growing. It's just that maybe we may be anticipating a couple of years
- 17:18
- Steve Hou: out, let's say 2028 and onward, that we're not going to keep accelerating. Right. And asset prices is always going to tell you what assumptions and expectations have been baked in, right? Uh and and and this was actually very telling, right? Uh the fact that memory stocks, you know, have uh sold off so so so
- 17:38
- Steve Hou: so hard and uh you know storage stocks like Sandisk sold off such that they are trading as single digit multiples or even low single digit for forward multiples tells you that uh we were pricing in a whole bunch of uh you know forward expectations so it's actually been interesting. So now a few months later now I think we are basically
- 17:58
- Steve Hou: um I I I would argue you know I've always we have flushed out a lot of the excess. Um there's I think I saw some chart down the way in. Uh someone was showing that uh, you know, uh leverage ETFs, you know, leverage has come significantly down and uh you have had multiple rounds of uh uh uh various news about uh
- 18:17
- Steve Hou: hedge funds uh running into issues and so on and so forth. So I think it's always always a trade-off between fundamentals and positioning slash expectation. expectations. Right.
- 18:28
- Phil Rosen: So if you look at the market, it's it's coming back, I would say, on the AI trade, but it's it's recovering from lower lows over the last month. Do you think valuations today are attractive, or you're still gonna wait for either a different catalyst or a different level?
- 18:48
- Unidentified speaker (UNKNOWN): I mean,
- 18:49
- Steve Hou: I uh it's hard for me to on I think, make such recommendations, so I wouldn't try, but I would say what we're just saying, like I think is always a race between fundamentals and uh, you know, uh expectations um and slash positioning. Um so I think fundamentals are actually showing up turning to be a little
- 19:09
- Steve Hou: improving a little bit, right? You know, in that uh we've been plagued with this whole question about whether or not there's gonna be going to return on investment, whether on the micro level or aggregate level, uh, whether or not the labs are going to have good margin. But uh most recently we have had the S1 filing from Anthropic and then the amended S1 filing, and we are seeing like a lot of I think very
- 19:29
- Steve Hou: promising you know signs, right? They have got very good margin, they have got very low burn uh burn rate, you know, they they've these run very efficiently, they even manage managed to be cash flow. positive, you know, for a quarter. Let's see whether or not it remains to be the case. Regardless, I think they're trading somewhere like uh like maybe 20%, 20% or maybe even
- 19:49
- Steve Hou: like the teen level uh you know uh four multiple right on revenue, which is I think high, but not crazy. Like what is and what is Plot Plallantier Palantir or SpaceX trading at, right? So that I think is actually an interesting sign that uh maybe, first of all, the economics is not as bad as a lot of the people were
- 20:09
- Steve Hou: making it out to be. And then there's the further question about uh uh adoption, right? Uh the biggest worry is that there is gonna be a gap of adoption, uh, that demand doesn't show up as quickly as supply comes here, right? Because we kind of have an idea when supply of build-out capacity is gonna get here, but the demand. has to do with how quickly
- 20:29
- Steve Hou: people adopt these tools. We've already sort of saturated a lot of the easy to saturate, like software engineers, coders, and and so on. Like is the next phase gonna be a lot harder? And I think basically with the new models and the fact that they've managed to capture a lot of the enterprise use case. And I I think there there's is also signs, the
- 20:48
- Steve Hou: reasons to be pro uh uh optimistic. optimistic on that front. So that's what I mean by things on on the um fundamental front. Combine that with the fact we just talked about how positioning has become less excessive, you put yourself in a maybe less, you know, uh uh yeah, a better position, so to speak. Yeah.
- 21:05
- Phil Rosen: Steve, I feel like you have one of the more unique vantage points on AI and and sort of the behind the scenes economics. So if we were to go through the value. chain, right? We have the hyperscalers, chip designers, uh, neo clouds, you have the frontier labs. Where do you think the most favorable economics are going to accrue to over the
- 21:25
- Phil Rosen: next year, let's say?
- 21:28
- Unidentified speaker (UNKNOWN): I
- 21:28
- Steve Hou: think it keeps shifting. And that's what's been interesting and challenging, right? Uh we've been through the phase when everyone's was saying, oh, like models are getting commoditized and open models just. gonna sweep through everything and Anthropic is going out of business and OpenAI is going to basically be bankrupt the next next quarter. And then you you realize, oh wait a second, actually
- 21:48
- Steve Hou: Anthropic not only are they not, are they actually improving, they've got this enormous capture of the enterprise the same way that Windows manager capture maybe enterprises and they actually, you know, they're the I think the retention rate from the S1 reveal is like 900%. You know, and and so there's a lot of, I think, good user experience that's actually being
- 22:07
- Steve Hou: rewarded, uh, such that uh the I think proliferation and the threat posed by open models actually remains a hypothesis to be proven, right? The null hypothesis remains the case that a US-based kind of a uh you know branded you know uh uh a uh clo close a proprietary model uh appropriately
- 22:27
- Steve Hou: priced, is going to be you know here to stay. Longer term remains to be seen, right? The Alex Carp argument about uh companies caring about the IP and just gonna wholesale adopt some kind of post-trained A open model. We'll see whether that happens. So I think the the the the worries over the labs actually gone away a little bit. Now we are seeing
- 22:47
- Steve Hou: on the other side of it uh you know the hardware, they suffer. from
- 22:52
- Steve Hou: the adj adjustment of expectations. You know, we have basically various reasons to expect uh AI builder to slow down, right? Not to mention uh the late with the latest being uh you know the midterm election and and the local populist pushback against the build out. That it's likely that uh even if we will continue, right, which we will I
- 23:12
- Steve Hou: expect us to continue, uh the the sp the the pace and the cost per megawatt of building out additional data center is likely to be uh uh l uh uh lower, right? So higher cost and slower pace. Uh the higher cost because you have to compensate local communities and you have to pay higher taxes, likely you won't be getting
- 23:32
- Steve Hou: as much tax rebate. Bait, That just means that for the same dollar amount in terms of capex, even if you see the same amount of dollar capex. increase, it doesn't buy you as much in terms of megawatts of compute, right? And that has implications for the downstream suppliers, whether you use a chipmaker, whether it's a testing equipment and memory and so on, right?
- 23:52
- Steve Hou: And then there is also uh the issue of competition. Uh, you know, uh, with NVIDIA being the dominant uh player in this space of CUDA uh uh for the longest time, this market was always NVIDIA's to dominate and theirs to also potentially, you know, uh uh uh see competition, right? Uh
- 24:12
- Steve Hou: AMD and others have have been trying, but for the longest time you have had CUDA being a, I think, a pretty powerful mold. But uh the with the latest, I think, you know, filing from Anthropic and even discussion from OpenAI that AI is actually allowing developers to make it a lot easier to transition from one type of chip to
- 24:32
- Steve Hou: another, because you can just sort of like rewrite the code. You know, AI is able to do it for you. And in fact, I think Anthropoc mentioned AMD explicitly when they use methods, how like they reduce the the the amount of time to sort of win off CUDA from you know a year to maybe a few days, uh, which is like a very, very
- 24:51
- Steve Hou: stark chart. I maybe even quoting the length a little bit off, but regardless, that's the intuition, right? So if that is the case, does that mean that uh you're gonna see a
- 25:00
- Unidentified speaker (UNKNOWN): you
- 25:01
- Steve Hou: know sort of uh uh a shift of, so to speak, the rent away from the hardware again back towards the labs and the software layer? Um so you remains to be seen, and that's part of the reason why I think the AI. trade has gotten a lot more volatile and and a lot harder, right? Compared to a couple of years ago.
- 25:20
- Phil Rosen: So may maybe I'm being too simplistic about this,
- 25:23
- Unidentified speaker (UNKNOWN): but
- 25:24
- Phil Rosen: if we think about the charts we showed earlier with GPU rental rates just going up and up and up, isn't that extremely bullish for NVIDIA? year And then maybe we just forget about everything else.
- 25:37
- Steve Hou: Again, uh we we were talking about earlier, like equity market prices, not what's happening today, but prices the future, in particular prices, the second derivative of uh the growth rate, right? Of uh acceleration. Is it is it growing at an accelerating pace or is it growing but at a decelerating pace? That remains
- 25:57
- Steve Hou: to be seen. Uh again, as I was saying before that uh uh NVIDIA or any hardware supplier, you know, uh he always faces competition. And um, I think so far NVIDIA has actually defended and even some some instances strengthened right their competitive place in the marketplace. But uh it remains to be seen because there's certainly
- 26:17
- Steve Hou: no shortage of competition coming from I think every single customer of NVIDIA, a large customer, also have their own ASICs custom chips, right? Whether it's TPU from Google, whether it's training from Amazon, whether it's uh, you know, uh all these uh uh accelerators like Cerebras and uh and so on. These are fast inference chips. So
- 26:37
- Steve Hou: competition is coming and especially combined with the power of AI on the coding front, uh it may just be a lot easier potentially, right? To to uh switch out, you know, hardware.
- 26:49
- Phil Rosen: When you think about the depreciation question how do you think that individual investors should be factoring that in as they're trying to let's say pick stocks in the AI trade?
- 27:02
- Unidentified speaker (UNKNOWN): So
- 27:03
- Steve Hou: the depreciation question, why is it so important? Why does it get brought up? I think it's probably a bit confusing to the average investor because they're not thinking about this question all that actively. There's a there's a couple of things. Number one, most direct directly is that uh from the perspective of the people that are buying the GPUs, uh, let's say the hyperscalers, uh,
- 27:23
- Steve Hou: depreciation, you know, uh cuts against your earnings, right? You make a bunch of money, and if you have to toss out the machines um uh you know very quickly, that means that you're not really making them on the business not that good, right? You're you're you're gonna you're going to uh I think earn very thin margins. Now what it what we have found is interestingly is that uh uh the whole
- 27:42
- Steve Hou: fear over uh fast depreciation that's actually been quite overblown, right? Uh we are finding in the data uh we can observe from not just today's on demand rental rates, but the forward curves, mark implied long term rental curves, that uh uh you can make some certain uh conservative assumptions about utilization rates of how many chips.
- 28:03
- Steve Hou: are being used at any given moment and also the decay to come up with a reasonable estimate on a market fair value. And this market value bar value is actually hovering, you know, even today for an A one hundred is a twenty two percent, which is a six year old chip. You should be long debt by now if you apply like a three year of straight line depreciation, right? You should be dead Debt within three years because it was released,
- 28:23
- Steve Hou: it was first announced in May of twenty 2020 twenty and then became publicly mass available. Uh and today, but six years plus later, it's still, I think, holding, you know, pretty significant rental rates. And then you look at H100 and B200 as the other two examples that we show you on that chart. I think H100 is about uh 50-60% of the original sale
- 28:42
- Steve Hou: price. And that's or also like you know, four, four or five years. And uh then you've got uh B200 is actually trading uh having a sh. a fair value of well over 100%, actually almost 200%, 180%. I think out by our estimate. Why what does that mean? That means that there's such a demand for it, and that the rental rate is so high that the
- 29:02
- Steve Hou: fair market fare value for it is actually substantially higher than his original sale price. Right. And I think that that that is actually a very interesting data point from the perspective of of evaluating. uh depreciation, that maybe we shouldn't just assume that these things go away in two, three years in a straight line fashion. Uh
- 29:22
- Steve Hou: that's
- 29:23
- Phil Rosen: very encouraging.
- 29:24
- Unidentified speaker (UNKNOWN): Well,
- 29:24
- Steve Hou: you can yeah, exactly. Right. So from that perspective, if you are, you know, uh s certainly from Nvidia's point of view and also from their customer point of view, uh this is good news, right? That means that uh these GPUs can run for quite a long time and they can be productive and generating tokens, producing revenue uh for you for a long time, and
- 29:44
- Steve Hou: uh uh uh that means better economics. And which in turn, of course, also matters for financing, which is the other reason why uh depreciation gets so much mentioning these days, is because so much of what's happening today, the bottleneck is actually coming down to capital, right? Interest rates are much higher, inflation is higher, and there is a lot of compete competing users. uses for capital, includ not
- 30:04
- Steve Hou: least from the government, right? So, how do you actually find that source of capital when lenders are becoming more and more, I think, cautious about the fact that there's so much lending that went into AI already in the last few years against the uncertainty of uh, you know, these these businesses. So, part of what we are trying to unlock here is to
- 30:23
- Steve Hou: provide data and help people understand. the economics of GPUs, connected GPUs as a capital asset that produces income. So that's maybe, maybe we can also think about lending towards the GPUs as a collateral in the same way when you obtain a mortgage on a home, some part of the lending is lent against
- 30:43
- Steve Hou: the home as a source, as a as a piece of collateral, right?
- 30:47
- Phil Rosen: Steve, is there anything that you're looking at today? that you think investors should be more worried about?
- 30:55
- Unidentified speaker (UNKNOWN): I
- 30:55
- Steve Hou: think it's the same set of worries that uh you had you already mentioned. I think the interesting things is just that you have to sort of watch it, right? Because we don't really know which way is going to go. And part of the reason why we at least have a bit of a self-plug, plug, our data is helpful, is you see it allows you to have that sort of daily high frequency view. Uh, you know, the question
- 31:15
- Steve Hou: about demand airpocket, right? You know, uh is the demand for AI uh uh mo growing fast enough to uh make you know sort of justify the the supply. In fact, you know, uh in fact I think justify the expectation that demand will outstrip supply for you know a long time. And
- 31:35
- Steve Hou: there's another question which we do not directly answer just yet, which is basically the quality of demand, right? A lot of the demand for AI today does indeed come from within AI industry or AI adjacent industry, right? You've got basically the labs, right, buying compute. And they also use a lot of the tokens. You know, Anthropic is one of the biggest users
- 31:55
- Steve Hou: of their own tokens for developing their models. Uh, and then you've got these inference platforms that are or or or AI.
- 32:03
- Steve Hou: related startups that are VC backed, right? And these VCs, you know, probably made a lot of their money from investing in anthropic OpenAI and early on. And they are able to raise funds and fund some of these new ventures, which in turn take the money and buy the compute.
- 32:20
- Unidentified speaker (UNKNOWN): You
- 32:21
- Steve Hou: know, therein you start hearing some of these like, uh, you know, I think some may interpret it it's as like bearish, you know, sort of a uh narrative of a circular financing. But in our view, like basically, I mean my personal view, that is, uh uh this is I think kind of a normal part of the cycle when you build out something like this, especially when the the the technology is very you
- 32:41
- Steve Hou: know uh capital intensive and it is going to come from some some in you know you uh a concentrated source. So some degree of this is uh inevitably going to happen. The question. is, does it actually proliferate? Does it actually diffuse? Do we start seeing uh non-AI companies actually spend meaningful amount meaningful amount of uh uh money on
- 33:00
- Steve Hou: AI in various forms because they actually derive value from it by they see actually an impact on their individual worker productivity. They can see an acceleration to their the top line growth, which means that they're willing to spend more on AI. And I think that they're in therein lies the question. And I've got my friend Era at Ramp AI. He has got all these charts and he sort of
- 33:20
- Steve Hou: has been pointing that, pointing out that in the trade rear view,
- 33:24
- Steve Hou: rear viewer, rear mirror view, right? Most of the spending is very concentrated in the top 1% of the company. And a lot of it is AI related. But you know, time will tell whether or not that that proliferates. And it doesn't really have to take that much for that picture. to actually just change, you know, enough.
- 33:41
- Phil Rosen: So what what stands out to me about a lot of the data and the research that you do, nothing necessarily is flashing a bubble signal. And maybe I'm not reading it the right way. But what do you think of the ongoing bubble question?
- 33:58
- Unidentified speaker (UNKNOWN): I
- 33:58
- Steve Hou: think the bubble is actually a little bit of a, I think, tiresome, oversimplication.
- 34:04
- Steve Hou: Um like
- 34:05
- Unidentified speaker (UNKNOWN): I
- 34:06
- Steve Hou: what is the bubble? The bubble is basically the worry that we're going to overbuild and that someday like when I wake up and realize oh we have got all this excess compute, right? That uh where nobody's gonna need. And does sort of people think about this dark fiber back in the dot com era and uh you know you basically end up building a ton of uh uh hardware that didn't get
- 34:26
- Steve Hou: used. Today that's certainly not the case, right? You know, all the GPUs are used are being used. In fact if you look at the on-demand uh rent arises actually going vertical the last few weeks it's not just going up it's literally m you know going up like the GPUs are melting. Now uh of course we just discussed you know whether or not the quality and composition of the demand but leave that
- 34:46
- Steve Hou: aside I I think you know sort of the the w the worry someone has to have is like you know, oh, are we gonna at some point have some kind of a magical algorithmic improvement that will make the models compute in a completely different way where you don't need multi matrix multiplications anymore. Maybe like a lot of the compute is not needed. Now we it's not like we haven't been seeing
- 35:05
- Steve Hou: algorithmic improvements, right? Mixture of experts and all these sort of uh you know uh uh uh uh multi-head attention and you know types of algorithmic improvements from me from China. DeepSeek and elsewhere, even within the US labs, um, that uh is actually making these efficiency gains and we are a lot more memory efficient. But
- 35:25
- Steve Hou: again time and again we are seeing this microJevons paradox jumping in all over the place. When we become more efficient at using memory, we just use a ton more memory. The conversation goes longer, right? People have these super long interactions with us within a single conversation to try to get a different, better output, which I'm sure you've experienced
- 35:44
- Steve Hou: yourself, right? You go back to the same session and you if we recall a year and change ago, you couldn't do that, right? That model was tap out. But today you can keep going. So but even now I still find for myself in running out of uh space in some of my you know sort of research pieces. So
- 36:02
- Unidentified speaker (UNKNOWN): you know
- 36:03
- Steve Hou: more algorithmic improvements would be welcome.
- 36:06
- Phil Rosen: Steve, you got a lot of knowledge here. Um and I know you're as in the weeds as anyone with the the AI economics. Where can we send people to find more of your work?
- 36:17
- Unidentified speaker (UNKNOWN): Well
- 36:17
- Steve Hou: they can come to our website silicondata.com and I'm starting to You publish our blog series much more frequently we're aiming for like a one once a week. Uh And also they can find me on my Twitter account, both uh Synacon Data, you know, the handle official account, and also my personal account, which of course contains a lot of my personal views on on AI and on
- 36:37
- Steve Hou: non-AI topics.
- 36:38
- Phil Rosen: Amazing. Uh Steve, I I think was this your third time on the show? Fourth time? I
- 36:43
- Steve Hou: think it might be my fourth time.
- 36:44
- Unidentified speaker (UNKNOWN): Fourth time,
- 36:44
- Phil Rosen: wow. And every single time we record, I just try to keep up with you. So I appreciate your time as always. And uh we'll do it again.
- 36:51
- Steve Hou: Thank you so much.
More from Full Signal
Veteran Strategist: 3 sectors to BEAT THE MARKET in 2027
Episode discussion: Veteran strategist outlines AI infrastructure risks, Fed policy, and sector rotation for 2027
Speakers’ claims: Speaker 1 argues that market fundamentals remain strong with broad-based earnings growth, attributing recent volatility to sentiment rather than economic deterioration. He identifies utilities as a key beneficiary of AI power demand and predicts a 'low hire, low fire' labor equilibrium driven by productivity gains. Speaker 2 highlights political and regulatory headwinds, including NIMBYism and potential policy reversals, as risks to the AI trade. Both speakers emphasize diversification beyond mega-cap tech and semiconductors to mitigate headline risk.
Wall Street Insider: 3 stocks WIN as the AI trade CRASHES OUT
Episode discussion: Speakers analyze AI trade risks and highlight Visa and Walmart as resilient picks amid market rotation.
Speakers’ claims: Speakers discuss the market reaction to Nvidia CEO scandal headlines, arguing the stock drop reflects headline risk rather than fundamental issues. They critique the sustainability of AI capex, noting cash flow concerns for hyperscalers and questioning productivity gains. The discussion shifts to Visa as a consistent pick benefiting from digital payment growth and stablecoin integration. Walmart is highlighted as a resilient consumer stock, outperforming amid an 'anti-Trump trade' hypothesis anticipating lower tariffs.