## Detailed walkthrough
### [00:00](https://www.bilibili.com/video/BV1E5L964EDo?t=0s) Course Setup, Stanford Logistics, and Personal Background
Course instructor Apoorv opens the inaugural session of MS&E435 (*Economics of the AI Supercycle* at Stanford) by welcoming students to the nine-week course. He outlines the agenda: personal introductions, course logistics, an immediate Day 1 quiz, and exploring the core question—"Where is the money in AI?"
Tracing his journey from India to Singapore and eventually to Stanford for grad school, Apoorv shares that he spent roughly 13 to 14 years at Palantir leading engineering teams (joking that they "wrote a lot of Spark in government buildings") before co-founding/leading Altimeter's investment operations. Altimeter operates across public and private markets with a concentrated investment approach focused entirely on this tech cycle. On a personal note, he reveals he became a father six months ago, calling it the "biggest investment" of his life with a guaranteed positive non-financial IRR.
- **Time Commitment**: Target of $\le 3$ hours per week total (1 hour of lecture + 1–2 hours of readings/debates with ChatGPT or Claude).
- **Format**: Weekly guest speakers operating under Chatham House rules (no recordings allowed), covering the entire stack—semiconductors, infrastructure, energy, foundation models (OpenAI, Anthropic), and applications/agents. Optional post-class dinners with speakers.
- **Grading**: 50/50 split between attendance/participation and a final course assignment released at the end.
> "Half of you are going to start an AI company, the other half are going to fund it. So at least you should know where to spend the Series A money that you're gonna raise."
---
[05:00](https://www.bilibili.com/video/BV1E5L964EDo?t=300s) The Structural Asymmetry: CapEx vs. Application Monetization
Apoorv shifts the discussion to the core framework of the course, presenting a visual comparison of economic distributions across tech cycles. Unlike previous platform shifts (Internet, Mobile, Cloud), where value accrued rapidly at the top application layer, the generative AI ecosystem currently resembles an upright pyramid (or inverted triangle): massive capital expenditure at the bottom substrate, narrowing down to a relatively small application revenue layer at the top.
Students propose several hypotheses for this structure: early-stage ecosystem lag, NVIDIA’s structural monopoly on compute, and fundamental differences in software delivery economics. Apoorv reinforces the economic shift: traditional SaaS benefited from Marc Andreessen's "software is eating the world" paradigm with zero marginal cost of distribution and 80–90% gross margins. In AI, incremental users consume compute, meaning high-scale software tools like Cursor or ChatGPT still face high operational expenditures even at multi-billion-dollar revenue scales.
- **NVIDIA Stranglehold**: Acknowledged as holding dominant market share over global compute capacity.
- **The "Five-Layer Cake"**: As conceptualized by Jensen Huang: Energy $\rightarrow$ Chips $\rightarrow$ Power $\rightarrow$ Interconnect $\rightarrow$ Memory.
- **Historical Comparison**: AWS took 8 years (from inception in 2004 to Amazon fully migrating in 2012) to flip its CapEx-heavy infrastructure buildout into clear application value creation.
---
[10:00](https://www.bilibili.com/video/BV1E5L964EDo?t=600s) Stack Breakdown, Incumbent Dynamics, and Ecosystem Economics
Addressing the individual layers of the AI supply chain, Apoorv details how profitability and competition vary across the market stack. He takes questions from students regarding how incumbent software giants (e.g., Salesforce, Palantir) and hyperscalers fit into the visual model, explaining that their spend is integrated via API/model consumption revenue.
```
[ Application Layer ] <- High user scale, margin pressures
│
[ Foundation Models ] <- OpenAI, Anthropic (Ads vs Subscriptions)
│
[ Inference / Middleware ] <- Most competitive; Feature vs. Platform debate
│
[ Infrastructure / GCP ] <- Cloud providers, TPUs, custom ASICs
│
[ Semiconductors / Hardware ] <- Highest profitability (NVIDIA)
```
- **Semiconductors & Infrastructure**: Characterized by massive capital commitments and high market concentration.
- **Foundation Models**: High user scale (e.g., OpenAI’s massive user base, Anthropic’s cloud reliance). The industry continues to debate whether subscriptions or ad-supported models will ultimately drive long-term profitability.
- **Inference Layer**: Identified as the most fiercely contested segment, caught between fast-growing startups and hyperscalers (AWS, Azure, GCP) absorbing feature sets into core platform suites.
> "The most profitable part of the stack is the semiconductor layer by a long shot."
---
[15:00](https://www.bilibili.com/video/BV1E5L964EDo?t=900s) Industry Equilibrium, Hyperscaler CapEx, and Training vs. Inference
The session concludes with an open discussion on what an "unsuccessful" or stable equilibrium looks like for AI economics, and what catalysts could cause the ecosystem pyramid to invert. Apoorv notes that while cloud stack shifts took roughly a decade to mature, the AI substrate’s hardware complexity may prolong the current heavy-CapEx phase.
He highlights two critical industry metrics and potential turning points:
1. **Custom ASIC Breakouts**: Success in hyperscaler internal chip design programs (e.g., Google TPU, Meta MTIA, Amazon, Microsoft) could reprice the semiconductor layer.
2. **CapEx Guidance Drops**: Ceasing aggressive capital expansion guidance on earnings calls would signal that current infrastructural returns are unviable.
3. **Training vs. Inference Workloads**: NVIDIA's fleet utilization is estimated at roughly 40% inference and 60% training. Inference workloads are inherently bursty and variable (dropping on holidays like Thanksgiving and Christmas), whereas training workloads maintain steady, high utilization over discrete blocks.
Apoorv advises students to listen to hyperscaler quarterly earnings calls four times a year to directly monitor how corporate leadership views CapEx expansion and unit economics.
[20:00](https://www.bilibili.com/video/BV1E5L964EDo?t=1200s) Profitability Gaps Across the AI Stack and Hyperscaler Market Dynamics
The financial distribution across the AI stack reveals a stark contrast between hardware infrastructure and application-layer software. NVIDIA’s data center revenue commands gross margins near $75\%$, though application-layer profits lag significantly behind, fluctuating between $0\%$ and $30\%$ depending on the provider. This massive margin divergence underscores a hyper-concentrated profit triangle where the semiconductor layer captures the lion's share of value, heavily driven by NVIDIA's dominant market position. While cloud providers undergo massive infrastructure investment cycles, application builders continue to struggle with lower unit economics and elevated compute costs.
When evaluating custom silicon initiatives—including Google's TPUs, custom chips from AWS, or OpenAI's pursuit of custom ASICs—the addressable semiconductor hardware market stands at approximately $\$300\text{ billion}$. As NVIDIA CEO Jensen Huang regularly highlights during earnings calls, roughly half of this revenue originates directly from a handful of major hyperscalers. Consequently, hardware startups entering the chip design landscape face an extremely concentrated market structure characterized by a tiny group of ultra-large enterprise buyers rather than a broad, diversified customer base.
The traditional internet supercycle demonstrates how vertical integration consolidates long-term market cap and pricing power. Google serves as the prime historical example, having built a dominant $3\text{ trillion}$ market valuation and a near $99\%$ share in web search by controlling the entire stack from custom server infrastructure up to the end-user interface.
- NVIDIA data center gross margins hover around $75\%$, while application-layer revenues yield $0\text{--}30\%$ margins.
- The addressable hardware opportunity totals roughly $\$300\text{ billion}$, with over $50\%$ concentrated among major hyperscale cloud operators.
- Silicon startups must target the top 5 hyperscale cloud providers as primary buyers rather than relying on enterprise long-tail adoption.
## Notable quotes
> [00:00](https://www.bilibili.com/video/BV1E5L964EDo?t=0s) Half of you are going to start an AI company, the other half are going to fund it. So at least you should know where to spend the Series A money that you're gonna raise. — Apoorv
> [10:00](https://www.bilibili.com/video/BV1E5L964EDo?t=600s) The most profitable part of the stack is the semiconductor layer by a long shot. — Apoorv
In the generative AI supercycle, economic value is heavily concentrated in the physical hardware layer rather than software applications, upending traditional software economics through persistent compute costs and massive infrastructure expenditures.
Structural Inversion of the Traditional Software Stack (5:00)
Instructor Apoorv illustrates the current generative AI ecosystem as an upright pyramid characterized by massive capital expenditures at the bottom substrate (energy, chips, and hardware) tapering to a comparatively small application revenue layer. This structure fundamentally breaks from the SaaS era, where software enjoyed near-zero marginal distribution costs and 80% to 90% gross margins. Because incremental AI users consume compute, applications like Cursor or ChatGPT incur ongoing operational expenses even at multi-billion-dollar scale.
Apoorv contrasts this with historical infrastructure transitions: Amazon Web Services required eight years (from 2004 to 2012) to convert its heavy CapEx buildout into clear application value creation. Given the complexity of the hardware stack—conceptualized by Jensen Huang as energy, chips, power, interconnect, and memory—the current CapEx-heavy infrastructure phase may last longer than earlier platform shifts.
Severe Margin Polarization Across the Stack (20:00)
Financial returns across the AI supply chain reveal extreme divergence between hardware suppliers and application developers. NVIDIA's data center business achieves gross margins near 75%, while application-layer margins range between 0% and 30% due to underlying compute expenses. This dynamic concentrates the vast majority of current industry profits at the semiconductor layer.
The intermediate layers remain under competitive pressure. Foundation model developers such as OpenAI and Anthropic are still navigating whether subscriptions or advertising models will drive long-term profitability. Concurrently, inference and middleware providers face an intense platform-versus-feature battle as hyperscale cloud providers (AWS, Azure, GCP) continuously absorb standalone inference capabilities into their core enterprise suites.
Hyperscaler Concentration and Catalysts for Rebalancing (15:00)
The addressable semiconductor hardware market stands at roughly $300 billion, but customer concentration is high: over 50% of this demand comes from a handful of major hyperscalers rather than a diversified enterprise base. Historically, deep vertical integration has captured enduring market capitalization, as seen in Google's $3 trillion valuation and dominance in search achieved by engineering its own server infrastructure through to the user interface.
Apoorv highlights that rebalancing this pyramid depends on hyperscaler in-house silicon initiatives (such as Google TPUs, Meta MTIA, and custom ASICs from AWS and Microsoft) and whether hyperscalers lower CapEx guidance on quarterly earnings calls. Workload composition also shapes economics: NVIDIA's fleet utilization is estimated at 60% training and 40% inference. Training provides steady, predictable utilization blocks, whereas inference demand remains variable and bursty, dropping noticeably during holidays.
Whether the AI economic pyramid eventually inverts to resemble historical software margins depends on whether end-user applications can overcome compute unit costs, or if hyperscaler custom silicon and cooling capital expenditures compress semiconductor pricing first.
[0:00] Apoorv: Half of you are going to start an AI company, the other half are going to fund it. So at least you should know where to spend the Series A money that you're gonna raise.
[10:00] Apoorv: The most profitable part of the stack is the semiconductor layer by a long shot.
Original transcript
[0:00] All right, folks. We're going to have some-- fun in this session. My name is Apoorv. I'm going to be your instructor for the next nine weeks or so. Um... Here's what we're going to go through today. I want to talk a little bit about myself. Why do I do this? some logistics on what to expect. Quiz, yep, I'm that guy. We're gonna have quiz on day one. And the biggest question that we've all been wrestling with, where's the money? Where's the money in AI? Some of you know me, but my journey started in India. I moved to Singapore. I met a couple of Singaporean folks here earlier. I started my career at Palantir with Sunil and a couple other folks about 13 years ago, 14 years ago. led a variety of engineering teams. All that to say we wrote a lot of spark in government buildings. And I came back to Stanford for grad school, which is when I got tired of writing Spark in government buildings. And now I lead Altimeter. I don't know how many of you have heard of Altimeter, but Altimeter is an investment firm. focus on a fairly concentrated form of investing. We've got two businesses, we've got a public business and a private business. And then I got the biggest promotion of my life six months ago. I'm now the proud dad. - Woo hoo! - Um, this is, uh... As people have told me, the biggest investment I will make, some have called it the one with the most guaranteed negative IRR. I think it's the most guaranteed positive IRR, not financial. But that's me. I live across the street. reach out with questions. I want to be here as a resource to you guys and make the most of it. We've got a great session lined up for you guys. The course is designed to be no more than three hours a week. and that includes class, readings all the time you spend arguing with Chad, GPT-weight, Claude, about whether it can do your assignment for you. So it's about an hour of class. It's about an hour or two of readings. And basically the format is we're going to do guest speakers every single class next class onwards Chatham House rules, a lot of guest speakers will share, maybe overshare. So please don't record what they're saying. We'll have an optional dinner with some of them right afterwards. You guys are welcome to join. Arrange the logistics. Grading is easy, 50/50. Show up to class. If Ali Goetze can show up to class, you can show up to class. And the other half has an assignment that we will release at the end of the course. Yeah, it's conversational. Ask questions, be involved. The more you're involved, the more you're going to get out of it. And honestly, what's in it for me is I'm going to learn the most from you guys. This is the course schedule over the next nine weeks. Lots of great speakers, one more impressive than the other. As for me, you know, honestly, putting this calendar together has been my my job for the last couple of weeks and months But be present. These are all incredible leaders running incredible businesses across the stack from semis to infra, to energy on the infrastructure side, to models, you would have folks from OpenAI, Ananthropic, and a bunch of applications and agents. So ask all your hardest questions, save them for the speakers, They're going to love it. and we're going to sign some readings. Thank you. So, Why should you take this course? What should you achieve in this course? What is a good thing to get out of this? You know, honestly... I thought about this and I was just telling one of the students here of how it all began. It's You know, I come back to campus once a year and I talk about all that's happening and you know typically it's in the context of finding great people it's I realized that This is such a big super cycle. We know it. We believe it at Ultimator. We have positioned our entire focus around it. And I did not find a course that goes deep in a way that I would have liked to be when I was an undergrad here. or a grad student here. And You know, I thought about in five years, everybody's going to ask you, hey, did you see it coming? You were at the start of it. You were around when ChatGPT was launched. You were around when, you know, the tectonic plates were forming and clay was forming. And I think you want to be able to say yes. So half of you are going to start an AI company, the other half are going to fund it. So at least you should know where to spend the Series A money that you're gonna raise. And at the very minimum, you know, you will have a sense of what not to go,
[5:00] These have mental models for "Hey, this business that I'm looking at, or considering starting, or considering funding, or considering joining, what are the right questions to be asking? What are the laws of physics that govern this business at this part of the cycle?" Thank you. I think it's gonna be the biggest one yet. And I'm excited that you guys are here to study alongside us. I'm going to spend some time on this slide because this is sort of the punchline. How many of you have seen a version of this before? Well, those of you did the readings, thank you. I appreciate it. We did include a Notebook LM for those who were more auditory inclined. But let's talk about this for a second. You know, what is going on here is actually probably the biggest the biggest question in generative AI right now, which is the If you listen to any of the earnings calls from the hyperscalers or or even NVIDIA and others is, We are investing so much into the capex we're investing so much into building these data centers where you know it's a five-layer cake as jensen calls it Energy. Chips. Power. interconnect memory, All that to give you a data center that you can either rent by the hour or by the token. that you can go train models on and serve those models. And then the question is, hey, these models that you built, are they creating economic value? That is basically the right hand side of this chart. And, To make an analog of the biggest technology revolutions that I have seen, you know, internet 25 years ago, mobile 20 years ago, Cloud, probably the most recent one 10 years ago, I put up one of those charts on cloud, but on the readings, you will see the same for internet and mobile and cloud. And that's the shape of the cloud ecosystem. The cloud ecosystem looks dramatically different than the AI ecosystem. Anybody have guesses as to why that's the case or your theory on why it's so different or reasons why this might look like? We're not going to call it a pyramid. We're going to call it a triangle, inverted triangle. Go ahead. Does it still hurt me, the site or AI? Yeah, definitely. Definitely, that's a good guess, yeah. Any others? Any other thoughts? But maybe because NVIDIA has like a monopoly so they can charge other Can you ask that question again next week when we have the film stream in video? But no good. It's a great point. They do have a stranglehold. One of the charts we had in the readings was the market share that MVA has on all of the compute right now. Um, Uh, It's up there. Any other thoughts or hypotheses on why this is so uh, different. Yeah, I don't know. Is it the cloud? stackers seem to be able to leverage that hardware to generate value for your apps. And then AI didn't really go up there. You know, we know how software ate the world, as Mark and recent said, software ate the world because You know, I could build software, could bid software, and I could distribute it to millions of people. And the marginal cost of running that software was close to zero. these software businesses ran at 80, some even at 90% gross margins. That is not the case with this new economic model of AI because if we have a set of users using cursor or using You hear all these stories about large-scale businesses that are still not profitable. at billions of dollars of revenue scale, is because of that, is because the incremental user of an AI application is not free, it's not marginally free, it's actually quite a bit more expensive to have AI users because Turns out you've got to burn those GPUs. And I would say everything you guys said from it being early to NVIDIA being dominant, we'll call it, to you know the the the physics of the problem are very different of how inferences run is certainly where we are right now. So I think that's the case right now. I might add another dimension to it, which I spoke about in the readings, was we analyzed what happened in internet. we analyzed what happened in mobile and cloud and How many years did it take? for these triangles to flip And, you know, One of the examples we take is AWS. AWS started in the year 2004. AWS has its first customer. in Netflix in 2010.
[10:00] And ultimately Amazon shifted fully to AWS in 2012. eight years from breaking ground, eight years from first CapEx investment cycle. I don't know if any of you were around reading earnings reports 20 years ago, but the big debate was, hey, is Amazon is going to go bankrupt. And that was the biggest That was the biggest question everybody had about the build out of AWS and Thankfully, nobody, at least yet, is on the verge of bankruptcy, but these are large numbers. Um, So we'll come back to this slide, but I would say this is the central theme of the course that we're going to explore. We're going to have speakers from some of the companies that are listed here to others and The central theme that we're going to pick around is like, hey, in your Field. with the NVIDIA speakers, are you A dominant. force, how long are you going to stay to be the dominant force? What are the forces that you're most worried about? Who are the ASICs that you're most worried about? What are the pricing compression vectors for your business? to the folks at Anthropic and OpenAI, who are gonna talk a lot about profitability. is you're serving a billion people user, franchise at OpenAI. With the Anthropic folks, honestly, 100% of this class is on cloud. So we'll ask them about, is this group of users profitable? How do you think about profitability? is ads gonna be a bigger source of revenue than subscriptions? And then for the folks in the middle, which is the inference layer, this is the most competitive part of the whole ecosystem. There's a lot of startups that are doing really well. They're winning so far. but you've also got the hyperscalers who want to have a dominant say in that So honestly, the jury's still out. And the biggest question there is, Are you a feature or a platform? a lot of new businesses that we are seeing on the infrastructure side that They feel like very good ideas. But if you ask yourself the question, hey, why is this not a part of AWS? you're thinking about maybe it should be a part of AWS. So for the speakers, we're going to talk a lot about that. Any questions before we jump into the quiz? Go ahead. I'm curious how you think about, so on the right hand side, like the triangle being like the application area being small, like how do you think about including like incumbent platforms into that? Like maybe like Salesforce, maybe revenue, like would you... Do you include them as part of analyzing that per minute how it shifts over time? It's a great question and I might add Salesforce, Palantir, there's a series of let's call them old economy businesses that are reinventing themselves to have SKUs of products that are, you know, in the case of Salesforce, Einstein, in the case of Palantra AIP, in the case of, you know, there's a series of these And the answer is yes, they should be. The answer is yes, that they should be. The way I solve for that in this calculation is I get the model revenue. And so if you were running Salesforce, you're probably running either one of the big models or running inference. So their spend is captured in the app layer by way of the substrate. It's very hard to extract that out from public disclosures, but yeah, we should. Yeah. Is a large part of the bottom part of the Ti-Gan side pyramid basically spying capacity for future revenue, which will accrue to the top, which is what we're not seeing in the future? - It's a great question. And maybe just to rephrase the question, the question is, hey, is there a timing mismatch in, the build out of the semis layer, because typically you build semis for a five year period or a six year period. but the application revenue is for right now? It's a great question and that's what makes the lower half of this color triangle somewhat cyclical. and you go through phases of uh, capex cycles. Think of it as like laying down the railroads. um that is very much the case and so there's a chart in the readings um for what happened in the mobile super cycle, something very similar happened. The first inning, had inflated market caps for a lot of the CapEx heavy businesses. And so if you think about a basket of CapEx names to call it steady state names. You should expect that. And so I suspect we are so early that that's happening as well. I'm curious to see how you think about Google, because I see that you label Google as Google Cloud or Google Cloud. but Google also has their own design models. How do you view layer composition in this triangle? any large conglomerate
[15:00] Google deserves to be, you know, we have to call business units. So I would put the TPU business unit in semis. We include that here. as we counted the revenue. Their GCP unit is in the infrastructure layer. And the Gemini unit is at the AppSleer. the you know, we have a chart later that we'll talk a little bit about. Gemini is actually one of the most used consumer applications. It's the second most used consumer application right now and The biggest question there is how much of that is coming from the distribution advantage that Google has to it meritocratically being such a good application. Jury's still out. But we'll get into it. Yeah. Well, let's go ahead. I just have a question about prediction. Right now, it looks like there's a triangle shape. And if it were to be successful, perhaps it should become inverted. But what does an unsuccessful new technology look like? Does it say a triangle? How would you be able to predict whether this would invert or not? Yeah. I'm not sure the... Maybe rephrasing your question, I might say that like, what is the... stable equilibrium of this industry. I think it is pretty clear that AI is unlikely to be a fad. It is unlikely to be an unsuccessful endeavor. And I think about the stable equilibrium of this chart quite a bit. In fact, I got into a little bit of a debate on Twitter with somebody quite smart and thought a lot about this, about this exact question of like, what is the stable equilibrium and You know, my guess is that it might stay this way for longer than I anticipated. In the cloud, I think that range is about a decade. I have a feeling this might stay longer, this way longer because of just how hard it is to get the substrate right But there will be one or two unlocks. I couldn't tell you what they are. But for example, if one of the ASIC programs at one of the hyperscalers-- be it Google's TPU or Meta's MTIA or, you know, the folks at Amazon and OpenAI and Microsoft and all the labs that we don't even know about exist. Um, has a breakout has breakout success i suspect that'll be the biggest repricing of that layer. Thank you. The other catalyst could be You know, I think about the hyperscaler CapEx guidance and earnings calls, which, by the way, I recommend everybody here to listen to four times a year. you'll have public company CEOs tell you they're biggest questions, their biggest things that they're thinking about and I recommend listening to those. if they stopped guiding to big numbers on CapEx, because that would imply that the current equilibrium does not work. So that's probably the second thing that could happen. And so you see there's a lot of news about the guidance that all the hyperscalers give about their capex for that reason. Go ahead. Is there also an algorithm of changing versus influence? because my sense is if, The only way this gives is if infants are meaningfully larger than training, Hi. and i'm curious to hear your thoughts on the very end you think that will happen because then you're seeing that and it needs to stop spending on training because they're not seeing what is the point of view of meaningfully and hence want to have a non-influence. It's a great question, and it is probably one of the nuggets of information that NVIDIA's earnings calls have, like the most sought-after nugget of information. What is NVIDIA's share of inference in their fleet? Last I checked it was about 40%, or they coded it to be about 40%, meaning that If they were selling a million GPUs, assuming full utilization, about 40% of them were used for inference and the other 60% for training. I suspect that number will increase over time in favor of inference. But I couldn't tell you when and how it'll happen because there's a lot of training still going on in the world. and the shape of the training workload as you know looks very different from the shape of the inference workflow workload. A training workload is very predictable, high utilization for a short period of time. The inference workload is very burst-like. usage, typically when humans are awake, until the agents take over. Maybe then it'll be 24/7. And harder to predict. It goes down around Christmas for some reason. It goes down around Thanksgiving for some reason. But I think that might be the case, though we At least in this calculation, we try to capture it because it's a mix. But it's a good hypothesis. Thank you. - I've seen the revenue, but where would probability be? - It's on slide 16. We'll come to it. We'll come to it. I'll give you the answer. The most profitable part of the stack is the same easier. by a long shot.
[20:00] NVIDIA's data center revenues are the most margin of about 75%. Don't quote me on it. It's like plus minus a couple percentage points from there. Whereas I estimate some of the application layer revenues to be somewhere between, depending on who you ask, like between 0% and 30%. And so, The gap is quite wide. And I think the reason to that, I mean, it's a theory that gentlemen here had is, there's one player who kind of runs the tables on the semi-slayer. And so it's very much the case. And in fact, if you looked at this from a profitability perspective, it's even more concentrated. the triangle is even more concentrated. um i'll i'll i'll flash that in a second when we when we get to it that's a great question Go ahead. Do you think this year is going to show the BI? Yeah, that is definitely a big part of it, is that we've gone through the investment cycle in cloud. It's definitely an element to it. Go ahead. - If all these infra companies like Google, AWS, and all their own TPUs, and you guys also do it for your most, OpenAI is also searching for some basics. Where are those of all these eight states different startups in the Seltz? So I'm just going to go on to There's $300 billion of revenue to fight about. But to answer your question, about half of that, as Jensen discloses on the earnings calls, is from the big hyperscalers. So those are probably going to be your primary customers. So if you were starting a ship company today, you would have a very, the shape of your customer base is a very small number of very large orders. Um, It's a very different shape from building a consumer business or an enterprise software business. And then you might have a long tail of other enterprises, though I wouldn't bank on it, because I think they just go to the cloud providers. If you were thinking about starting a chip company, it should be your number one consideration. Which of the five are you going to Sell to first. Thank you. Last question. I was able, you get small and full of miners in each of these layers, it takes multiple years for that to fray out. I feel like in the past, there hasn't been a fully vertically integrated one. I think that Google is fully vertically integrated on the right hand side. We're wondering how that shifts the value of power this time. What a great question. The biggest winner on the internet super cycle is probably Google. It's about 3 trillion in market cap, has near 99% market share in search. I would say that that's a pretty vertically integrated player, right? They're on their own file server to search.