Nvidia Deepdive

SummaryTariffs are the main headline these days, but they are just a catalyst to tighten/reverse/raise the cost of capital and tip the real risk in the market – the AI infrastructure ecosystem – into a negative reflexivity tailspin. I think if you look back at any bear market, the “real risk” at the time was where the leverage / reflexivity was. I believe that in this cycle, the leverage / “real risk” is in the AI infrastructure ecosystem. I believe this segment of the market, as evidenced by the massive valuations in big cap tech and extreme concentration in the stock market, has been in the boom part of a classic boom/bust cycle, akin to Internet / telecom infrastructure in ~2000. <br><br>Apple CAPMApple Forward P/EApple Gross MarginApple Net MarginApple ROE Return on EquityApple ROA Return on AssetApple Graham NumberApple Current RatioClassic signs of a boom/bust cycle are present: high capital intensity, external funding, lack of focus on commercial viability, KPIs that have nothing to do with profit, and sketchy behavior by corporate executives at key companies. NVDA is the posterchild for the AI infrastructure boom and lies at the heart of the reflexivity and leverage in the ecosystem. The recent market weakness, catalyzed by tariffs, as well as AI-related headlines like DeepSeek, have tipped the AI infrastructure ecosystem firmly into the bust part of the cycle. Over-ordering, lengthening lead-times, rising prices, accelerating revenues and profit margins, accelerating valuation multiples and stock prices, and the resulting larger and richer capital raises of privately-funded AI companies are in the process of reversing and becoming self-reinforcing on the way down as they were on the way up. Short NVDA. ContextI think that AI will very likely revolutionize many aspects of our lives over the next several decades, but I also think we’re at a pivotal moment in the AI capital cycle. I do not believe that most of the long-term value creation resulting from AI will be in the companies that have attracted the most capital to-date - similar to how capital investment and value creation played out in the telecom / Internet cycle. Despite my belief that AI will likely set the stage for significant value creation in the future, I believe we’re at a pivotal moment today and that many of the perceived beneficiaries of AI are poised for massive declines in their stock prices. I think that history provides great context for how we got here and how things will play out - in particular, both the long-term characteristics and performance of the overall stock market, as well as the history of the Internet bubble. In terms of the long-term characteristics and performance of the stock market, I’d like to set the stage with two charts below courtesy of Morgan Stanley and Yardeni Research. The first is a historical chart of stock market concentration in the U.S., and the second is rolling 10yr returns of the S&amp;P 500. <br><br>What you see is three coincident peaks, suggesting that market cycle returns are generally driven by the largest stocks in the market. In other words, exuberance in a subsegment of the market drives overall market returns, until that exuberance reverses, and causes returns to reverse. Leverage is great on the way up and hurts on the way down<br> I think it’s important to step back and ask what the markets and press are focused on today as a benchmark of AI success. In the Internet bubble it was “eyeballs” despite the fact that this KPI was in no way related to a financial return on all the capital investment being made by telecom companies. Similarly, today there is very little attention paid to, let alone concrete indication of, ROI on AI models. Also similarly, there is a new “eyeballs” driving capital raising: model performance. Just as Worldcom pushed “Internet traffic doubles every 100 days” in the 1990s, model companies today push scaling laws to justify additional capital investment. Here’s Dylan Patel, well-respected author of SemiAnalysis, on a recent podcast with Lex Fridman (&gt;5hr listen, but a good one - https://lexfridman.com/deepseek-dylan-patel-nathan-lambert/): “I think only Nvidia is actually making tons of money and other hardware vendors, the hyperscalers are all on paper making money, but in reality they’re spending a lot more on purchasing the GPUs, which you don’t know if they’re still going to make this much money on each GPU in two years, right? You don’t know if all of a sudden OpenAI goes kapoof and now Microsoft has hundreds of thousands of GPUs they were renting to OpenAI that they paid for themselves with their investment in them that no longer have a customer. This is always a possibility. I don’t believe that. I think OpenAI will keep raising money. I think others will keep raising money because the returns from it are going to be eventually huge once we have AGI.” Let’s leave aside that these same experts not only acknowledge that “AGI” is anyone’s definition, that it is likely 2030 at the earliest, and that it will come at a massive expense with no clear-cut commercial model. <br><br>Amazon CAPMAmazon Forward P/EAmazon Gross MarginAmazon Net MarginAmazon ROE Return on EquityAmazon ROA Return on AssetAmazon Graham NumberAmazon Current Ratio<br>The main point is that, as in the Internet bubble, only the equipment providers are making money – that’s not sustainable. Here’s a quote from a recent FT article (https://on.ft.com/422tW4I) titled “Who needs revenue when you’re a multi-billion-dollar AI start-up?” (also reminds one of this great scene from Silicon Valley - https://www.youtube.com/watch?v=BzAdXyPYKQo). “Last month, Thinking Machines Lab, an AI “research and product” company launched by OpenAI’s former chief technology officer Mira Murati, was reported to be seeking $1bn in funding at a $9bn valuation. Assessing that on traditional metrics such as a multiple of revenue is impossible. Not only does Thinking Machines Lab generate no revenue, it has yet to specify what it might sell. Murati’s former colleague Ilya Sutskever, ex-chief scientist at OpenAI, is going one further. His pre-revenue, pre-product AI company Safe Superintelligence is in talks to raise funds at a $30bn valuation.” In the Internet bubble, eyeballs did not end up being correlated with the capital cycle; Internet traffic grew exponentially and unabated throughout the bust. Furthermore, while Jevons paradox ended up holding for the Internet in the long-run, in the short run, 97% of fiber that was laid remained unlit for a period of time and as a result capital investment plummeted and as we remember, related stock prices declined by epic amounts. I believe this happened because a) capital investment was focused on demand drivers that weren’t associated with economic value, and b) use of bandwidth got more efficient. I think the same thing is happening today with a) capital investment focused on model performance and not monetization / profitability, and b) models are getting more efficient (DeepSeek, etc.). Reflexivity in AIGeorge Soros’s theory of reflexivity essentially states that there is a feedback loop between expectations, prices and fundamentals. I believe that this feedback loop is amplified / turbocharged when the ecosystem is not self-funding. Soros and Stan Druckenmiller have made fortunes identifying and trading both sides of the boom-bust cycles that are a function of the theory of reflexivity. I believe that there is a significant amount of reflexivity in the AI infrastructure capital cycle. To frame this concept, I think it’s helpful to break the model developers into two buckets: what I’ll call the Self Funders, and the YOLOs (because they have been raising, spending and losing money like there is no tomorrow). The Self Funders are the likes of Meta, Google, Microsoft (xAI is debatable, I suppose) – large entities with very successful and cash generative legacy / core businesses. While these businesses are self-funding, they’re still investing against hope and vision, not clear-cut ROI. This is ok in the boardroom and investor meetings and calls while the reflexivity is positive, but once it’s negative, the calculators come out and sober financial analysis is brought back into the discussion. The YOLOs are the likes of OpenAI, Anthropic, Mistral, Cohere, etc. Not only are these companies losing billions of dollars, but they don’t have a path to profitability either. As the AI hype has built, they have raised more and more capital at richer and richer valuations, spending and losing more and more money. Driving this positive reflexivity has been a combination of FOMO and a focus on KPIs that have nothing to do with a path to profitability. These companies are the leverage in the system. They are not sustainable in their current, or any near/medium term form, and they are likely to disappear / get rescue funding / become acqui-hires for the Self Funders. As that bust plays out, not only will these companies stop ordering GPUs, but the GPUs that they own or rent will become available. Furthermore, the acceleration of the capital cycle for these models is amplified downstream to their capital equipment providers – the data center, networking, electrical, cooling companies, and of course NVDA. As the model companies raise money, they recognize they are in a race and they invest first and ask questions later. Lead times go out, prices go up, margins expand, stock prices go up, more money is raised…the self-reinforcing good times roll. As this has happened, not only have revenues accelerated for the capital equipment companies and NVDA, but margins have expanded and so have valuation multiples. Leverage is so nice on the way up. The “DeepSeek moment” is real, I believe – I think we’ll look back and recognize that moment as when the reflexivity in the AI infrastructure boom/bust cycle went from positive to negative. Since then you’ve had MSFT making negative comments re: capex, you’ve had intense skepticism around a very sketchy-looking CoreWeave IPO, and now you’re having follow-on DeepSeek-like announcements with Manus and likely another DeepSeek release imminent (https://www.reuters.com/technology/artificial-intelligence/deepseek-rushes-launch-new-ai-model-china-goes-all-2025-02-25/#:~:text=Deepseek%20had%20planned%20to%20release,reason%20in%20languages%20beyond%20English). And now you’ve got tightening financial conditions across the entire capital markets ecosystem as tariff concerns abound. As a result, stock prices are down, valuation comp tables are looking tougher, and capital raising prospects are looking dicey. The bust part of the cycle has started. Why NVDA?My recommendation is to short NVDA, however I think that many companies across the industrial, power and networking landscape will also be subject to significant corrections in their stock prices as the capital cycle in AI infrastructure investment reverses. I’m focused on NVDA because it’s big, it’s liquid, I think it is easily recognizable as the Cisco of its era, and it is a case study in reflexivity within the context of a boom/bust cycle. NVIDIA was instrumental in kickstarting the positive reflexivity flywheel – they’ve funded many of the YOLO models, CoreWeave, etc. - money-losing customers who took NVDA’s cash, turned around and bid up the frenzy for NVDA’s own GPUs. Jensen, strutting in his black leather jacket and making claims like “software is eating the world, but AI is going to eat software” has only turned up the heat and lathered up the investment community. Not only that, but it looks like they may have knowingly allowed companies like Super Micro (Jensen and SMCI’s Charles Liang are well-known friends - https://fortune.com/2024/06/04/super-micro-computer-fortune-500-debut-data-center-servers-training-ai-models-nvidia-chips/) to smuggle chips into China – a quote from Dylan Patel of SemiAnalysis on the aforementioned Lex Fridman podcast: “I saw a photo from someone in the semiconductor industry who leads a team for networking chips that competes with NVIDIA, and he sent a photo of a guy checking into a first class United flight from San Francisco to Shanghai or Shenzhen with a super micro box that was this big, which can only contain GPUs, right?”. I believe two factors are going to drive materially negative profit revisions for NVDA in the short to medium term: 1) volumes and 2) ASPs and margins. On volumes, I think most if not all of the YOLO model companies will cease to exist. The podcast I just mentioned was super bullish re: the potential of AI but was also full of cognitive dissonance including acknowledgement of the amount of money that companies are currently losing, the lack of commercially viable products, and the fact that models are getting better and cheaper. Towards the end of the podcast Lex asks: “But X, Google, and Meta have these other products. So isn’t it likely that OpenAI and Anthropic disappear eventually?” I believe the answer is yes. I think Sam Altman will end up being chastised for not hitting the bid on Elon’s $97.4b. I think that lower stock prices will lead to down rounds and eventually rescue funding / acqui-hires by the self-funding models of Anthropic, OpenAI, Cohere, Mistral, etc. Not only will the volumes from these models go away / decline materially during this process, but it will also take the heat off of the self-funding models and cause them, encouraged in addition by more judicious capital markets, to be more discerning with their spend. On prices, I think that ASICs and other new chips will pressure ASPs in addition to volumes. Amazon is developing its new Trainium chip supposedly for $4k ASP (https://www.theinformation.com/articles/amazon-undercuts-nvidia-aggressive-ai-chip-discounts). Clearly Trainium and Blackwell chips have different characteristics and capabilities, but this dynamic, along with lower demand, will put intense pressure on NVDA’s prices and margins. Google is developing cheaper chips - https://www.theinformation.com/articles/google-taps-mediatek-cheaper-ai-chips?Fds-Load-Behavior=force-external. Startups like Groq are developing alternatives to NVDA -With lower volumes and some ASP pressure towards ASIC economics, earnings at NVDA could fall materially. Multiples typically compress with earnings revisions, likely accelerating the downside. NVDA shares could fall another 60% and it would still be a trillion dollar market cap company. On a related note, here is Stan Druckenmiller from an interview with NBIM’s Nicolai Tangen last November: “I did an interview on Nvidia, I think it was like 370 or something, and I said, this is one we're probably going to own for a few years. But I didn't think it was going to go to 900 in a year, and to over a $2 trillion market cap, I think it just started like 100 billion or 150 billion. It was something crazy. So no, I don't necessarily sell early. I'm a technician, so I usually wait for tops. Nvidia had no top. A top is something the rate of change of it's going up changes and it tends to flatten out for quite some time. The trick is in the technical world, that can end up being a bull flag where it's just consolidated for a bit and then it's hit a new leg or it could be a top where that was it. And how do you know which is which? You don't. You have an opinion and you express it and sometimes you're right and sometimes you're wrong. With Nvidia, there was no top, but I just, I've analyzed the semiconductors industry, not particularly well, but since the 1970s, and it's a cyclical industry. And I knew Nvidia had staying power and they had 4,000 software engineers, so it wasn't just hardware. You know, they have a CUDA, this thing called CUDA software that they do to make their GPUs. But I just thought once it went through $2 trillion, this is just too much. And worst case, it'll have a big correction, I'll get another chance.” Why does Stan call himself a technician and look for tops? I believe it’s because he’s looking for the switch from positive to negative reflexivity. His comments re: the cyclicality in the industry suggests he recognizes the potential for a boom/bust cycle here. Not to mention these two quotes, which I think support the framework I’m putting forward: “We've never had a bear market start without the leadership narrowing, and it's narrowing enough that you're starting to get toward a necessary condition being satisfied. But it's early, but it's a yellow light, it's not a red light. That's how I read it. The AI boom is going unabated, Nikolai. I think the private sector just sees it as an existential threat to their business if they don't spend money on it, because if they don't spend money on it because if they don't spend money on it and their competitors do and their competitors are right, they're going to have a big, big competitive problem.” “we started with picks and shovels, which is Nvidia, and to some extent Microsoft, but now we are seeing just massive amounts of capital being spent by these modelers. And if AI is for real and I think it is, they are all going to give you the same answer, so we are going to have four or five companies will spend massive amounts of capital, but I don't see it as a winner-take-all model. On the other hand, I think there are applications that I haven't even thought of, and nobody has thought of that are going to spring up. I mean, who would have thought of Uber or Facebook when the internet started. So we're very bullish on AI, but we're not bullish currently on exactly where we're supposed to be and how to play it aggressively. Not unlike the internet in 2000, 2001, you could have believed in the internet not been exposed and then got your exposure on a more timely basis.” Interestingly, from the Lex Fridman podcast I referenced before – “I don’t think it’s winner take all”. I think both Stan’s comments and the Lex Fridman podcast strongly support the idea of a reflexive, boom/bust cycle in AI infrastructure, eventually leading the way to value creation at the application layer. Altimeter Capital Partner Apoorv Agrawal is more succinct and articulate than I am on this thought process - https://www.tiktok.com/@cnbc/video/7468806821243964718. <br>best stock research sitesbest stock research sitesWhile that value creation at the application layer is an exciting prospect, in the meantime, just as with fiber, we could have a temporary glut of GPUs, Jensen Huang could end up looking like the Jack Welch of tech – having artificially (no pun intended!) inflated demand by financing unsustainable customers - and NVDA could go down another &gt;50%. ,