Michael Burry Is Short Nvidia. He's Not Short AI — He's Short the Crowd.

Michael Burry bought puts on Nvidia. The same man who shorted subprime mortgages before the world fell apart. The same man who saw the housing bubble when everyone was buying second condos with no income docs. Now he's standing across from the loudest trade on the planet. And the crowd? The crowd is still chasing the alpha before the liquidity dries up. Scion Asset Management filed its 13F with the SEC, and there it was: a fresh batch of put options on NVDA. Not small. Not symbolic. The Big Short guy pointing his finger at the AI darling. The internet did its usual thing — retweets, outrage, hot takes, and a thousand keyboard analysts calling him a boomer who doesn't understand exponential growth. But here's the clue everyone missed: a blockchain media outlet picked up the story and published a deep technical analysis of Nvidia's semiconductor supply chain. Wait, since when do crypto desks cover chip fabs? Since Nvidia became the ultimate cross-collateral asset. A trade that touches AI, crypto mining, data center capex, GPU scarcity, and the entire global liquidity matrix. Once a symbol becomes that big, technical analysis loses to flows. Price action stops being about the company and starts being about the crowd. That's exactly where Michael Burry likes to strike. Let's back up and build the context properly. Nvidia is a fabless semiconductor company. It designs some of the most advanced AI accelerators on Earth, but it doesn't manufacture a single wafer. TSMC makes the chips. SK Hynix, Samsung, and Micron make the HBM memory. ASE and Amkor handle some of the packaging. Nvidia sits at the top of the value chain, collecting gross margins north of 70%, because it owns the architecture, the software stack, and the market narrative. That's a beautiful business. It's also a fragile one. Burry's short isn't a bet against the technology. It can't be. The technology is winning. Every hyperscaler on the planet is waving billions of dollars at Nvidia, trying to secure enough H100s, B200s, and GB200s to train the next generation of models. The company's data center revenue is growing at triple-digit rates. The demand curve is real. The bottleneck isn't a lack of customers; it's a lack of manufacturing capacity. So why short it? Because the market is paying a price that assumes perfection, and perfection doesn't exist in semiconductor supply chains. I've covered enough launch events, token sales, and liquidity traps to know that the moment a trade becomes a consensus, the floor becomes a rumor. Hype is the fuel, but fundamentals are the engine. Let's get into the technical core. The first thing you need to understand is that Nvidia's manufacturing moat isn't really about process nodes. It's about system-level integration. The current AI workhorse, the H100, uses TSMC's 4N process, which is basically a customized 5-nanometer node. The Blackwell generation, including the B200 and GB200, uses TSMC's 4NP, another refined version of the same family. Those chips are still FinFET transistors. TSMC doesn't move to gate-all-around, or GAA, until the N2 node. So Nvidia is never the leader in raw process technology. TSMC is the leader. Nvidia is just the best customer. That's a critical distinction that gets lost in the headlines. If you're buying NVDA as a proxy for 'the most advanced chips in the world,' you're buying the wrong layer. The real power sits with the foundry and the memory makers. Nvidia's value comes from pulling all those pieces together into a cohesive system, then wrapping it in CUDA so every engineer in the world knows how to write code for it. That doesn't sound like a short thesis. But here's where it gets messy. The thing that makes Nvidia invincible on paper — the software ecosystem, the network interconnects, the CoWoS packaging, the HBM allocations — is also the thing that creates severe concentration risk. Nvidia is the largest buyer of TSMC's CoWoS advanced packaging capacity. That's a fancy way of saying the real bottleneck isn't the GPU die. It's the packaging that stacks memory next to the processor. TSMC has been scrambling to add CoWoS capacity for years, and every AI chipmaker on Earth wants a piece of it. If TSMC's CoWoS expansion slips, Nvidia's revenue slips. Not because demand isn't there, but because physical production can't keep up. Then there's HBM. High Bandwidth Memory is the other silent choke point. SK Hynix, Samsung, and Micron essentially control the entire supply of HBM3e and HBM4. Nvidia has long-term agreements and pays premium prices, but it doesn't control the production line. In a bull market, that's fine. In a down cycle, that's a cost disaster. And if any one of those memory suppliers has a yield problem — which has happened before — Nvidia's accelerators become paperweights waiting for memory. I can already hear the rebuttal: 'But Nvidia has pricing power. It can pass the costs through.' Sure. In a shortage, scarcity pricing works. But the stock already prices in that scarcity. The market has given Nvidia a valuation that assumes the scarcity lasts forever, the CUDA moat never erodes, and the hyperscalers never blink. That's not a technical analysis. That's a religious belief. Now let's talk about the competitive landscape, because this is where the contrarian angle gets sharp. Nvidia's biggest threat isn't AMD, though AMD's MI300 is legitimate. It isn't Google's TPU, though TPU has been running production workloads for years. The real threat is the hyperscalers themselves. Microsoft has Maia. Amazon has Trainium. Google has TPU. Meta has its own MTIA. These are custom ASICs designed specifically to handle the workloads that matter to their own clouds. They don't need to beat Nvidia on every benchmark. They need to be good enough and cheap enough. And the moment Nvidia's price per GPU starts to feel like a tax, those custom chips become more attractive. That's the part the crowd refuses to see. The same logic that made Nvidia dominant — the GPU's flexibility for training — becomes a liability in the inference era. Training is the hard, glamorous work. Inference is the massive, repetitive, ongoing work that happens every time a model responds to a prompt. Once AI models stabilize, a huge portion of computing shifts to inference. And inference is exactly where custom silicon and ASICs can be far more cost-efficient than a general-purpose GPU. Nvidia will retain a huge slice of the market, but the monopoly-style margins will erode as the workload matures. That's the slow kill. Speed kills, but slow kills too in this game. I've watched this movie before. In 2017, I led a rapid-response team covering the ICO boom. We chased every token from zero to moon and back to zero. Every project had a white paper, a Telegram group, and a hero narrative. The ones that survived were the ones with real protocols. The ones that died were the ones that traded like they were already worth a billion dollars before they had a product. Nvidia has a product — an incredible product. But the stock is not the company. The stock is a claim on future cash flows, and the future cash flows are already priced as if the next ten years will be exactly like the last two. Burry isn't betting on a business collapse. He's betting on a repricing. He's saying the current price embeds too much certainty, and certainty is exactly what an AI-accelerated world doesn't have. Let's go deeper into the supply chain. The source analysis I read pointed out something that most semiconductor coverage misses: Nvidia doesn't directly control materials, equipment, or manufacturing. It relies on TSMC for advanced lithography, CoWoS packaging, and many other process steps. It relies on SK Hynix and Samsung for HBM. It relies on Synopsys and Cadence for EDA software. It relies on Arm for CPU architectures. None of that is an insult to Nvidia — it's just the reality of the fabless model. But it means Nvidia's 'moat' is actually a series of alliances, not absolute ownership. If TSMC has a bad quarter, Nvidia has a bad quarter. If CoWoS capacity doesn't expand fast enough, Nvidia ships fewer accelerators. If HBM4 experiences yield challenges at SK Hynix, Nvidia's roadmap slips. And the more complex the chip, the more ways it can break. The Blackwell B200 is a massive product with two GPU dies bridged together. That's twice the silicon area, twice the thermal stress, twice the packaging complexity. Every step multiplies the chance of something going wrong. Now, the bullish rebuttal is always: 'TSMC and Nvidia have been navigating this for years, they know what they're doing.' True. But the market is not trading 'they know what they're doing.' The market is trading 'nothing will ever go wrong again,' which is a very different statement. Here's the hidden layer that a blockchain journalist is uniquely positioned to see: Nvidia has become the crypto trade of this cycle. I'm not talking about GPU mining. I'm talking about the psychology. The way retail and institutional investors talk about Nvidia — the FOMO, the 'blue chip' label, the 'you're either with AI or against it' framing — it's identical to the way people talked about Bitcoin in late 2021 and NFTs at the top. The crowd moves fast, but the ledger moves faster. In that sense, Burry's put position is a liquidity trade disguised as a semiconductor short. When the global liquidity tide goes out, high-multiple assets get crushed regardless of earnings. And Nvidia is no longer just a semiconductor stock. It's the poster child for the AI bubble trade, the index heavyweight, the margin portfolio's favorite tech name. Everyone who has made money in the last eighteen months has made it by being long Nvidia or something correlated to Nvidia. When a trade is that crowded, the exit door has a very limited width. I've seen the moon, now I'm looking for the exit. Let's turn to the geopolitical angle, because it strengthens the contrarian case. Nvidia is caught in a two-track world. On one track, it sells to the West — Microsoft, Meta, Google, Amazon, Oracle — and the AI arms race continues. On the other track, US export controls restrict the most advanced chips from going to China. That creates a vacuum in the Chinese market, and Chinese companies like Huawei and Cambricon are stepping in with their own accelerators. They're not as powerful as Nvidia's best, but in a controlled market, they don't need to be. They just need to be available. Every environment that closes to Nvidia is a training ground for its competitor. This is not a near-term doom story. Nvidia will continue to post spectacular numbers. But the two-track world means Nvidia's addressable market is smaller than it would have been, and its competitors are getting home-field advantage in the world's second-largest economy. That's a structural drag that isn't reflected in the current multiple. The other thing no one is talking about is the capex cycle. The AI boom is being financed by an extraordinarily few number of companies. Microsoft, Meta, Alphabet, Amazon, and Oracle are spending hundreds of billions of dollars on data centers and chips. They're doing it because they believe AI will drive future returns. But if one of those giants misses earnings, or if the ROI on AI capex takes longer than expected, the boards will start asking questions. And when bean counters get involved, the most expensive part of the budget gets scrutinized first. That's Nvidia. I've seen this exact pattern in DeFi Summer 2020. Lenders were handing out liquidity with no collateral. Everyone felt rich because yields were absurd. Then the leverage got rattled, and the floor fell out from underneath the protocols that had the best narratives. The fundamentals didn't matter in the moment — the liquidity retreat mattered more. Where the yield is sweet, the risk is steep. Let me be clear: I'm not calling Nvidia a Ponzi. Nvidia is a real company with real revenue, real profits, and real technological leadership. It's not a meme token. But the stock's current price embeds a very specific prediction about the future — that AI capex keeps accelerating, that HBM and CoWoS bottlenecks get resolved, that hyperscalers don't integrate their own silicon, and that China stays locked out without meaningful consequences. That's a lot of things that have to go right. Michael Burry's short is a statement about margin of safety. He doesn't need Nvidia to collapse. He needs the price to revert to something closer to the historical norm for a hardware company, even one as extraordinary as this. The market has been arguing with him for months. The stock has rallied since his position was disclosed. But in a frothy environment, being early looks like being wrong — until suddenly it looks like being prescient. The real danger isn't Nvidia's technology. It's the belief that the technology makes the graph exponential forever. That's the same belief that drove tulip prices, internet stocks, ICOs, DeFi tokens, NFTs, and every other mania that ended with a burst. Hype is the fuel, but fundamentals are the engine. And when the engine begins to sputter, the fuel doesn't matter. I've been through enough cycles to know this: the winner isn't the person who gets in early. It's the person who knows when the crowd is too crowded. Michael Burry has made a career of that exact instinct. He shorted the housing market when everyone said housing never falls. He has been right about more things than most people want to admit. And now he's looking at Nvidia the way he looked at banks in 2007 — a position that makes the entire market call him wrong, until the entire market discovers he was right. Here's what I'm watching next. The next round of hyperscaler earnings is the first big catalyst. If any of the big four mention 'AI capex normalization' or 'efficiency investments in our own silicon,' that's the first crack in the narrative. Second, I'm watching TSMC's monthly revenue and CoWoS capacity announcements. If packaging capacity grows significantly faster than expected, the scarcity premium on Nvidia's chips will erode. Third, HBM pricing is the canary. If HBM prices flatten or drop, it means the memory shortage is over, and the urgency to buy Nvidia systems will cool. None of those data points require a recession. They just require a slowdown in the rate of increase. That's all it takes for a stock trading at 35 times forward earnings to compress to 25 times. And when a stock compresses from that high a base, the drawdown is brutal. The crowd will keep buying the dips. They'll keep saying 'this time is different' and 'Burry is early.' They'll keep looking at the massive data center numbers and ignoring the concentration risk. But I've watched enough floors drop in this industry to know that every market has a moment where the narrative stops being enough. Nvidia is not a token. It's not a fake protocol. But in a bull market, everything trades like a token. And in a bull market, everyone is a genius until the liquidity dries up. So here's the takeaway. This isn't a short thesis on AI. It's a short thesis on consensus. If you believe AI is real, then Nvidia is a genuinely important company. But the stock has already priced in that belief — plus a hundred percent of the upside, plus a premium for certainty, plus a premium for the crowd's inability to imagine a different future. Michael Burry is betting that the premium will evaporate. And in this market, where speed kills and slow kills too, the only way to win is to know whether you're holding the asset or holding the narrative. The crowd moves fast, but the ledger moves faster. I've seen the moon, now I'm looking for the exit.