Mispriced

Jensen Huang: this is a "technological miracle"  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌

March 12th, 2026   Donald Doge In 2024, Jensen Huang stood on stage at NVIDIA's GTC conference, held up a chip smaller than his palm, and called it a "technological miracle." He wasn't talking about a GPU. He was talking about a memory chip. "Without the HBM memory," he said, "there is no AI supercomputer." Nobody clapped like they understood what that meant. The crowd wanted processing power. The next GPU. Models are getting smarter. Memory chips? That's plumbing. That's boring. Except it's not boring anymore. Right now, the global supply of memory chips is collapsing. Supplier inventories have cratered to three-to-six-week levels, down from the healthy ten-to-twelve-week norm. DRAM contract prices surged 45%-50% last quarter. This quarter, they're projected to jump another 90%-100%. And the AI infrastructure buildout driving it all is still in its early stages. This is the worst memory-chip shortage in roughly a quarter-century. And the downstream consequences, for the phone in your pocket, the laptop on your desk, and the entire trajectory of artificial intelligence, are just beginning to appear. Jensen wasn't hyping. He was warning. Most people just weren't listening. The Wall Nobody Sees If you've been reading me for any length of time, you know I've been hammering one theme: the AI revolution has a physical problem. I've written about the energy crisis. The grid can't keep up. I've written about the critical minerals. Adversaries control the metals that make chips possible. Every time, the market acts surprised. Every time, the pattern is the same: Wall Street obsesses over software and models while ignoring the physical world those models depend on. Memory is the latest chapter. And it might be the most urgent one yet. Engineers call it the "memory wall." Large language models, the engines behind ChatGPT, Claude, Gemini, and every other AI system racing to reshape the world, aren't just constrained by how fast they can think. They're constrained by how fast data can move between chips. Picture the most powerful engine ever built. Now connect it to a fuel line the width of a coffee straw. That's your AI supercomputer without high-bandwidth memory. High-bandwidth memory chips, HBM, are the fix. They stack DRAM chips directly next to the GPU using advanced 3D-packaging technology, creating a wider fuel line. Demand for them is exploding. TrendForce estimates HBM demand will increase 70% year-over-year in 2026, after rising 130% last year. The total addressable market is forecast to grow from roughly $35 billion in 2025 to approximately $100 billion by 2028, a compound annual growth rate of about 40%. Those numbers sound like opportunity. They are. But they're also the source of a crisis already rippling into your daily life. The Squeeze You're Already Paying For HBM chips consume roughly three times as much silicon per gigabyte as standard DRAM. Every time a manufacturer shifts production capacity to HBM, which they must because that's where the money and the demand are, they pull capacity away from the conventional memory chips that go into everything else. Your phone. Your laptop. Your car's infotainment system. Your kid's tablet. HBM production is set to account for 23% of the global DRAM manufacturing base this year, up from 19% last year. Data centers overall are projected to consume 70% of all memory chips produced globally in 2026. That leaves 30% for the rest of the world. The fallout is already showing up on price tags. International Data Corporation projects a 12.9% decline in global smartphone shipments this year, down to 1.12 billion units. The average selling price is expected to rise 14% to a record $523. The sub-$100 smartphone segment, 171 million units shipped last year, could become permanently uneconomical. The PC market faces the same squeeze. Global shipments are expected to decline 11.3% this year. Gartner expects the entry-level PC segment, anything below $500, to effectively disappear by 2028. The cheapest computers and phones are being priced out of existence because AI is gobbling up all the memory. Growing up in a lower-middle-class household, I know exactly what it means when the affordable option disappears. It means falling behind gets easier and catching up gets harder. The AI revolution that's supposed to democratize intelligence is quietly making the basic tools of modern life more expensive for everyone. 🚀 Get Moonshot Alerts Instantly Markets move in minutes. Your inbox can't keep up. Join the Moonshot VIP Text Service and get real-time alerts, portfolio updates, and first-move intelligence directly to your phone. Be first to know. Anywhere. 👉 [Join the Text List →] A Gap That's Accelerating, Not Closing Transformer-model sizes, the architecture behind every major AI system, have grown roughly 240x every two years. Each successive generation of HBM delivers only about 2x bandwidth gains over a similar period. That gap will only grow bigger. A single NVIDIA Blackwell B200 GPU contains 192 GB of HBM3e, up to 12x more memory than an entire traditional server. A modern GPU-server rack like NVIDIA's NVL72 consumes more memory than approximately 1,000 high-end smartphones or several hundred PCs, and that's just one rack. And inference, the process of actually using AI models in real time, the thing that happens every time you ask ChatGPT a question, is becoming the dominant workload. Lambda's 2025 "AI Wrapped" report found that average consumption of reasoning tokens increased 320x year-over-year. McKinsey projects inference will account for more than half of all AI compute by 2030. Inference is hungrier for memory than training. At standard DDR5 speeds, a single inference pass takes 7.5 seconds, completely unusable for real-time applications. At HBM3e speeds, the same pass takes milliseconds. Add AI agents, longer contextual windows, and multi-step reasoning chains, and memory requirements compound exponentially. The industry is in a structural deficit. And every new AI model, every new agent, every new reasoning chain makes it worse. 🚀 Get Moonshot Alerts Instantly Markets move in minutes. Your inbox can't keep up. Join the Moonshot VIP Text Service and get real-time alerts, portfolio updates, and first-move intelligence directly to your phone. Be first to know. Anywhere. 👉 [Join the Text List →] The Chokepoint Nobody's Pricing In If this all sounds familiar, it should. It's the same pattern I've been writing about with critical minerals, just one layer deeper in the stack. It's the same pattern that's been handing us double and triple-digit gains in the Moonshot Minute Portfolio. Memory chips don't just need silicon. They need gallium and germanium to tune how quickly and efficiently circuits switch on and off. They need rare-earth elements for specialized components. And increasingly, they need scandium, one of the smallest and most strategic metals markets in the world, with global production of only about 60 tons per year. And who controls the supply of these critical minerals? Anyone following my writing already knows the answer: China. Since 2022, Beijing has imposed 29 rounds of export license restrictions on critical minerals. Heavy rare-earths, gallium, germanium, scandium, the very materials that memory-chip production depends on, are subject to tight controls. So the constraints I've been writing about aren't separate stories. They're layers of the same story. Energy limits where you can build data centers. Minerals limit what chips you can make. And now memory limits how fast AI can actually run. Each one compounds the others. The physical world doesn't care about software roadmaps. I've said it before and I'll keep saying it: the future isn't built on code alone. It's built on dirt. On rocks. On the physical materials that most investors are too distracted to notice. The memory-chip crisis is the latest proof. Where the Money Is Moving The memory-chip sector's revenue has risen ninefold in four years. It now makes up 55%-58% of the global semiconductor market, up from 25%-30% historically. TrendForce projects global memory-market revenue will reach $551.6 billion in 2026, up 134% year-on-year, and peak at $842.7 billion in 2027. While capital chases hyperscalers, GPU producers, and data-center builders, memory may quietly hold the best pricing power in the AI stack. The HBM market is dominated by a handful of players. SK Hynix leads with about 62% of shipments. Samsung holds a strategic advantage in conventional DRAM. Their 2026 production is contracted and sold out. But there's one company in this space that I think the market is dramatically mispricing. It's trading at a deep discount to its peer group. Its entire 2026 HBM production is already spoken for. It has pricing power, contracted demand, and structural tailwinds that Wall Street hasn't caught up to yet. And it sits at a strategic intersection that makes it irreplaceable in the AI buildout. I just added it to the Moonshot Premium portfolio. Premium Members get the full breakdown below, my entry price, and the thesis behind why I think this is one of the most asymmetric setups in the semiconductor space right now. When every major AI chip maker is locked into multi-year purchase agreements and inventories sit at quarter-century lows, the leverage belongs to the supplier, not the buyer. This company is one of those suppliers. And people are starting to wake up to it so your window is closing fast. The Takeaway Here it is, plain and simple… AI doesn't run on algorithms. It runs on physical infrastructure. Energy. Minerals. And now, memory. Each one is a chokepoint. Each one is tightening. And memory might be tightening fastest of all. The memory-chip shortage isn't a temporary supply hiccup. It's a structural crisis driven by insatiable AI demand, constrained manufacturing capacity, and geopolitical chokepoints on the critical minerals that make production possible. It's already making your phone more expensive, your computer harder to buy, and the AI revolution more fragile than anyone wants to admit. The investors who understand this, who see that the real power in the AI stack isn't the model but the physical layer underneath it, are the ones who'll be positioned when the rest of the market catches up. The one thing you can do right now: Look at your portfolio and ask yourself a simple question. Do I own anything connected to the physical infrastructure that AI actually runs on? Not the apps. Not the chatbots. The chips. The minerals. The memory. If the answer is no, you're exposed to the biggest structural risk in the global economy without even knowing it. The wall is here. The question is which side of it you're standing on. Double D P.S. The write-up on our next pick is live below for Premium Members only. The ticker, the thesis, my buy-up-to price, and a full breakdown of the risks, including the headline that just rattled the stock and the earnings report around the corner. Upgrade to Premium now. P.S. #2 Here's a screenshot of the current Moonshot Minute Portfolio. I've blurred out the tickers since that information is only for Premium Members, but you can see how we've done so far: 🔓 Premium Content Begins Here 🔒 In today's Premium Section, you'll find the new recommendation we're adding to during this memory shortage… because when Jensen Huang, who's running the biggest seller of AI chips in the world, tells you this is a "technological miracle", believe him. I hope you've been paying attention because many of our picks are currently beating the S&P by up to 4-to-1 this year. Most financial newsletters charge $500, $1,000, even $5,000 per year. Why? Because they know they can. I don't. I built my wealth the old-fashioned way, not by selling subscriptions. That's why I priced this at $25/month, or $250/year. Not because it's low quality, but because I don't need to charge the typical prices other newsletters charge. One good trade, idea, or concept could pay for your next decade of subscriptions. The question isn't 'Why is this so cheap?' The question is, 'Why would I charge more?' 👉 Upgrade to Premium Now P.S. If this newsletter were $1,000 per year, you'd have to think about it. You'd weigh your options. You'd analyze the risk. But it's $25 a month. That's the price of a bad lunch decision. And remember, just one good idea could pay for your subscription for a decade. 👉 Upgrade to Premium Now   Update your email preferences or unsubscribe here © 2026 Moonshot Minute 228 Park Ave S, #29976, New York, New York 10003, United States Terms of Service