![[images/Why Everyone Is Wrong About the AI Bubble.webp]] **Creator:** Maxinomics · **Published:** 2025-12-23 · **Length:** 21:43 · [Watch on YouTube](https://www.youtube.com/watch?v=Wcv0600V5q4) > *This video looks at three things the AI boom is betting everything on. If any of them turn out to be wrong, the whole thing unravels.* (video description) ## 1) Detailed outline ### [0:00] Intro: the coffee pot and the critical lie - The first viral video was a live stream of a coffee pot. Cambridge students, tired of climbing two flights of stairs to find it empty, hooked a webcam to the internet; by 1995 millions had watched it and used it to explain the internet to friends and family, feeding the hype behind the dot-com bubble. - The question people now ask: is AI a bubble, and can they safely ignore it? - Phil's thesis: every bubble shares one ingredient, **a lie the majority believes**. Three claims about AI could be that lie: **AI will keep getting better**, **we need more data centers**, and **a lot of people are using it a lot**. - To judge them, first understand the lie of the dot-com bubble: **dark fiber**. Sponsor: Public.com (read comes later). ### [1:01] The dot-com bubble - The promise: the internet would be how we did everything (talk, bank, shop, date, watch movies), even though almost no one was online yet. - Being online meant slow loading of one of only a few thousand websites; web developer was barely a job, and sites had no easy way to get paid. Internet access arrived on discs in the mail and ran over phone lines. - **Copper telephone wire** carried about **100,000 times less information** than fiber-optic cable, which could also carry data thousands of miles. Copper was the bottleneck. - The real bubble wasn't software or websites; it was **fiber-optic cable**, "eerily similar" to data centers today except in one critical way (explained later). - History of the US data network: a mesh that began as single iron-strand telegraph lines, upgraded to two copper strands for phone calls. By **1995** the two main New York–LA trunks were fiber, with branches to regional hubs, which was more than enough; nearly all traffic was phone calls. - As the hype built, companies raced to lay fiber. Within about five years there were **13 new cross-country fiber cables**; Houston–LA went from 1 to 5 routes, New York–Washington from 2 to 10, US–Europe from 3 to 12. By **2001**, about **$500 billion** had been spent. - None of it addressed the **last mile**: the couple of miles of telephone wire from each home to the central office. Example: how was **Pets.com** supposed to reach customers still on dial-up? - By 2001, **less than 10%** of the fiber was lit; the other 90% became **dark fiber**. Investors noticed. - Speed matters: research shows people abandon a computer task at around 10 seconds, and sites routinely took 30+ seconds to load in 2001. - The clincher: most homes had **one phone line**, so going online meant no calls in or out, and family fights over the line. - The pop was driven by **telecom companies**, not web companies. They went to zero; news clips cite **Enron** heading into Chapter 11. About **half a million telecom jobs** disappeared, the stock market fell about **50% in 18 months**, and the economy followed. ### [5:56] Claim 1: AI will keep getting better - Humans grasp a word like "third" with all its context (the third planet, Earth, and its neighbors); AI "just knows the word." - How a model is trained: all text humans have created is converted to numbers; random chunks have a word removed; the model guesses (ranking options), the answer is revealed, and its assumptions are adjusted, billions of times. Training is "a giant glorified flashcard session." - The released **model** is just that file of tuned assumptions: "the smartest child that has ever existed." - Half-empty vs. half-full view: humans don't peak in intelligence until around 70, and a child knows physical things AI doesn't (what rain feels like, how water spills, when to run to kick a ball, how much paper weighs). - How much the eyes take in: NFL broadcasts run about **12 Sony HDC cameras** at about **$80,000** each; your two eyes gather roughly as much data as all 12 combined, so identical hard drives would fill at about the same time. - Text "just scratches the surface" of knowledge. Humans estimate distance, predict shadows, and imagine the unseen side of an object. - The gold-nugget test: asked to rotate an image of the largest gold nugget ever found 360°, AI fails as soon as it must imagine the unseen side, because it hasn't handled thousands of rocks the way people have. - Vision, touch, smell, and sound are far denser than text, so the question splits into **can it** (yes) and **will it**, which depends on whether it gets to "play with thousands of rocks." ### [10:36] Sponsor: Public.com - **Generated Assets:** type an investing idea (for example AI-powered supply-chain companies with positive free cash flow, or defense tech firms growing revenue 25%+ a year); AI agents scan every US stock, build an index around the thesis, explain each pick, and let you backtest against the **S&P 500**. - Public also offers stocks, bonds, options, and crypto, plus an uncapped **1% match** on transferred accounts (public.com/max). Paid endorsement. ### [11:35] Claim 2: We need more data centers - Clips of tech leaders promise multi-gigawatt "AI factories," a world covered in data centers, even gigawatt data centers in space, and a site covering a large part of Manhattan's footprint. - Why so big: Manhattan is about **23 square miles**; a data center that size might hold **5,000 rows × 50 racks × about 70 chips**. Predicting each single word spreads the math across dozens to hundreds of chips, per word, not per question. - This is why **Nvidia** can't meet demand, and it's the **critical departure from dot-com**: the fiber sat dark, but every Nvidia chip runs around the clock, sometimes hard enough to melt. "There is not a dark Nvidia chip in the country right now." - Data density: a word is a one-dimensional **vector**; an image section is two-dimensional; video adds time and is three-dimensional. One frame of video is about **a million times** as information-dense as an average word; two seconds of video holds about as much raw information as the entire **Harry Potter** series. - Companies are building for what comes after chatbots (predicting the other side of the gold nugget), egged on by a **2017 paper** written about text that turned out to work on nearly every kind of data. The question isn't whether it works but how to get enough compute. - Possible red flag: "we need more data centers" really means **we need more electricity**. Grid operators keep a safety margin above peak demand (a midday heat wave); announced data centers would nearly erase that margin permanently, 24/7, because data centers never turn off. - Scale of the need: roughly **100 nuclear plants**, or tens of thousands of square miles of solar plus batteries, or windmills covering half of Texas. Electricity demand has never grown this fast in human history. - Companies are restarting **Three Mile Island** or building their own power stations. Like an arms race, everyone has the same paper and knows it works, so no one can fall behind. - Electricity is AI's version of the dot-com **telephone-wire problem**. If it isn't solved, expectations, products, and valuations come into question. It can be done, so Phil raises a **yellow flag**, not a red one. ### [15:49] Claim 3: A lot of people are using it, a lot - Early search engines in 1997–98 were clunky portals. **Google** offered one box, two buttons, and ten blue links, powered by **PageRank** (rank pages by how many other pages link to them). - Google's retention: people who used it three times in one week stayed for life; about **95 of 100** triers became permanent users. Facebook, Instagram, Uber, and WhatsApp came closest but fell well short. - **ChatGPT** initially wasn't on that list: many tried it and never came back. - Then something that "is not a thing" happened: users who'd dropped off five months earlier came back, bending the **retention curve** upward. Partly the model improved; partly OpenAI offered a smarter **paid tier**, and people were glad to pay for more. - The signal to industry: build data centers as fast as possible. ### [17:35] DeepSeek and Jevons paradox - A little-known Chinese company, **DeepSeek** (backed by a hedge fund, "15 guys in an office"), published a paper showing comparable results with about a tenth of the chips; reports said the model cost under **$6 million**. - Panic: the Nasdaq fell, and **Nvidia** dropped about **17%**, roughly **half a trillion dollars** of market value, as many concluded data centers were unnecessary. - People who understood how words, images, and video are represented thought instead: apply that efficiency to all the chips we have, and AI gets about 10× cheaper. It did, and paying users got 10× more for the same price. - **Jevons paradox:** when something gets cheaper, people don't buy the same amount for less; they spend the same and use much more. - Whale oil example: a whale held about **2,000 gallons** of oil; enough to light one lantern nightly for a year cost about **$1,200** in today's money. Whaling ships eventually stayed out three or four years per voyage, and most people went without light after dark. - **Kerosene** from crude oil arrived at about a tenth the price: homes went from one lantern to a lamp per person. Electricity came about 70 years later: a lamp per room became dozens of bulbs. The same pattern holds for cars, refrigerators, coal, oil, and farming. - Data shows that one year after the DeepSeek shock, AI use rose about **50×** ("not 50%, 5,000%"). ### [20:21] The forest of AI - At its core, AI repackages data people already consumed, over the same cables, protocols, and devices, largely from the same companies. - Metaphor (Phil's): incumbent tech companies are giant redwoods unfazed by storms; AI startups are seedlings, and most won't make it without the forest noticing. - A second trait of bubbles: the promised thing is new, a green field. The early internet was; today it's a mature forest, and AI is that forest expanding. - Conclusion: if the **electricity gap** isn't solved, a wildfire rolls through. If it is, given usage and headroom to improve, calling AI a dot-com-style bubble about to burst "is likely missing the forest for the trees." ## 2) Things mentioned ### Economics, markets, and policy ideas - **Speculative bubbles:** the claim that every bubble rests on a lie most people believe and on a brand-new, untested "green field." - **Dot-com bubble and crash (about 1995–2002):** hype-driven investment followed by a roughly 50% market drop over 18 months. - **Infrastructure overbuild / dark fiber:** about $500 billion of fiber laid by 2001 with under 10% in use. - **Last-mile bottleneck:** the home-to-central-office telephone wire that capped demand. - **Telecom bankruptcies and job losses:** about half a million jobs lost. - **AI capex arms race:** competitors compelled to build compute because rivals know the same techniques work. - **Grid reserve margins and peak demand:** the safety buffer data centers could permanently consume. - **User retention curves:** Google's ~95% retention vs. ChatGPT's unusual "smile" as lapsed users returned. - **Freemium / paid tiers:** willingness to pay for smarter models. - **Market shock:** the DeepSeek selloff erasing about half a trillion dollars of Nvidia's value. - **Jevons paradox:** efficiency and lower prices increasing total consumption. - **Energy price history:** whale oil to kerosene (about a tenth the price) to electricity. ### Companies and organizations - **Public.com** (sponsor): investing app; Generated Assets AI index tool; 1% transfer match. - **Nvidia:** AI chip maker that can't meet demand; hit by the DeepSeek selloff. - **DeepSeek:** Chinese AI lab backed by a hedge fund; its efficient model sparked the January 2025 panic. - **OpenAI / ChatGPT:** chatbot whose retention rebounded with model upgrades and a paid tier. - **Google:** search engine with PageRank and exceptional retention. - **Facebook, Instagram, Uber, WhatsApp:** high-retention products that still trail Google. - **Enron:** shown in news footage heading into Chapter 11 during the telecom and energy collapse. - **Pets.com:** emblematic dot-com company stuck behind dial-up customers. - **Sony:** maker of the HDC broadcast cameras used for NFL games. - **NFL:** broadcast example for visual data volume. - **Telecom companies** (unnamed) that laid fiber and went to zero. ### Markets and indexes - **Nasdaq:** tech-heavy index that fell on the DeepSeek news. - **S&P 500:** backtest benchmark in Public's Generated Assets. ### Technology and products - **Trojan Room coffee pot webcam** (Cambridge): the first viral video. - **Dial-up internet** over copper phone lines; internet setup discs mailed to homes. - **Fiber-optic cable:** roughly 100,000× the capacity of copper phone wire. - **Large language models:** next-word prediction trained like flashcards; a model as a file of tuned weights. - **Vectors and tensors:** 1D words, 2D images, 3D video. - **The 2017 paper** that kicked off modern AI (background: the transformer paper "Attention Is All You Need"). - **Multi-gigawatt data centers**, a Manhattan-sized campus, and space-based data centers floated by tech leaders. - **GPUs** running continuously, sometimes melting. - **Sony HDC broadcast camera:** about $80,000 each, 12 per NFL game. - **ChatGPT paid tier.** - **Public.com Generated Assets.** - **PageRank.** - **Whale oil lanterns, kerosene lamps, electric light bulbs.** ### Energy and infrastructure - **US power grid capacity** and realistic available output (maintenance, clouds, calm wind). - **Three Mile Island** nuclear plant being brought back online for data-center power. - **Equivalents for AI demand:** about 100 nuclear plants, tens of thousands of square miles of solar plus batteries, or wind farms covering half of Texas. - **Company-owned power stations** for data centers. ### Places - **Cambridge, England** (coffee pot); **Manhattan** (23 sq mi data-center comparison). - **Fiber routes:** New York–Los Angeles, Houston–LA, New York–Washington, US–Europe. - **Texas** (wind comparison); **China** (DeepSeek). ### Figures and cultural references - **Harry Potter book series:** about the raw information of two seconds of video. - **The largest gold nugget ever found:** used as the AI "unseen side" test (background: usually identified as the Welcome Stranger, found in Australia in 1869; the video doesn't name it). ## 3) Biographies ### Phil Andrews (presenter) Producer and on-camera host of Maxinomics ("the guy in the videos and the comments," per the credits). He explains economics, technology, and history through visual storytelling, here framing the AI boom against the dot-com crash. ### The Cambridge coffee-pot students (subjects) University of Cambridge computer scientists who in 1991 pointed a camera at the shared coffee pot in the Trojan Room so they could check it remotely; it later went on the web and became a famous early internet attraction. (Background: the setup is usually credited to Quentin Stafford-Fraser and Paul Jardetzky; the video doesn't name them.) ### William Stanley Jevons (referenced) English economist (1835–1882) whose 1865 book *The Coal Question* observed that more efficient coal use raised total coal consumption. The effect, Jevons paradox, is the video's explanation for why cheaper AI led to far more AI use. (Background; the video names only the paradox.) ### DeepSeek's founders (subjects) The video describes DeepSeek as "15 guys in an office" at a hedge fund rather than a tech company. (Background: DeepSeek was founded by Liang Wenfeng and is backed by the Chinese quantitative hedge fund High-Flyer; its low-cost R1 model triggered the January 2025 Nvidia selloff.) ### Google's founders (referenced) PageRank, which the video credits for Google's simple, sticky search, was developed by Larry Page and Sergey Brin at Stanford in the late 1990s. (Background; the video doesn't name them.) ### The kerosene inventor (referenced) The video says "a guy figured out how to make kerosene" at a tenth the price of whale oil. (Background: kerosene refining is usually credited to Canadian geologist Abraham Gesner in the 1840s–50s; the video doesn't name him.) ### Tech leaders in news clips (subjects) Unidentified executives and commentators shown promising multi-gigawatt AI data centers, data centers in space, and a Manhattan-scale campus, alongside news anchors covering DeepSeek and the market drop. The video doesn't caption their names. ### Production credits Producer Phil Andrews; video editor Christie Muldoon; motion graphics Seth Laupus; director of production services Sam Wolf; thumbnail Luca Depardon; franchise content producer Tariq Abdellatif; chief content officer Devin Emery.