Why AI Isn't Coming to Your Desk (It's Physics, Not Greed)

Do we actually need $100 billion AI data centers, when a ~$4,700 box on your desk can already run serious AI models? I work in operations and supply chain, so I went through the evidence and checked four popular claims. Short answer: AI is splitting in two, and only half of it is coming home. 📝 Article + charts: https://www.actvli.com/decentralized-... 📄 White paper PDF (free, full sources): https://www.actvli.com/wp-content/upl... ⏱️ CHAPTERS 00:00 Do we need $100B AI data centers? 02:08 Nvidia's ~$4,700 "personal AI supercomputer" 05:23 History: computers you visited, not owned 07:27 What AI training actually is (the dials) 11:37 Training vs. inference 12:57 Why training needs a whole building (it's physics) 15:40 How fast home AI hardware is improving 17:36 The 4 claims · Claim 1: data centers are overbuilt 22:10 Claim 2: is local AI greener? (PUE + Jevons paradox) 25:33 How chips are made: wafers and TSMC 27:46 Claim 3: chips will keep getting cheaper 31:16 Claim 4: history repeats (Grosch's law) 34:23 The verdict + 3 questions before you pick cloud or local 36:12 Does training ever finish? 36:57 Is today's AI already "enough"? 🔎 THE 4 CLAIMS, AND WHERE THEY LANDED 1️⃣ "Data centers are a bubble that home hardware will replace" → REJECTED. Training a frontier model needs thousands of GPUs syncing every step. Inside one server that sync runs at about 900 GB/s. Between servers it is roughly 9 to 18 times slower. Home broadband sits another two to three orders of magnitude below that. That's physics, not pricing. But inference (using a finished model) really is moving to your desk. 2️⃣ "Running AI on your own device is better for the planet" → UNPROVEN. Big data centers report PUE around 1.09 to 1.2 against an industry average of 1.52, and the IEA expects data center electricity use to roughly double by 2030. No study has measured energy per AI query at home vs. in a data center. The indirect evidence leans against the claim. 3️⃣ "Chips keep getting cheaper, so local AI will too" → REJECTED. TSMC's 2nm wafers cost about $30,000, more than 50% above 3nm, with the next node rumoured at $45,000. Cost per transistor stalled around 2012. Now it's going up. 4️⃣ "Mainframe to PC proves AI will decentralize too" → REJECTED as usually argued, TRUE in a narrower form. History decentralized independent, single-user jobs (word processing, spreadsheets, and today, AI inference). It never decentralized one giant job that has to stay in sync, which is what AI training is. 🧭 IF YOU RUN A BUSINESS, ASK THREE THINGS BEFORE CHOOSING CLOUD OR LOCAL AI 1. Do you need frontier capability, or would a smaller local model do the job? 2. Is your demand steady, or occasional? Idle hardware is the most expensive kind. 3. Does the data have a legal or practical reason it can't leave the device? 👤 WHO AM I I'm Hani Mebar: 24 years in operations, supply chain and programme management (Nokia, HMD, Alma Career), MBA, Lean Six Sigma Black Belt, founder of Äctvli Responsible Consulting. I've spent my career on the manufacturing side of the chips this video is about. 📚 KEY SOURCES (full list in the white paper) • IEA: Data centre electricity use surged in 2025 (April 2026) • Douillard et al. (2023): DiLoCo, Google DeepMind • Jaghouar et al. (2024): INTELLECT-1, Prime Intellect • Lidin et al. (2026): Covenant-72B • Or-Bach (2014): Moore's Law has stopped at 28nm • Uptime Institute: large data centers are mostly more efficient • Epoch AI (2026): training compute growth ~4x per year Corrections to a few on-camera numbers are in the pinned comment. Where do you land: central, local, or both? Tell me in the comments. #AI #DataCenters #LocalAI #Nvidia #AIInfrastructure #DGXSpark #TSMC #Semiconductors #CloudComputing #LLM #Actvli