Anindya Ghose

The market is obsessed with spotting an AI bubble. Meanwhile, a stampede is forming behind them... Five months ago we built the only evidence-based framework to answer: is AI a bubble? We tracked 5 indicators: economic strain, industry strain, revenue momentum, valuation heat, funding quality. Our conclusion then: AI is a productive boom, not a bubble. The evidence has only gotten stronger. --- The bear case sounds reasonable: Capex growing faster than revenue. Model costs falling (DeepSeek proved it). Most enterprise AI is still chatbot stuff. @michaeljburry shorted Nvidia and Palantir. Fund managers call AI overexposure their top tail risk. The FT has published 100+ articles invoking "AI bubble." But reality moved the other way. --- We follow the money. Monthly AI revenue: $772M (Jan 2024) to $13.8B (Dec 2025). An 18x increase in two years. The industry strain ratio – investment to revenue – dropped from 6.1x to 4.7x in five months. If it holds, it crosses below our 3x threshold by Q2 this year. This will mean that revenue is starting to carry the installed base. --- The hyperscalers are the engine: Google Cloud: +48% YoY to $17.7B AWS: +24% to $35.6B Azure: +39%, backlog up 110% to $625B @sundarpichai, @satyanadella, and @ajassy all say the same thing: AI is the main growth driver. When the CEOs of the three largest cloud companies tell you the same story, the attribution question answers itself. --- But revenue alone isn't enough. Is anyone actually using this stuff? We analysed 6,000+ S&P 500 earnings calls from Q4 2022 through Q4 2025, and extracted 30,000 AI-related statements. Companies making quantified AI claims – specific numbers, specific outcomes – jumped from 1.9% to 13.2%. --- Boring adoption is bullish... Bank of America: AI coding tools cut dev time 30%, saving 2,000 full-time engineer equivalents. Norway's $2T sovereign wealth fund automated portfolio monitoring with Claude, saving $17-32M/year. Meta: 30% increase in engineering output. Power users at 80%. Western Digital, hard disk manufacturer: productivity gains up to 10%. --- This is a technology that has crossed from experiment to infrastructure. Deloitte (Jan 2026): 25% of orgs have 40%+ of AI projects in production. 54% expect to reach that within 6 months. KPMG: 67% of CEOs expect AI returns within 1-3 years – down from 3-5-year horizons just a year ago. --- Something else changed at the end of 2025. Models crossed a threshold. They can now work reliably on tasks lasting 1-2 hours. Agentic AI has arrived. Claude Code became a $1B revenue business in its first six months. Today it's likely generating $3B+ ARR. Its GitHub usage grew 3x in six months. Nicholas Carlini used Claude Code to build a C++ compiler for $20,000 in API costs. The same project built by humans: 5-10 engineers, 18-24 months, ~$2-3M. At Exponential View, we've written several hundred thousand lines of code this year. Software that might have cost $1M has cost ~$500. --- Here's why this changes everything for infrastructure: The chatbot era: a few hundred tokens per interaction. The agentic era: hundreds of thousands to millions of tokens per task. We went from snacking to feasting. And the kitchen was built for snacks. --- The supply crunch is real: 40-50GW of data center capacity in queue, only 11-14GW online. New data center interconnection requests: 3-5 year wait times. AWS lost a $10M Fortnite contract because it couldn't guarantee compute. Microsoft had to choose between its own products and Azure customers, wiping 13% off its share price. All because demand WAS there. --- The bottleneck is physics: power generation, chip supply, construction timelines, trained workforce. GE Vernova's turbine backlog: filled through 2030. Vertiv's cooling backlog: doubled YoY, orders up 250%. Quanta Services construction backlog: $40B. Amazon bought a copper mine. @elonmusk wants data centers in space. --- Using OpenAI's heuristic of 1GW = $10B ARR: Current online capacity supports $110-140B in revenue. At my personal usage of $50/day in tokens, that's capacity for only 6-7 million users at that intensity. There are 20-30 million developers in the world. Not even enough for them, let alone everyone else. --- The bears were answering yesterday's question all along. The revenue trajectory has closed the bubble debate – industry strain is falling, enterprise adoption is deepening. The boring phase is finally ending. The real question: can we build fast enough? cc @Jsevillamol @DKThomp @tglocer @p_ferragu @mattturck @ttunguz @AlexandraMousav @erikbryn @EconoScribe @deedydas Full analysis: exponentialview.co/p/bubble-…