I am not a finance guy. I am an engineer. But I\u2019ve been in this industry for forty+ years, and I\u2019ve got enough scar tissue to look at a P&L sheet and ask some questions. And frankly, I don’t like what I’m seeing.<\/p>\n\n\n\n
When I look at the current AI landscape, I see a trillion-dollar venture. As I wrote recently in AI Bubble? : AI Skeptics Hard Look at the Value Equation,<\/a> the math of this investment requires scrutiny, not just optimism. You have reports like Leopold Aschenbrenner\u2019s “Situational Awareness” paper<\/a> arguing that AGI by 2027 is a certainty based on straight lines on a graph. You have accelerationists claiming that winning AI is a geostrategic categorical imperative\u2014that the U.S. must secure superintelligence before China does to ensure national security. (If you want to lose some sleep, watch the AI 2027<\/a> video).\u00a0 The narrative is that it is effectively a winner take all race.\u00a0 But breathy narratives rarely hold up in the real world. We need to focus on Getting the AI Hype Cycle Right<\/a> – distinguishing between inevitable progress and the economic reality.<\/p>\n\n\n\n I always like to look at the physics of a situation. And the physics suggest that this massive spending isn’t building a defensible fortress. It\u2019s building a money pit.<\/p>\n\n\n\n Because here is the hard truth: AI ages like fresh fish.<\/strong><\/p>\n\n\n\n\n\n\n\n Let\u2019s talk about the physics of capital expenditure. In a normal times, you build an asset\u2014a factory, an oil rig, a data center\u2014and you depreciate that asset over 10 or 20 years. That asset generates value for decades.<\/p>\n\n\n\n In the current AI climate, compute hardware and trained models depreciation is not a matter of decades, but quarters and maybe even months.<\/p>\n\n\n\n\n\n A while back, a memo leaked from Google that said, “We Have No Moat, And Neither Does OpenAI.<\/a>” It was an eye-opener. (Note: proprietary data might provide a moat.)<\/p>\n\n\n\n We are seeing the proof of this right now. Look at the recentKimi K2.<\/a> It achieved near-frontier capability in reasoning and coding shortly after the market leaders, but it was built at a fraction of the cost (as in < $5m). There is no secret sauce here. Techniques like RLHF or Chain-of-Thought leak into the research ecosystem instantly. I always tell engineers: don’t start with the tool, start with the problem. What is the customer actually trying to do?<\/p>\n\n\n\n The question here is this: how good does the AI need to be for the problem the customer has? When you want to add 2+2, there is no difference between a $10 calculator and a $100 million supercomputer.<\/strong><\/p>\n\n\n\n Just to be clear, I\u2019m passionate about AI for Material Science. Here I’m a hard core AI accelerationist – we have many many miles to go.\u00a0 Those are supercomputer<\/strong> jobs.\u00a0 And I’m absolutely convinced that new materials will transform our world and bring huge economic benefits (over decades!). \u00a0 But how many problems are like that?\u00a0 Lots of businesses just need AI to summarize PDFs, improve a manufacturing process, or route customer service tickets. These are calculator<\/strong> jobs.\u00a0<\/p>\n\n\n\n This obsession with building the God Model\u2014the single most capable Foundation Model\u2014misses the point of technology. The real metric of success isn’t the perplexity score of a model; it is economic growth. Satya Nadella nailed it when he said, \u201cthe real benchmark for AI, if it’s a new industrial revolution, is achieving \u2018Industrial Revolution type of growth\u2019 in GDP<\/em>\u201d<\/p>\n\n\n\n Economic growth does not come from having a model that can write a sonnet in French. It comes from taking a model and applying it to a boring, messy, high-friction problem domain like manufacturing supply chains, marketing analytics, or insurance claims processing.<\/p>\n\n\n\n And here is the kicker: We probably already have the models we need to solve those problems.<\/strong><\/p>\n\n\n\n The current generation of models are more than capable of driving massive efficiency gains in these sectors right now. We don’t need to wait for AGI or the Trillion Dollar Cluster to fix a broken procurement process or improve our manufacturing. We just need to do the hard, dirty, unglamorous work of integration and application development. . This is the ‘Phase 2’ I discussed in AI Underpants Gnomes<\/strong><\/a>, without this missing step, there is no profit, only a collection of expensive underpants.<\/p>\n\n\n\n I heard someone recently say that if AI progress stopped today, we would be able to take existing capabilities and generate trillions of dollars of progress with it over the next 10 years. That rings true to me.<\/p>\n\n\n\n The ultimate winners of this revolution won’t be the labs lighting billions of dollars on fire to build the best<\/em> model. The winners will be the application layer companies that utilize this newly cheap, commoditized intelligence to solve actual problems for actual customers.\u00a0<\/p>\n\n\n\n Stop focusing on the fresh fish. I am not a finance guy. I am an engineer. But I\u2019ve been in this industry for forty+ years, and I\u2019ve got enough scar tissue to look at a P&L sheet and ask some questions. And frankly, I don’t like … Continue reading The Economics of Fresh Fish<\/strong><\/h1>\n\n\n\n
The No Moat Reality<\/strong><\/h1>\n\n\n\n
The pioneers pay the crazy costs of R&D exploration; the fast followers just pay for implementation<\/p>\n\n\n\nThe Calculator vs. Supercomputer Problem<\/strong><\/h1>\n\n\n\n
The Real Success is Economic Growth<\/strong><\/h1>\n\n\n\n
Focus on the meal you’re cooking for the customer.<\/p>\n\n\n