{"id":"2093074599995187682","url":"https://x.com/DavidOndrej1/status/2093074599995187682","text":"if you're building software,\n\nread this... https://x.com/i/article/2093073177916448768","author":{"name":"David Ondrej","username":"DavidOndrej1","avatarUrl":"https://pbs.twimg.com/profile_images/1648700022563696642/fkEMa1qo_200x200.jpg"},"createdAt":"Thu Aug 27 20:34:06 +0000 2026","engagement":{"replies":17,"retweets":30,"likes":365,"views":127952},"article":{"title":"Inference will eat the world.","previewText":"Every month, we spend more and more money on AI inference. This is true for me, for all the people on my team, for all the people in the AI space I know, and it's probably true for you too.\nWe are not","coverImageUrl":"https://pbs.twimg.com/media/HQwYFSuXEAAbn8i.jpg","content":"Every month, we spend more and more money on AI inference. This is true for me, for all the people on my team, for all the people in the AI space I know, and it's probably true for you too.\n\nWe are not paying for more vibe-coded AI SaaS tools... we are paying for tokens.\n\nAnd this trend is not stopping anytime soon.\n\nSure, there are going to be optimizations -- for example, open-source AI models from China often greatly over-perform closed models from America, giving you far more value for far less cost. Subscriptions like OpenCode Go give you access to near-frontier AI models, for 20x less than Claude Max subscription.\n\nBut despite all these optimizations, the amount of tokens burned keeps rising month after month. No matter where you look -- OpenRouter data, Vercel AI gateway data, estimates from OpenAI's and Anthropic's API -- the number of tokens generated keep growing. And it's not growing slowly, it's not growing linearly, it's growing exponentially.\n\nThis is why INFERENCE will eat the entire software world.\n\nThinking from first principles makes this obvious. Models are improving every 1-2 months, and there is no sign of those improvements slowing down (if anything, they are speeding up.) Agentic Harnesses are also improving rapidly -- a year ago, we had basically only Claude Code, and now we have Pi, Codex, Hermes, OMP, Droid, Prime Agent... and many others.\n\nThese are just two factors pointing towards the same conclusion: Agents are becoming more and more powerful.\n\nTranslated into plain English: every month, agents are able to do a larger percentage of the total software tasks there is. This is obvious to anyone with two eyes. It doesn't take a genius to see where this leads... a world where agents are doing the vast majority of interactions with all software that exists.\n\nPeople are choosing banks based on which one has an API. People are ordering food via DoorDash CLI. Why? Because instead of using some random shitty UI, they can just talk to their personal agent, tell him what they want, and the agent handles the rest.\n\nAnd the more APIs there are, the better Computer Use gets, the more connectors exist, the more the agents will be able to do. Period.\n\nSo not only is each agent being able to handle more tasks, and more difficult tasks, but each person (at least those who are locked-in) is now running multiple agents in parallel. Just ask yourself: are you running more agents now, than you were 6 or 12 months ago? I don't have to know you, to know that the answer is Yes.\n\nThat said, it should be now clear to you that more agents = more tokens, and more tokens = more money spent on inference.\n\nIn the past, even if you had a simple legal question, you had to pay a human lawyer to find out. Now, you just ask ChatGPT Pro. Just a few years ago, if you had to categorize thousands of accounting transactions, you'd pay an accountant to do that for you. Now, you just tell Claude Code to write you python script.\n\nAre you starting to see the broader point?\n\nMoney, that would otherwise go to other non-inference computer work, either products or services, is now simply going to INFERENCE.\n\nYou don't pay for human translators anymore, you just ask an AI chatbot to translate it. You don't pay someone for to research something for you, you just ask Perplexity. You don't pay a frontend developer to polish your website, you just have Fable do it. All of that spend, all of that purchasing power, all of that revenue, is going to INFERENCE.\n\nThis is why I believe that inference will eat the software world, slowly but surely. (or more like \"quickly but surely\"...)\n\nNow, after wrestling with this idea for many days, I did manage to come up with some things that cannot be replaced by just \"throwing more tokens at it\" -- industries, fields, tasks that Agents, no matter how smart, will not be able to do. However, that is a topic for another article.\n\nFor now, understand this: Anything that is easily verifiable, anything that has a clear objective function, will be done (sooner or later) by AI Agents, whether you like it or not.\n\nInference will be one of the biggest markets in the world, and it will be the biggest market in all of computer-work. What you do with this information is up to you.\n\n-written by David Ondrej \n\n(manually, by hand, keyboard press after keyboard press)"},"adhxContext":{"savedByCount":1,"publicTags":[],"previewUrl":"https://adhx.com/DavidOndrej1/status/2093074599995187682"}}