AI Power Scaling Cuts Both Ways

AI Power Scaling Cuts Both Ways ← Back AI Power Scaling Cuts Both Ways August 22, 2026 Since the industrial revolution, much has been said about technology's ever-increasing power scaling functions. The power density and power projection of states and corporations has ballooned to mind-boggling levels thanks to ubiquitous communications networks, transportation networks, and economies of scale. But, for example, even as these huge organizations have used small-scale prototyping and manufacturing technologies like CNC and 3D printing to speed up their R&D iteration, so too have these technologies made it possible for end-users to hit the eject button on entire classes of manufactured parts and products. You can't sell nearly as many proprietary injection-molded plastic parts if any Joe can borrow a 3D printer at work and print an open-source STL made by other people who own the product. Plus, garage engineers have been able to accomplish mind-bending feats of open-source product development, without needing the infrastructural inputs of a corporate hierarchy above them to supply hard-to-manufacture parts. AI and autonomous agents might do the same. 1,000,000x I've discussed with coworkers ad nauseum that unknown unknowns abound in AI. There are doubtless many undiscovered algorithmic optimizations or entirely original approaches that will take the same capabilities we spend so much energy and money to use today and multiply their performance by a factor of 10, 100, 1000, even 1 million. Hardware companies are even experimenting with hard-etching model architectures and weights as processors. Inputs on one end, outputs on the other virtually instantaneously. No need to emulate inference via accelerated matrix multiplications when we can do the whole pipeline physically. If those performance gains are still out there, there are two major takeaways: Everybody everywhere will eventually have access to a capable autonomous agent locally in their pocket. That then calls into question the data centers, if everyone can run a model locally on their 5W phone processor. The infrastructure will be obsolete. Like the rest of the software industry, AI already is and will continue to be slowly consumed by the marathon pace of open-source. There will be no business model in AI provision, and its continued R&D will be an academic pursuit once gain, except this time with real-world results immediately in everyone's hands when a paper is published. A New Hope For every leak about how the corpo-state is consolidating analysis of dragnet data collection using AI, identifying standout patterns and filtering through volumes of information that we can't even conceive, there will be more stories about how the open-source agents in everyone's pockets will undercut many of those advances. We'll have Halo's dumb AI or Star Trek's nameless "computer", and it'll be capable of anything we ask. The simple scenario here is enshittification. Imagine: You hate your car's surveillance and data collection features. You also want all its disabled subscription-tied features. You ask the open-source agent on your phone to slurp crowd-sourced knowledge bases about your car. There are protocols for agents to query knowledge bases, and even interface with other agents if necessary. The knowledge bases are built from agentic analyses of ROM dumps of your manufacturer's firmwares. You grab your phone's USB-to-whatever dongle and hook it up to the appropriate interface on your car's motherboard. You ask your agent to use the knowledge base to reflash your car's firmware with a common open-source alternative, whose drivers were built using the agentic decomps. Alternatively, it writes the new firmware on the spot. You configure it with sensible privacy and feature defaults. Congratulations, your car's dash and features are now actually yours. It'll be like Neo downloading kung-fu. Vendor lock-in and rent-seeking enshittification via software will essentially become useless wheel-spinning. The only viable business model for a given product category will be to make working firmware and maintain it until the open source alternative catches up and abrogates it. Expending any more effort doesn't make the revenue graph budge. John Deere will go back to making just tractors. Volkswagen will go back to making just cars. Bambu will go back to making just printers. Now apply this scenario's logic to countless others. This might just be a high-energy kick in the direction of a distributist utopia. ← Back