A Commoditized AI is a Good AI
Milton Friedman advocated strongly against centralized power for it leads to oppression. Instead, decentralized systems—whether in governance, markets, or technology—help preserve individual freedom. That is because competition, choice, and dispersed decision-making safeguard against tyranny by reducing the ability of a single entity (or a small group of them) to exert control over people’s lives. AI is not an exception. When just a few players own the models, applications, inputs and, eventually, outputs, there is no limit to the size of the jackboot on every person’s neck.
The cost of developing and operating a large language model like ChatGPT, for example, is comparable to the price tag of a Boeing 747 aircraft, i.e. approximately 100 million GBP, depending on the model. Let’s break it down. The research and development costs require between 10-30 highly qualified professionals whose salaries range between 100 to 300 thousand USD, making the personnel cost for 1 or 2 years in the region of 10 million USD. The data collection and preparation phase mean sourcing high-quality datasets and cleaning, annotating, and preprocessing the data, setting anyone off for 1 million USD. For software licenses and tools, we’re talking between 200 and 500 thousand USD. Then, in order to train the LLM, large-scale cloud computing is required, using high-performance GPUs like the NVIDIA A100 which costs up to 20 million USD for a single training run, and multiple training runs are required, gathering a fat check of around 50 million USD. The energy consumption would be around half a million USD. Once everything is ready for launch, deployment costs involve infrastructure setup, API integration, and that is in the region of 1 million USD. If all of this happens in Europe, complying with GDPR and implementing security protocols would cost around half a million on top. And, finally, the ongoing yearly costs would include cloud hosting and scaling for 12 million USD, continuous model improvement at around 5 million USD, monitoring and troubleshooting for 2 million USD, and 1 million USD for the lawyers, of course—because we are worth it 😊 Grand total? 90 million USD just in the first year.
With numbers like these, it is almost assured that AGI will not emerge from the garage of a college dropout living with his parents. However, this is changing and it is a good thing. Open source (the Midas touch!) is starting to change the landscape of that teen in a garage. Models like LLaMa 2, Mistral 7B, and the DeepSeek of recent fame, are all open source and democratize the game. Moreover, more efficient AI architectures are emerging like MoE or Quantization maintain performance while being designed to be smaller and faster. Add to that the stampede away from big providers’ cloud computing towards marketplaces for AI compute sharing and the availability of cheaper consumer hardware, and you get the perfect storm that can commoditize AI.
The commodification taking place matters a lot because it avoids AI monopolies controlling what models get built and who can access them. Also, such a landscape enhances transparency by virtue of a decentralized AI ecosystem that leads to even more open models and, with that, there are more localized solutions and faster technological evolution. As a cherry on top, if small players can train and run their own models, decoupling from cloud-based services reduces data harvesting.
So, by overcoming compute power, data monopolies, energy costs, and technical complexity, the door is open for a virtuous circle of more and more open-source models, the triumph of federated learning, mushrooming of compute cooperatives, crowdsourced datasets, and further experimentation with blockchain projects. Imagine a world where AI is as accessible as electricity! There would be community-driven AI projects (like Linux is for the OS oligopoly) and low-resource models chiseling away the monolithic power of a few players.
Even for a non-technical person the idea of having an AI monopoly evokes (correctly) the feared “SkyNet” scenario. Power corrupts. Concentration of power corrupts faster. If the power is fragmented, nobody can exert sufficient chunks of it to subjugate everybody. In our age, when we are inching closer to true behemoths that may reach AGI, the solution against shoggoths (look it up 😊) can very well emerge from the garage of a college dropout living with his parents. LFG!
