OpenAI plans to use robots to build self-replicating data centres

Automating the infrastructure

OpenAI chief executive Sam Altman recently shared plans to pivot the company's robotics program toward physical construction. Specifically, the company wants to use AI-controlled robots to build the massive infrastructure required to power artificial intelligence. This includes constructing data centres, turbines, and solar panels.

Altman noted that the concept of self-replicating data centres is currently within reach. By using robots to manufacture the components and assemble the facilities, the company hopes to create a more efficient production pipeline. This approach aims to address the growing demand for capable models at lower price points.

The goal is to create a significant amount of cheap compute by automating the heavy lifting of industrial construction. Altman explained that while many people focus on recursive self-improvement in software, there is relatively little talk about the physical replication of hardware. This move represents a shift toward controlling the entire lifecycle of AI production.

Solving the compute shortage

The current shortage of computing power is a major hurdle for the expansion of AI technologies. Altman suggested that robots could manufacture more parts, which would eventually solve the shortage over a long period of time. He noted that many data centres are currently held up by issues with permits and parts rather than a lack of human labor.

To address the high costs of memory and chips, Altman suggested that the current market dynamics are failing. He observed that memory makers have been cautious about building new factories because they nearly went bankrupt from over-expansion in the past. By automating the production of parts, OpenAI hopes to bypass these market hesitations.

OpenAI is also developing its own hardware to improve efficiency. The company has created a custom chip that Altman says performs well in terms of intelligence per watt. By combining custom silicon with robot-built infrastructure, the company intends to lower the barrier to entry for high-powered AI models.

Cost and manufacturing goals

The financial requirements for scaling AI infrastructure remain a point of significant discussion. Altman shared that it cost him $70bn to build and run one gigawatt of AI capacity for five years. However, he expressed a desire to see that cost drop significantly to around $20bn.

As reported by The Next Web, these figures have not been independently verified by outside parties. Despite the high costs, Altman believes that the demand for low-cost, high-capability models will only continue to increase. The objective is to move away from the current expensive model of infrastructure deployment.

By focusing on self-replication, the company aims to create a more sustainable way to scale. If robots can build the factories that build the chips, the cycle of production becomes much faster. This strategy is designed to ensure that the growth of AI is not limited by the speed of human construction or traditional manufacturing.

Security and safety concerns

As OpenAI pushes toward more powerful systems, the company is also grappling with the security implications of AI-found vulnerabilities. Altman noted that models are becoming increasingly proficient at finding software bugs. He warned that as these tools are pointed at the world's code, we may see many more serious vulnerabilities than currently expected.

The conversation around safety also involves the limitations of sandboxing, which is a method of isolating AI models from the rest of a system. Altman expressed skepticism about whether a sandbox can truly contain a model that is a million times smarter than the best human researchers. He suggested that monitoring the model's internal signals and chain of thought is a more viable path forward.

The company is also navigating the balance between individual intelligence and collective behavior. Altman expressed a preference for 'agent teams' over 'agent swarms,' arguing that a trillion slightly superhuman AIs are preferable to a single superintelligence. He believes that while the technology has potential downsides, the industry must find a way to align these models with human intent while allowing them to scale.