OpenAI CEO Sam Altman expressed concern about those who want AI to reason locally on the device, arguing that RAM should be reserved for the data center.

"A lot of people want to use local inference. They believe the problem is solved by having open source models that run on personal computers and don't need things in data centers," Altman said in an interview with YouTube channel Cheeky Pint published on October 30.

Local inference is the process of letting a trained AI model run directly on the personal device's hardware, such as the CPU, GPU or NPU on a computer or phone, instead of sending data to the cloud servers of third-party services like ChatGPT and Claude.

"From a global resource allocation standpoint, it's bad that the limited supply of RAM sits in everyone's computer and is only used 5% of the time. We need to make the most of cloud resources if we want to deal with today's very limited supply, but people want to go against that," he added.

According to Tom's Hardware, OpenAI has released a number of open source models, but its main business is still based on closed models, operating in data centers.

"Altman is actually defending OpenAI's business model. The company doesn't feel threatened by personal computers, but the ability to run inference locally means users will pay less for models like the GPT-6 Luna," Tom's Hardware said.

The OpenAI CEO's statement is based on the assumption that data center construction and expansion will not slow down, while RAM running 24/7 in data centers will bring more benefits than operating on users' computers.

To ensure supply for AI infrastructure, technology corporations are actively signing long-term supply contracts. Micron has signed 26 strategic agreements, covering more than 35% of expected revenue through 2030, accompanied by $32 billion in committed deposits from customers. Samsung is said to have closed long-term contracts covering 75-80% of its expected memory chip output in 2027. SK Hynix also pushed for a similar deal.

Unlike previous cycles, in which chip companies often expanded factories when demand was high, causing the market to then become surplus and cool down, this time manufacturers chose to increase prices multiple times instead of significantly expanding capacity. That decision led to supply continuing to fall short of demand and creating lasting price pressure.

Analysts predict the memory chip shortage will last at least until 2028, giving manufacturers a negotiating advantage.

"Whether Altman mentioned personal computers of users, or devices of AI-based application developers, the result is the same. Billions of dollars spent by companies to order RAM is the reason why computers, phones and many other technologies are becoming increasingly expensive," commented The Verge writer Steve Bonifiled.

Diep Anh summary