OpenAI's Gigaflop Supercomputer Starts Operation: Costing $50 Billion, Computing Power Price Reduction Still Insufficient

According to the Beating Monitor, during a live interview on CNBC's "Power Lunch" program, OpenAI CEO Sam Altman confirmed that the super data center project led by OpenAI and developed in collaboration with Oracle (codenamed Stargate) has officially broken ground in Saline, Michigan. The total planned capacity of the project is 1 gigawatt, and based on estimates from partners and the market, the initial investment required for site preparation and construction alone is $16 billion. Additional investment of at least $30 billion will be needed for subsequent hardware and software procurement, bringing the total investment for this phase to a range of $45 billion to $50 billion.

Responding to questions about the ROI of the massive computing power investment, Altman stated that the revenue and actual usage of large models in the enterprise sector are growing rapidly. He pointed out that although hardware upgrades have led to a continuous decrease in Token invocation prices, the pace of computing power consumption growth far exceeds the rate of price reduction. With future AI agents running in the background 24/7 and processing massive amounts of data, global computing power demand is expected to continue facing a severe shortage in the next two to three years, making computing power a highly sought-after asset that still outstrips supply.

Regarding some companies' complaints about the current status quo of "high input, low output" in large model applications, Altman believes this is the most pertinent criticism of the current AI industry. He admitted that there is indeed inefficient usage and waste of computing power among businesses currently, but he anticipates that within 1 to 2 years, as application scenarios and business loops mature, companies will rationalize their spending, aligning large model budgets more precisely with specific business outputs.