According to the Financial Times report, Oracle will spend approximately $40 billion to purchase Nvidia's latest superchips, with plans to provide computing power for OpenAI. These superchips will be deployed in the first U.S. "StarGate" data center located in Abilene, Texas. However, whether this data center can provide enough power to meet such a huge computational demand remains in question.
Image Source Note: Image generated by AI, authorized service provider MidJourney
This massive investment will be used to purchase around 400,000 Nvidia GB200 superchips. The chip was first released at Nvidia's GTC event, equipped with a pair of powerful Blackwell GPUs and its own Grace CPU. A system of 36 GB200s forms an NVL72 system, capable of providing 1.4 exaFLOPS of sparse FP4 computing power. Based on this calculation, Oracle will deploy approximately 11,000 rack systems within a 1.2 gigawatt facility, with overall computing power reaching nearly 16 zettaFLOPs.
Each NVL72 rack has a maximum power consumption of 120 kilowatts. Considering electricity and cooling losses, it is estimated that about 1.45 gigawatts of power would be needed to fully utilize the potential of these superchips. However, this data center can only provide about 200 megawatts of power this year, which is sufficient to support about 1,500 NVL72 racks, or approximately 54,000 GB200 superchips. The remaining power is expected to be available by 2026, and Oracle plans to lease this site for 15 years.
Although the power issue may pose a challenge, Oracle and data center operator Crusoe are likely to find solutions. Not all racks will reach maximum power consumption at the same time. Considering that the Abilene campus consists of eight buildings, Oracle is unlikely to concentrate all superchips in one training cluster. We anticipate that some chips will be used for inference, synthetic data generation, reinforcement learning, and other workloads, which are less likely to push the system to its limits.
If Oracle and Crusoe can overcome the power limitations, the Abilene data center will become one of the most powerful AI supercomputers in the United States, with computing power 10 to 20 times greater than Elon Musk's Colossus supercomputer. The Colossus system located in Tennessee is equipped with nearly 800 exaFLOPS of sparse FP8 computing power, and currently only one substation is completed, with partial power still relying on multiple natural gas turbines.
In addition, OpenAI's "StarGate" project will also expand internationally, collaborating with companies like Oracle, Nvidia, and Cisco to build new computing capabilities in the UAE. This project is expected to go online with the first 200-megawatt power phase in 2026.
Key Points:
🌟 Oracle will invest $40 billion to purchase 400,000 Nvidia GB200 superchips to assist OpenAI.
⚡ The Abilene data center may face challenges due to insufficient power, with only 200 megawatts of power currently available.
🌍 OpenAI's "StarGate" project is expanding internationally, planning to establish more computing capabilities in the UAE.
