Oracle is investing heavily in AI infrastructure, with the company investing no less than $40 billion in Nvidia’s most advanced GB200 superchips. This enormous budget is earmarked for the first American Stargate data center in Abilene, Texas. OpenAI will be the main customer for all this AI computing power, as it wants to use the infrastructure for AI training. The question remains whether the data center will be able to supply enough power for all these superchips.
According to various sources, Oracle will receive approximately 400,000 Nvidia GB200 superchips for this investment. The GB200 superchips use Nvidia’s most powerful Blackwell GPUs and its own Grace CPU. Thirty-six of these chips will be combined in an NVL72 system, which can offer 1.4 exaFLOPS. That amounts to enormous computing power for AI applications.
Power consumption is a challenge in Abilene
Oracle will equip the 1.2 gigawatt data center with approximately 11,000 racks, which translates to nearly 16 zettaFLOPS of computing power. Each rack has a price tag of around $3.5 million. However, there are doubts about whether the data center can supply enough electricity to enable all systems to operate simultaneously.
Each rack consumes a maximum of 120 kilowatts, on top of which comes the overall consumption of the building and the cooling of the data center. The energy demand of modern data centers is increasing rapidly due to the rise of AI. In theory, this AI infrastructure would consume more capacity than the data center can handle. Calculations are heading towards 1.5 gigawatts.
In practice, however, infrastructure never runs at 100 percent capacity. Some of the superchips will also be used for tasks such as inference, data generation, and reinforcement learning. These consume less power than intensive AI training. This should enable the data center to operate fully in the future.
Phased rollout until 2026
Before the data center is fully operational, its capacity will need to be expanded. This year, only 200 MW of data center capacity will be available. That is enough for approximately 1,500 NVL72 racks, or 54,000 GB200 superchips. The remaining gigawatt is not expected until 2026. The question is when Nvidia will deliver the 400,000 chips to Oracle. Demand for Nvidia chips is many times higher than the company can supply. Oracle is leasing the data center for at least 15 years, which probably means that OpenAI has also committed for multiple years, given the investment. This seems to be accelerating the cooling down of the relationship between OpenAI and Microsoft.
OpenAI Stargate
Meanwhile, OpenAI is expanding its Stargate program internationally. The company announced plans for additional gigawatts of computing power in the United Arab Emirates, in collaboration with Oracle, Nvidia, and other partners.