OpenAI is making a major change to its ambitious 'Stargate' supercomputer plan, but it's not a step back. Instead, it's a strategic pivot from building its own massive data centers to securing computing power through large, long-term deals with specialized partners.
Originally, the plan involved joint-venture projects to construct custom data centers, like the ones proposed in the UK and Norway. However, OpenAI is now shifting towards bilateral offtake agreements. This means they are essentially renting vast amounts of computing capacity from hyperscalers like Microsoft and other specialized cloud providers. This move allows OpenAI to avoid the immense risks, costs, and long timelines associated with building and operating its own infrastructure.
Several factors triggered this strategic shift. First, recent events were the immediate cause. OpenAI paused its UK project due to high energy costs and unclear regulations around AI training and copyright. At the same time, Microsoft took over the data center capacity in Narvik, Norway, that was originally intended for OpenAI. This was a clear signal that OpenAI was moving to access compute through its biggest partner rather than building it alone.
Second, broader market dynamics made this pivot a logical choice. Key component supply chains, especially for HBM (High-Bandwidth Memory), are extremely tight. Hyperscalers who have already secured large allocations are in a much stronger position. Furthermore, with Nvidia's next-generation 'Vera Rubin' GPUs set to be deployed first through major cloud providers, renting is the fastest way for OpenAI to get its hands on the latest technology.
Ultimately, this isn't a retreat from OpenAI's goal of securing massive amounts of compute. It's a smarter, more flexible way to get there. By shifting the construction and operational risks to its partners, OpenAI can focus on what it does best: developing cutting-edge AI models, while ensuring it has the computational power it needs for the future.
- Glossary
- Hyperscaler: A large-scale cloud service provider that can offer massive computing, storage, and networking services. Examples include Microsoft Azure, Amazon Web Services (AWS), and Google Cloud.
- Bilateral Offtake Agreement: A long-term contract between two parties, where one agrees to purchase a specific amount of a product or service (in this case, computing capacity) from the other.
- HBM (High-Bandwidth Memory): A type of high-performance computer memory used in high-end GPUs, essential for training large AI models.