Alphabet, Meta, and Oracle Challenge Nvidia CUDA Dominance

Alphabet, Meta Platforms, and Oracle are advancing efforts to disrupt the dominance of Nvidia's proprietary CUDA software platform, which has long locked data centers into using Nvidia's graphics processing units. According to YahooStocks0, open-source developments around PyTorch and new cloud infrastructure deals are weakening Nvidia's grip on data center spending.

PyTorch and Hardware Diversity

Meta Platforms utilizes the open-source PyTorch framework for deep learning and large language model development. Because PyTorch has historically been optimized for Nvidia's CUDA, it reinforced customer lock-in. However, Alphabet recently announced TorchTPU, an open-source tech stack allowing developers to run PyTorch workloads on Tensor Processing Units with minimal code changes. Amazon also introduced a backend supporting PyTorch on its Trainium chips.

According to Jana van Greunen, Meta's director of PyTorch engineering, advancing AI requires hardware diversity without being constrained by a single architecture. While these competing chips do not yet run PyTorch as efficiently as Nvidia's hardware due to deep CUDA optimization, the software updates pave the way for future portability and competition.

Oracle Cloud Infrastructure Partnership

Alphabet has also expanded access to its custom TPUs. After starting TPU sales to select third parties earlier this year, Alphabet announced an agreement with Oracle at the end of last month to bring TPUs to Oracle Cloud Infrastructure starting in 2027. This arrangement allows developers hosting databases with Oracle to deploy workloads on TPUs, offering potential advantages in performance and pricing over Nvidia GPUs.

Nvidia shares currently trade at about 22 times forward earnings expectations, reflecting potential risks as alternative AI accelerators gain traction and utility across major cloud providers.

Key facts

• Alphabet, Meta, and Oracle are developing alternatives to weaken Nvidia's CUDA software lock-in.
• Alphabet introduced TorchTPU to run PyTorch workloads on Google Tensor Processing Units.
• Alphabet's TPUs will come to Oracle Cloud Infrastructure starting in 2027.
• Meta works with multiple hardware providers to avoid reliance on a single architecture.