Meta Scraps Advanced AI Chip, Turns to Nvidia
Meta has reportedly scrapped its advanced in-house AI chip project, opting to deepen its reliance on partners like Nvidia and AMD. The move highlights the immense difficulty and cost of competing in the chip arms race, forcing even tech giants to make pragmatic build-vs-buy decisions for their AI infrastructure.
The now-abandoned project was Meta's most advanced training chip, internally codenamed "Olympus," which the company has now scrapped. This follows the cancellation of a previous second-generation training chip, "Iris," in 2024. The primary reasons for shelving Olympus included its complex design, concerns about software stability, and the immense risk of falling behind competitors like OpenAI and Google. Meta's internal chip development, part of the Meta Training and Inference Accelerator (MTIA) program, has faced significant hurdles. The company's first-generation MTIA chip, introduced in 2023, showed some efficiency for simpler AI tasks but lagged behind GPUs on more complex workloads. The challenges with its more advanced successors highlight a deeper issue: the lack of a sufficiently large engineering team to manage the complexities of advanced chip design and ensure commercial viability. Pivoting from its internal efforts, Meta is committing billions to external suppliers. The company has inked a "multigenerational" deal with Nvidia to purchase millions of its next-generation chips, including the Vera Rubin GPUs and Grace CPUs. This deal is part of a larger plan to invest up to $135 billion in AI infrastructure in 2026 alone. This strategic shift isn't limited to Nvidia. Meta has also secured a multi-year, $100 billion deal with AMD for its MI450 GPUs and is even reportedly renting AI chips from its direct competitor, Google. This diversified sourcing strategy underscores the immense computational power required for Meta's AI ambitions, which CEO Mark Zuckerberg noted involves deploying "tens of gigawatts" of computing power.
