Meta Platforms has acquired Rivos, a RISC-V-focused semiconductor startup, in a move that signals its intent to accelerate the development of custom compute engines tailored for its massive AI and social media workloads. The acquisition, confirmed by Walden Catalyst—the investment arm of Walden International—marks a pivotal step in Meta’s long-term strategy to design its own CPUs, XPUs, and interconnects, reducing reliance on third-party hardware and driving down infrastructure costs. Rivos, founded in 2021 by a team with deep roots in semiconductor design—including former Apple, PA Semi, Broadcom, and Google veterans—had been working on high-performance RISC-V-based processors and a data-parallel accelerator designed for AI and analytics workloads. The company reportedly developed a 3.1 GHz processor and built a CUDA-compatible software stack, aiming to enable AI models to run efficiently on its custom silicon. This compatibility is critical, as it could allow Meta to run existing CUDA-optimized AI code on non-Nvidia hardware, potentially disrupting Nvidia’s dominance in AI acceleration. Meta’s interest in Rivos began through a prior collaboration to help design the Meta Training and Inference Accelerator (MTIA) series, which are based on RISC-V cores and used for AI inference. The MTIA 2i, launched in 2025, reportedly delivered 44% lower total cost of ownership (TCO) than Nvidia GPUs for certain deep learning recommendation models—key workloads powering Meta’s ad systems. These models require massive memory for embeddings, exceeding the capacity of even Nvidia’s Grace CPU-based solutions, making custom silicon with integrated high-bandwidth memory (HBM) and DDR support a strategic necessity. The acquisition is reminiscent of Amazon’s purchase of Annapurna Labs in 2015, which became the foundation for AWS’s Graviton CPUs and Trainium/Inferentia XPUs. Similarly, Rivos is expected to become the core of Meta’s future compute engine roadmap, potentially leading to a RISC-V-based “superchip” akin to Nvidia’s Grace-Hopper and Grace-Blackwell hybrids—combining CPU and GPU functionality with unified memory access across DDR and HBM. While Rivos had raised $250 million in Series A and additional funding, with plans for a $500 million Series B that would have pushed its valuation above $2 billion, Meta’s offer proved irresistible. The deal likely valued Rivos at over $850 million, reflecting its technical talent, IP, and strategic alignment with Meta’s AI ambitions. One key challenge lies in the CUDA compatibility claim. Nvidia’s licensing agreements prohibit the use of translation layers that allow CUDA binaries to run on non-Nvidia hardware. While Rivos may have used source-to-source compilation or emulation, this could raise legal concerns. However, if Meta uses the technology internally and does not distribute it, the legal risk may be mitigated—especially if it can argue that Nvidia’s software-hardware bundling constitutes anti-competitive behavior. With its $50 billion in R&D and $66–72 billion in capital spending projected for 2025, Meta has the resources to push this vision forward. The integration of Rivos’ team could enable Meta to build a cost-competitive, RISC-V-based AI infrastructure stack—potentially even entering the systems market with high-performance, software-compatible hardware that could undercut Nvidia’s pricing. The move underscores Meta’s transformation from a social media company to a full-stack compute innovator, aiming to control every layer of its AI infrastructure. If successful, it could redefine the economics of AI at scale.