Meta Scales Video with FFmpeg, Scraps In-House AI Chip
Meta is heavily leveraging the open-source FFmpeg library to handle complex video processing and manipulation at a massive scale. Meanwhile, the company has reportedly scrapped its advanced in-house AI chip project, choosing to deepen its reliance on Nvidia and AMD for high-performance compute.
Meta's reliance on FFmpeg involves executing its command-line tools tens of billions of times daily. The company recently moved away from a long-standing internal fork of the software, collaborating with FFmpeg developers to integrate custom features like threaded multi-lane transcoding directly into the upstream open-source project. This software strategy is paired with custom hardware. Meta has integrated support for its own video transcoding ASIC, the Meta Scalable Video Processor (MSVP), into FFmpeg via standard APIs. This approach allows for unified tooling across different hardware, including accelerators from Nvidia, AMD, and Intel. The scrapped AI training chip, codenamed "Olympus," is the latest in a series of setbacks for Meta's custom silicon ambitions. An earlier inference chip was cancelled for underperforming, and a second-generation training chip, "Iris," was also abandoned due to design and efficiency issues. Reasons cited for shelving Olympus included its design being too complex for mass production, insufficient software stability, and the high risks involved while competing directly with OpenAI and Google. This decision reflects a broader challenge, as Meta's chip programs have reportedly struggled to build engineering teams sufficient for such complex hardware development. In response, Meta is committing up to $135 billion in capital expenditures for 2026 AI infrastructure, creating a tripartite supplier structure. This includes a multi-year deal with Google to rent its Tensor Processing Units (TPUs) for model development. The company also signed a multi-year agreement with AMD worth over $100 billion for up to six gigawatts of MI450 GPUs, which will include custom silicon tailored for Meta's Llama workloads. This is in addition to a separate deal to purchase millions of next-generation GPUs and CPUs from Nvidia, which currently dominates the AI accelerator market with an approximate 80% share.
