The social notworking outfit has rolled out a “small deployment” of in-house silicon, developed in partnership with Taiwan’s TSMC, with plans to ramp up production if things don’t go pear-shaped.
Historically, Meta has dabbled in custom AI chips but only for running models—not training them. Previous attempts at designing its silicon have ended in disappointment, with projects being scrapped or scaled back after failing to meet internal benchmarks.
But with AI workloads growing and Nvidia’s prices soaring, Meta is desperate to reduce the absurd $65 billion it plans to spend on capital expenditures this year—most of which is going straight into Jensen Huang’s pockets.
Meta was one of the first companies to experiment with RISC-V-based AI inference chips years ago, trying to cut costs and reduce its dependence on Nvidia’s hardware stranglehold.
But now, according to Reuters, the company has gone a step further—designing its own AI training accelerator, presumably with Broadcom’s help. If this chip meets Meta’s performance and efficiency goals, it could start weaning the company off Nvidia’s high-end AI GPUs, such as the H100, H200, B100, and the upcoming B200.
Meta and Broadcom have already taped out their first AI training accelerator, with TSMC producing the initial working samples.
According to the report, the partners have successfully brought up the chip, and Meta has started a limited deployment to assess its real-world performance. It remains unclear whether Meta is running full benchmark tests or the chip is already performing actual AI workloads, but at the very least, the hardware is no longer just a concept.
While the exact specifications of the chip remain a mystery, AI training hardware typically relies on a systolic array architecture—a grid of processing elements designed to chew through massive datasets in matrix and vector operations.
Given the nature of AI training, it’s likely that Meta’s accelerator features HBM3 or HBM3E memory to handle the immense bandwidth demands. Since this is a fully bespoke processor, Meta has probably fine-tuned its supported data formats and instructions to optimise die size, power consumption, and overall efficiency.
If it wants to compete with Nvidia’s latest offerings, the chip must deliver top-tier performance-per-watt, keeping up with—or surpassing—the H200, B200, and possibly even the next-gen B300.
This new accelerator is part of Meta’s ongoing Meta Training and Inference Accelerator (MTIA) programme. But history hasn’t been kind to Meta’s in-house silicon efforts.
The company previously scrapped its internal inference processor after it failed to meet performance and power efficiency targets during early deployment tests. That failure forced Meta to abandon its short-term custom hardware dreams in 2022, leading to massive orders for Nvidia GPUs to keep its AI ambitions afloat.
