Meta claims custom chips will reduce costs and energy use compared to Nvidia

Meta is betting that building its own silicon can meaningfully undercut the cost and energy footprint of relying on Nvidia’s GPUs. The company’s custom Meta Training and Inference Accelerator chips, known as MTIA, are already in production, and Meta has laid out a roadmap spanning four generations that it plans to roll out at a pace the semiconductor industry rarely attempts.

The headline numbers are eye-catching: a 44% reduction in total cost of ownership and 40% improvement in power efficiency compared to general-purpose Nvidia GPUs, at least for the specific workloads Meta is targeting. Those workloads include ranking and recommendation systems, ad optimization, and generative AI inference, which collectively represent the engine room of Meta’s business.

Four generations in rapid succession

Meta unveiled its roadmap for the MTIA 300, 400, 450, and 500 chip generations back in March. The MTIA 300 is already running in production environments, handling the inference tasks that power everything from your Instagram feed to the ads you see on Facebook.

The next chip in line, codenamed Iris (also called the MTIA 400), is scheduled to enter production in September 2026 after clearing initial testing phases. Meta says those tests ran for six weeks without surfacing major issues.

Meta is targeting a new chip generation roughly every six months. For context, the semiconductor industry typically operates on cycles of one to two years between generations. The architecture enabling that speed is a modular chiplet design, which essentially means Meta can mix and match building blocks rather than redesigning entire chips from scratch each time. Broadcom collaborated on the program, and fabrication is handled by TSMC.

The economics behind the silicon bet

Meta’s projected capital expenditures for 2026 sit between $115 billion and $135 billion. The company also plans to scale its computing capacity to 14 gigawatts by 2027.

These are application-specific integrated circuits, or ASICs, purpose-built for a narrow set of tasks. They sacrifice the flexibility of Nvidia’s general-purpose GPUs in exchange for doing a smaller number of things more efficiently.

Nvidia is not getting replaced

The MTIA chips are designed for inference, the process of running already-trained AI models to generate outputs. Training those models in the first place still relies on Nvidia and AMD GPUs. Meta continues to maintain partnerships with both Nvidia and AMD worth billions of dollars. The MTIA line is positioned as a complement to those external chips, not a wholesale replacement.

Inference is where the volume lives. Training a large language model happens once (or periodically). Running it for hundreds of millions of users happens constantly. By targeting inference with custom silicon, Meta is attacking the workload that consumes the most compute cycles and therefore the most money.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.