The AI industry runs on Nvidia GPUs. For years, every major tech company — Meta, Google, Microsoft, Amazon — has depended on Nvidia's H100 and B200 chips to train and run large language models. But that near-monopoly comes with eye-watering costs: a single B200 GPU can cost tens of thousands of dollars, and hyperscalers spend tens of billions annually. Meta alone projected $125 to $145 billion in capital expenditures for 2026, mostly for AI compute.
Meta's answer is the Iris chip, the latest generation of its MTIA (Meta Training and Inference Accelerator) program, also known as Project Artemis. Designed in partnership with Broadcom and manufactured by TSMC, Iris will enter production in September 2026. Samsung supplies the high-bandwidth memory. The chip is purpose-built for AI inference — running trained models — rather than the training phase, where Meta will still rely on Nvidia GPUs for now.
The strategic logic is straightforward. By designing its own silicon, Meta can optimize the hardware-software stack end to end. Its PyTorch framework, Llama open-weight models, and the Iris chip are designed as a unified system. This vertical integration mirrors Apple's approach with its M-series chips and Google's TPUs, but applied to the largest social media infrastructure on the planet. The chips will power recommendation algorithms on Facebook and Instagram, as well as Llama-based AI features across WhatsApp, Messenger, and the metaverse.
Energy efficiency is a core driver. Meta plans to bring 7 gigawatts of compute capacity online in 2026 and double that in 2027. At that scale, even modest efficiency gains translate into hundreds of millions of dollars in electricity savings. Custom silicon designed specifically for Meta's workloads — rather than general-purpose GPUs — can achieve far better performance per watt.
The move also reshapes the supplier power dynamic. While Meta will not completely sever ties with Nvidia, having an in-house alternative strengthens its negotiating position and insulates it from supply-chain shocks. The semiconductor industry has experienced chronic shortages in recent years, and geopolitical tensions around TSMC's Taiwan foundries add another layer of risk. Producing its own chips gives Meta more control over its destiny.
For the broader AI ecosystem, Meta's pivot reinforces an accelerating trend: the big tech companies are all becoming chip companies. OpenAI is reportedly exploring custom silicon. Amazon has its Trainium and Inferentia chips. Google has TPUs. Microsoft partnered with AMD for custom accelerators. The era of one company supplying AI compute for everyone is drawing to a close.
Knowledge takeaway: Meta's Iris (MTIA/Artemis) chip, entering mass production in September 2026, represents a strategic shift from reliance on Nvidia GPUs toward vertical integration — custom silicon designed specifically for Meta's workloads, manufactured by TSMC, and tightly coupled with the Llama AI model family and PyTorch framework.