Meta
META · NASDAQMeta is trying to do something no other hyperscaler has managed yet: build a chip good enough to train its own frontier AI models, not just run them afterward.
Who they are
Meta Platforms, the parent company of Facebook, Instagram, and WhatsApp, uses AI for two very different jobs: ranking and recommending content and ads across more than 3 billion daily users, and training and running its open-weight Llama model family, the most-downloaded large language model family in the world. CEO Mark Zuckerberg has framed the current buildout as one of the largest infrastructure bets in the company's history.
What they do
To reduce how much it pays Nvidia, Meta designs its own custom silicon, the MTIA (Meta Training and Inference Accelerator) family, and on March 11, 2026 revealed four new MTIA generations at once — MTIA 300, 400, 450, and 500 — built on a RISC-V architecture and fabricated by TSMC. MTIA 300 already handles ranking and recommendation work in production; MTIA 400 has finished lab testing for generative AI inference. A chip capable of training frontier models, the hardest workload and the one where Nvidia's ecosystem is stickiest, is still in development.
How Meta makes money
Meta's core business remains advertising, but its 2026 capital spending, guided as high as $125–145 billion, nearly double 2025's level, is almost entirely AI infrastructure: training and serving the models that power ad targeting, content recommendations, and Llama. Its flagship Hyperion data center in Richland Parish, Louisiana, is targeting 5 gigawatts of capacity, alongside other builds underway in Ohio and Indiana.
The bigger trend
Meta remains one of Nvidia's largest GPU customers even as it races to build MTIA, because moving any single workload onto in-house silicon is a multi-year effort and Meta's compute needs are growing faster than its own chip production can cover. Its Reality Labs division, the metaverse and AR bet, lost $4.03 billion in the first quarter of 2026 alone, adding pressure on investors already uneasy about the pace of AI capex versus visible revenue return.
Whether Meta ships a training-capable MTIA chip and actually reduces its Nvidia GPU spend rather than simply adding capacity on top; how the Hyperion, Ohio, and Indiana data center builds track against the $125–145 billion 2026 capex guide; whether investor patience holds if Reality Labs losses continue alongside AI infrastructure spending.
Related companies
Frequently asked questions
Partially. Meta uses custom MTIA chips for some workloads, particularly content ranking and recommendations, but remains one of Nvidia's largest GPU customers for AI training, since a chip capable of training frontier models is still in development.
Hyperion is Meta's flagship AI data center under construction in Richland Parish, Louisiana, targeting 5 gigawatts of capacity, one of the largest single AI infrastructure builds Meta has announced.
Meta uses AI to power ad targeting and content recommendations across its platforms, and also develops the open-weight Llama model family. Management has said underinvesting in compute capacity is a bigger risk than overinvesting, given how directly AI improvements have driven revenue.
This page describes public value-chain positioning for informational purposes only. It is not investment advice, and inclusion here is not a recommendation to buy or sell any security. Figures reflect public reporting as of mid-2026 and may have changed since.