Meta Launches Production of Its Own AI Chip to Cut Nvidia Dependence

Meta will begin production in September 2026 of the latest version of its in-house artificial intelligence chip, the Meta Training and Inference Accelerator, as part of a strategy to reduce reliance on external chip suppliers, primarily Nvidia and AMD. The news was reported by Reuters, citing an internal company document.

MTIA is a project Meta has been developing for several years as an alternative to Nvidia's costly graphics chips, which currently underpin most of the company's AI infrastructure. The new generation of the chip, set to enter production in September, is meant to let Meta shift part of the workload tied to training ranking and recommendation models, as well as inference in apps like Facebook and Instagram, onto its own hardware.

Scale of the investment

Meta says it will bring 7 gigawatts of computing power online in 2026 and plans to double that figure in 2027. That is an enormous scale for a single company's infrastructure, comparable to the energy needs of a medium-sized city. Capital expenditures for 2026 are expected to reach between $125 billion and $145 billion, a significant share of which will go toward AI infrastructure, data centers, and development of the company's own chips.

The company, cited by tech media, describes its approach to chip design as iterative: each new generation of MTIA builds on the previous one, uses a modular chiplet architecture, and incorporates the latest insights from AI workloads and available hardware technologies.

Each generation of MTIA builds on the previous one, using modular chiplets and incorporating the latest insights from AI workloads and hardware technologies. - Meta

Why Nvidia should take notice

With this move, Meta joins the ranks of Nvidia's biggest customers that are also investing in their own AI chips, alongside Google with its TPU line and Amazon with its Trainium chips. For Nvidia, it's a signal that even the largest buyers of its graphics processors see in-house chips as a way to curb rising computing costs, not merely a side experiment.

The partnership with Broadcom on chip design and with TSMC on manufacturing shows that Meta isn't building its own fabs, but instead tapping the existing semiconductor supply chain, much as other companies designing their own AI chips do. Samsung's role in supplying RAM and Sandisk's in supplying storage memory round out the picture of a complex, multi-tiered production chain.

Implications for Poland's market

For Polish companies using cloud computing and AI models, Meta's decision has an indirect but real significance. If major players manage to cut infrastructure costs more effectively through their own chips, that could eventually translate into cheaper services based on recommendation and inference models, the kind relied on by advertisers and e-commerce platforms operating in the Polish market.

Meta's expansion of its own chip infrastructure also fits a broader trend of large tech companies moving away from full dependence on a single AI hardware supplier, a shift being watched closely by analysts tracking Nvidia's valuation after the company's recent stock price swings.

Sources: TechCrunch (techcrunch.com), Aroged (aroged.com)