Meta's AI Chips: Mass Production Begins in September to Reduce Reliance on Nvidia

The Race to Break Free from Nvidia's Dominance

Meta is moving rapidly toward building an independent computing infrastructure. Reports from internal sources indicate that the company will commence production of its latest custom AI chips this coming September. This move comes amid an unprecedented shortage of computing components and a sharp rise in GPU costs, pushing the company to accelerate its journey toward technological self-sufficiency.

Multiple Industry Partnerships to Achieve the Goal

Meta is not pursuing this path alone — it has forged an extensive network of industry partnerships to ensure the quality of its final product. The company is collaborating with chip design giant Broadcom on chip architecture, while entrusting manufacturing to TSMC of Taiwan, the global leader in semiconductor manufacturing. The partnerships extend further to include:

• Random access memory from Samsung
• Storage units from Sandisk
• Fiber optic equipment from Sumitomo Electric

Notably, one of the chips passed the testing phase in approximately six weeks — an achievement that reflects the maturity of the company's development process.

The MTIA Program: A Modular Architecture for the Future

These chips fall under the Meta Training and Inference Accelerator (MTIA) program, which the company unveiled last March, encompassing four generations of specialized chips. Meta's approach in this program relies on a modular design based on what are known as "chiplets" — small, assemblable semiconductor units — giving the company the flexibility to adapt to the rapidly evolving demands of AI without needing to redesign from scratch each time.

The company describes its philosophy by stating that each generation builds on the previous one, leveraging the latest hardware technologies and AI workload requirements, with shorter and more efficient deployment cycles. Meta intends to use these chips for training ranking and recommendation models, general AI workloads, and inference operations tied to its applications.

Billions in Spending to Secure Computing Capacity

This initiative comes within the context of an ambitious investment strategy Meta announced last April, allocating capital expenditures of between $125 billion and $145 billion this year, with a significant portion directed toward strengthening its AI infrastructure. The company is aiming to deploy computing capacity of 7 gigawatts this year, with plans to double that figure next year.

Meta's investments are not limited to developing its own chips — they run in parallel with major deals that include:

• An agreement with ARM to secure computing capabilities for recommendation systems
• A multibillion-dollar deal with AMD for Instinct GPU units
• A contract with Amazon to leverage its proprietary processors for AI tasks

A Wave of Technological Independence Sweeps Through Tech Giants

Meta is not alone in this trend — the development of proprietary chips has become increasingly widespread across the global tech landscape. OpenAI has announced an inference processor being built in partnership with Broadcom, while Anthropic is exploring the development of its own independent chips in collaboration with Samsung. Google and Amazon, meanwhile, already have dedicated chips for training and inference. This landscape reveals a clear strategic direction: the drive to reduce near-total dependence on Nvidia, which dominates the AI GPU market.

Ultimately, Meta's move toward producing its own chips represents a significant shift in its technology strategy — one that combines medium-term cost reduction with long-term technological sovereignty, at a time when competition over AI tools is intensifying at an unprecedented pace.

✦ بقلم فريق دروب أيديا