One telling sign of this commitment came from Broadcom CEO Hock Tan, who, during a recent earnings call, directly addressed analyst concerns. "Now contrary to recent analyst reports, Meta’s custom accelerator MTIA road map is alive and well," he stated, adding, "We’re shipping now." Broadcom contracts with Meta to assist with certain chip design elements, according to Reuters.
The new MTIA roadmap emphasizes inference workloads, which involve running AI models to make predictions or recommendations. These tasks often have more predictable computational patterns compared to AI training, which is the process of building the models themselves and is currently dominated by Nvidia's powerful Graphics Processing Units (GPUs). This is a crucial distinction. Yee Jiun Song, Meta's Vice President of Engineering, explained to CNBC that custom chips allow Meta to "squeeze more price per performance" across its data center fleet.
Meta has already deployed its MTIA 300 chip, which is being used for ranking and recommendations training within its systems. The upcoming MTIA 400, 450, and 500 chips are designed to handle a broader range of workloads. However, Meta's blog post indicates they will "primarily use these chips to support GenAI inference production in the near future and into 2027." For example, one Meta data center rack will incorporate 72 of the MTIA 400 chips, specifically optimized for AI inference tasks.
While the immediate focus is on inference, Meta's long-term vision includes expanding custom chip design to encompass training models as well. Meta CFO Susan Li noted at Morgan Stanley's tech conference earlier this month that the company "eventually" plans for this expansion. This suggests a methodical approach: tackle the more predictable inference workloads first to gain expertise and cost efficiencies, then potentially take on the more demanding training chip development later.
For Developers
Expect Meta to continue optimizing its platforms for GenAI inference. If you're building applications for Meta's ecosystem, understanding the efficiency gains from these custom chips could inform your deployment strategies. For Founders/Investors in AI Infrastructure: This move highlights the intense competition and strategic importance of specialized hardware in AI. Meta's continued investment, even alongside external deals, signals a long-term trend towards custom silicon for specific AI tasks, creating both challenges and opportunities for niche hardware providers. For Consumers: The ultimate goal of these custom chips is to power more precise ad targeting and more engaging user experiences. You can anticipate more relevant content and potentially faster AI-driven features within Meta's apps as these chips roll out across data centers. Frequently Asked Questions Why is Meta developing its own AI chips? Meta is developing custom AI chips to reduce its dependence on external vendors, lower overall costs for its massive AI workloads, and specifically optimize performance for its unique data center needs, particularly for inference tasks like content ranking and recommendations. What is the difference between AI inference and training chips? AI training chips are designed for the computationally intensive process of building and teaching AI models. AI inference chips, on the other hand, are optimized for the more predictable task of running trained AI models to make predictions or generate content, which Meta's new MTIA chips are primarily targeting. Does this mean Meta will stop buying chips from Nvidia and AMD? No, Meta continues to sign multi-billion dollar deals with external chipmakers like Nvidia and AMD. Its custom chip strategy is about diversification and optimizing specific workloads, not a complete replacement of its existing hardware supply chain. Research Sources
Meta is developing its own custom AI chips, called MTIA, to handle AI inference workloads and reduce reliance on external chipmakers like Nvidia and AMD. They plan to deploy four new generations of these chips by the end of 2027, optimizing performance and cost for their AI infrastructure. While still purchasing chips from Nvidia and AMD, Meta's custom chips are tailored for specific, high-volume tasks.
Meta scrapped its more advanced AI training chip, Olympus, due to design hurdles, but is still committed to custom silicon. They are focusing on inference workloads first, which are more predictable, allowing them to gain expertise and cost efficiencies before potentially tackling the more demanding training models. This allows them to optimize price per performance across its data center fleet.
Meta has already deployed its MTIA 300 chip for ranking and recommendations within its systems. The upcoming MTIA 400, 450, and 500 chips are designed to handle a broader range of workloads, primarily supporting GenAI inference production in the near future and into 2027. For example, one Meta data center rack will incorporate 72 of the MTIA 400 chips, specifically optimized for AI inference tasks.
AI inference refers to running AI models to make predictions or recommendations. These tasks often have more predictable computational patterns compared to AI training, which is the process of building the models themselves. Meta is focusing on inference workloads to optimize price per performance across its data center fleet.
More insights on trending topics and technology
