Select Language:
Meta Platforms announced plans on Wednesday to develop four new in-house chips as part of their ongoing efforts to expand data center capacity. Similar to other tech giants like Alphabet and Microsoft, Meta has heavily invested in establishing a team capable of designing custom chips alongside purchasing standard components from Nvidia and AMD.
Creating tailored chips suited to specific data processing tasks allows Meta to improve energy efficiency and reduce costs. These new chips are integrated into the Meta Training and Inference Accelerator (MTIA) initiative. The debut chip, labeled the MTIA 300, is already operational and powers Meta’s ranking and recommendation engines.
The remaining three chips are scheduled for release throughout this year and into 2027. The MTIA 450 and 500 are designed to enhance inference processes, which involve AI models—like the one behind ChatGPT—responding to user inputs.
Yee Jiun Song, Meta’s VP of Engineering, emphasized the rising demand for inference, stating, “We see inference demand exploding right now, and that’s what we’re focusing on.” While Meta has experienced some success with inference-specific chips, it has faced challenges in creating a large-scale generative AI training chip capable of building expansive AI models.
Starting with the MTIA 400, which Meta plans to deploy within its data centers, the company has developed a comprehensive system around these chips. This system is roughly the size of multiple server racks and features liquid cooling technology for efficient heat management.
Meta intends to introduce these chips at six-month intervals to keep pace with its rapid data center expansion. As Yee Jiun Song noted, “That’s the reality of how quickly our infrastructure is being built out.” The company has projected capital expenditures between $115 billion and $135 billion for the current year.
Meta collaborates with Broadcom for some design elements, though specifics weren’t disclosed. The chips are manufactured by Taiwan Semiconductor Manufacturing Co. Additionally, Meta reserved large procurement deals in February with Nvidia and AMD, totaling tens of billions of dollars.
ChatGPT
Perplexity
Gemini AI
Grok AI
