The race to build artificial intelligence (AI) infrastructure is entering a new phase as Meta decides to expand long-term cooperation with Broadcom to develop customized AI chip generations.
The agreement is extended until 2029, showing Meta's great ambition in autonomizing computing capabilities and reducing dependence on external chip suppliers.
According to the announcement, the two sides committed to building computing infrastructure with an initial capacity exceeding 1 gigawatt, equivalent to the amount of electricity supplied to about 750,000 households in the US.
This is considered the first step in a strategy to expand to a scale of many gigawatts, in order to meet the huge computing needs of new generation AI models.
Meta leader Mark Zuckerberg said this partnership is the foundation for the company to build a "giant computing system", serving the goal of developing "super personal intelligence" for billions of users worldwide.
In the context of AI increasingly playing a central role in digital products, infrastructure capacity has become a decisive factor in the competitive position of technology giants.
The trend of self-designing chips is spreading in the technology industry. Besides Meta, corporations such as Google or Amazon are also promoting the development of their own chips to reduce costs and avoid dependence on suppliers such as Nvidia, which is dominating the AI chip market with high prices and limited supply.
In this wave, Broadcom emerged as one of the biggest beneficiaries. The company not only participates in designing custom chips for customers but also provides accompanying software infrastructure solutions.
In addition, Broadcom's Ethernet network technology will be used by Meta to connect large-scale AI computer clusters, helping to increase processing efficiency and optimize operation.
As for Meta, the company is implementing its own chip development program called Meta Training and Inference Accelerator (MTIA).
The first generation, MTIA 300, has been put into operation to serve the rating and content proposal system. The next three versions are expected to be launched before 2027, focusing on inference tasks, which is the stage where AI processes and responds to user requests in real time.
Along with technological changes, Meta is also adjusting senior personnel. Board member Tracey Travis will not run for re-election in the upcoming shareholders' meeting, marking a shift in the leadership apparatus as the company enters a period of strong investment in AI.
Analysts believe that Meta's expansion of cooperation with Broadcom is not only aimed at accelerating AI development but also demonstrates a long-term strategy from controlling the entire value chain from hardware to software.
In the global AI race, businesses with large-scale and flexible computing infrastructure will have a superior advantage.
