Social media giant Meta is facing unprecedented pressure on AI infrastructure costs, with AI-related spending expected to soar to an astonishing $65 billion, and the overall annual total spending may reach $119 billion. Faced with such as an astronomical bill, the technology giant finally couldn't sit still and resolutely embarked on the road of self-developed AI chips and has made significant progress in this field. The latest report shows that Meta is about to launch a small-scale self-developed chip deployment, a move that indicates that Meta will gradually get rid of its dependence on Nvidia and its expensive GPUs to reduce the huge cost of training artificial intelligence models.
Meta did not make the decision to develop its own chips on impulse. In fact, they once put aside the project, and the reason may be related to the many challenges in the R&D process. But now, Meta executives seem to have overcome these obstacles and have high hopes for self-developed chips. They expect this chip to be put into use in 2026, first used for training AI models, and then expand to generative AI products, such as AI chatbots.
According to Reuters, Meta's self-developed AI chip is a dedicated accelerator, which means its only mission is to efficiently handle various heavy tasks related to artificial intelligence. In addition to directly "hardly" Nvidia's GPU procurement costs that are "sky-high" at every price, self-developed chips can also significantly reduce the energy consumption of infrastructure. Since it is tailored for specific AI tasks, this chip will far exceed that of a general-purpose GPU in terms of energy efficiency, and will undoubtedly save Meta a lot of electricity bills.
It is reported that TSMC will be responsible for the production and manufacturing of this customized chip, but the specific process technology will be used is not clearly mentioned in the report. However, details revealed that Meta has successfully completed the first tape-out of this AI chip, which is usually a complex process that costs millions of dollars and takes up to six months. Even so, the success of the die-casting does not mean that everything is going well. There is still uncertainty as to whether the chip can fully meet the performance requirements of Meta. Once the chip fails to meet expectations, Meta will have to invest more time and energy to troubleshoot problems, diagnose faults, and may need to perform redischarge, which will undoubtedly further increase R&D costs.
Previously, due to difficulties in R&D, Meta once suspended its own AI chip plan, but now it seems that they have successfully overcome these obstacles. Meta's senior management expects that self-developed chips can start to play a role in 2026. Its primary goal is to train Meta's own AI system and then gradually apply it to generative AI products, such as the highly anticipated AI chatbots. Nvidia is still enjoying the dividends brought by the surge in GPU sales. Meta is still one of its most profitable customers, but with the advancement of Meta's self-developed chips, this situation may change in the near future.