Meta Starts Trial Of Own-made AI Chip, Challenges Nvidia's Domination

JAKARTA Meta, parent company Facebook, Instagram, and WhatsApp, has started a trial of homemade chips to train artificial intelligence (AI) systems. This move marks an important milestone in the efforts of the California companies to develop their own custom chips and reduce dependence on external suppliers like Nvidia.

According to sources cited by Reuters, the chip is being tested on a small scale, with plans to increase production if trials succeed. The development of the in-house chip is part of Meta's long-term strategy to reduce the huge cost of infrastructure, along with their massive investment in AI technology.

Meta estimates that the total expenditure in 2025 will reach between US$114 billion to 119 billion (Rp1,874.7 trillion-1,956.9 trillion), including up to US$65 billion for capital expenditures that will mostly be used for AI infrastructure.

One source said that this AI Meta chip is a dedicated accelerator, which means it is specially designed to handle AI tasks only. This approach is believed to be more power efficient compared to conventional GPUs widely used for AI workloads.

The chip is produced by Taiwan Semiconductor Manufacturing Company (TSMC), one of the largest chip manufacturers in the world. The chip production process reaches the tape-out stage, which is the initial design delivery to factories for production, which is an important milestone in silicon development.

However, this stage is not without risk. If the initial trial fails, Meta must analyze the error and repeat the tape-out process, which could take three to six months and cost up to tens of millions of dollars.

This latest chip is part of the Meta Training and Inference Accelerator (MTIA) program, which has experienced various obstacles in recent years. Previously, Meta had canceled a similar chip project after failing to undergo a small-scale testing phase.

However, since last year, Meta has started using the MTIA chip for the inference process of running the AI system as users interact with it on a recommendation system that determines content on Facebook and Instagram.

Going forward, Meta plans to use this chip not only for recommendation systems but also for generative AI products, such as Meta AI chatbots. Meta targets widespread use of this chip for AI model training by 2026.

Meskipun Meta tengah mengembangkan chip sendiri, perusahaan ini tetap menjadi pelanggan utama Nvidia. Pada 2022, setelah chip kustom sebelumnya gagal dalam uji coba, Meta justru memesan GPU Nvidia senilai miliaran dolar.

The GPUs are used to train various AI models, including their advertising system and big language model, the Llama Foundation Model Series. In addition, the Nvidia chip also handles inference for more than 3 billion users of the Meta app every day.

However, Nvidia's dominance in the AI industry is starting to be questioned. AI researchers increasingly doubt the effectiveness of a "scale-up" approach that relies on more data and computing power to improve AI models.

This doubt has grown stronger since DeepSeeek, an AI startup from China, launched a cost-effective model that optimizes computational efficiency through inference, compared to conventional AI models.

DeepSek's launch even caused panic in the global AI stock market, with Nvidia's stock dropping to 20% before finally recovering. Even so, investors still believe that Nvidia chips are still an industry standard for AI training and inference.

Meta is now entering a critical phase in their AI strategy. If this AI in-house chip is successful, Meta could reduce its dependence on Nvidia and reduce their operational costs. However, if these trials fail as before, Meta will likely again rely on Nvidia GPU to meet their AI needs.

With big ambitions in AI, this Meta move can become a game-changer in the artificial intelligence industry, as well as a new challenge for Nvidia's dominance in the AI chip market.