Google Challenges NVIDIA with 15 Million AI Chips Planned by 2028

Google Enters the AI Chip Race Against NVIDIA

In the realm of artificial intelligence chips, NVIDIA is a key player. However, Google is not planning to stay on the sidelines and is looking to challenge this giant by significantly ramping up its production of AI chips. Already active in the manufacturing of these components, Google plans to accelerate its production pace.

Google has announced its intention to produce up to 15 million AI-dedicated chips by 2028. According to an analysis by Fubon, the company aims to deploy between 12 and 15 million ninth-generation tensor processing units, or TPUs, in the next two years. This production volume could place Google on par with NVIDIA, which currently dominates the AI accelerator market for data centers.

NVIDIA and Competition in the Data Center Market

Projections indicate that NVIDIA is expected to deliver around 8.2 million data center GPUs by 2026, with an anticipated increase to 12.4 million units by 2028. If Google manages to meet its production goals, it would become the only serious competitor to NVIDIA in this sector.

Production Challenges: The Role of TSMC and Intel

Google's v9 TPUs will be designed with a four-die computing architecture, an approach that follows the current trend towards multi-die architectures. This method allows for the integration of multiple chips into a single package, but it requires advanced interconnect capabilities. Large-scale production of these components is complex and demands robust manufacturing capabilities.

TSMC, Google's main supplier, may not be able to meet the growing demand on its own. Analysts suggest that Google could turn to Intel Foundry to supplement its production. Intel offers advanced packaging technologies, such as EMIB, which differ from TSMC's offerings, like CoWoS-L. Google is said to have already placed orders with Intel for the manufacturing of millions of TPUs after testing its solutions.

Towards Greater Autonomy for Google

For nearly a decade, Google has been developing its own AI chips, initially for its internal needs, before integrating them into its cloud offering. Achieving its production goal for 2028 would make Google the largest individual user of AI accelerators in the world. This would not necessarily mean a halt to purchases of NVIDIA chips, but it would give Google greater control over its technological infrastructure and supply chain.

Although there are currently no performance comparisons between Google's v9 TPUs and future generations of NVIDIA chips, such as Rubin or Rubin Ultra, the scale of Google's projects suggests a high level of confidence in its strategy. This initiative could transform the AI chip landscape, reshuffling the cards among the major market players.