Big tech companies such as Google and Meta are launching counterattacks against Nvidia, which has dominated the artificial intelligence (AI) chip market. In a bid to reduce dependence on Nvidia chips, they are accelerating their development of custom chips, and there is speculation that as early as next year, shipments of big tech's application-specific integrated circuits (ASIC) will surpass Nvidia's AI graphics processing units (GPU).
An application-specific integrated circuit (ASIC) is a semiconductor designed for specific applications, focusing performance on particular AI services instead of the versatility of AI GPUs. By optimizing hardware for necessary computations, it offers higher energy efficiency and significant expense savings compared to GPUs.
◇ 'ASIC total shipments will surpass Nvidia next year'
According to Nomura Securities in Japan on the 27th, Google is expected to ship between 1.5 million and 2 million of its own AI chips, Tensor Processing Units (TPU), and Amazon Web Services (AWS) is predicted to ship 1.4 million to 1.5 million ASICs. Even the combined shipments from these two corporations are nearing half of Nvidia's annual AI GPU supply (over 5 million to 6 million units).
With other big tech companies such as Meta also joining in earnest next year, there is a prevailing outlook that total ASIC shipments will exceed Nvidia's AI GPUs. Currently, Nvidia holds over 80% of the AI server market, and ASIC-based server market share is still only around 8% to 11%. However, when viewed through the lens of 'chip count' in terms of shipments, the market structure is changing rapidly.
Big tech companies are accelerating their development of custom chips largely due to the desire to escape the high expense structure referred to as the 'Nvidia tax.' Jensen Huang, CEO of Nvidia, emphasized the superiority of their AI chips at an event this month, stating, "If it’s not (performance-wise) better than buying, why would anyone make ASICs?" However, big tech is strengthening independent chip investments to secure long-term cost efficiency and supply stability.
The biggest advantage that big tech companies are focusing on is the expense savings from 'total cost of ownership (TCO).' The industry analyzes that ASICs can achieve a TCO reduction of 30% to 50% compared to equivalent AI GPUs. This can significantly reduce not only the purchase cost of GPU chips but also operational expenses, including power consumption. In fact, Google reported that its TPU has up to three times the power efficiency compared to Nvidia GPUs. Additionally, utilizing self-designed chips allows for stable procurement in line with service schedules, unaffected by external variables.
◇ From Google to Meta... Intensifying 'AI chip war' among big techs
Google is the pioneer in the development of its own AI chips. In 2016, Google first unveiled its dedicated AI chip, TPU, and has been applying it to its services. Currently, most big tech corporations, including Meta, Amazon, and Microsoft, have actively entered the ASIC development arena. JPMorgan estimates that the global AI ASIC market size will reach approximately $30 billion (about 41 trillion won) this year, and predicts it will grow by over 30% annually in the future.
Recently, the industry has been closely monitoring Meta's actions. Meta plans to release the next-generation high-performance ASIC chip 'MTIA T-V1,' designed by Broadcom, in the fourth quarter of this year. This chip aims to surpass the performance of Nvidia's next-generation AI GPU 'Rubin.' By mid-next year, they plan to introduce 'MTIA T-V1.5,' which will double the chip area, and by 2027, the 'MTIA T-V2,' which is expected to have overwhelming performance consuming the same amount of power as 50 households (170kW).
However, Meta's ambitious plans come with practical challenges. Meta aims to ship between 1 million and 1.5 million ASICs from the end of this year to next year, but the current supply capacity of advanced packaging (CoWoS) from TSMC stands at only 300,000 to 400,000 units, making supply bottlenecks inevitable.
Nvidia is not sitting idle either. Recently, it opened up its previously proprietary 'NVLink' protocol to the outside, allowing for easy integration of third-party central processing units (CPU) or ASICs with Nvidia GPUs. This is interpreted as a strategy to prevent full departures of big tech companies and encourage them to remain within the Nvidia ecosystem. The software ecosystem 'CUDA,' accumulated over decades, is also considered a unique weapon that big tech companies cannot easily match with just ASICs.
An industry insider stated, "When AI operates on customized chips optimized for each service, consumers will have a faster and more sophisticated experience," noting, "The reduction of dependence on Nvidia by big tech is an inevitable process as the AI industry matures, particularly given the accelerated rate of innovation in AI services."