Nvidia has been performing exceptionally well in recent times. Over the past 20 months, the increased demand for AI compute power from chatbots (such as ChatGPT) and Generative AI for Images has surpassed Nvidia’s production capabilities. Training models has become more challenging, and the adoption of AI technology has grown at a faster rate than before. Furthermore, the emergence of GAN for Video will only further intensify this demand.
However, Nvidia’s current situation is not entirely unexpected. A prime example is Apple, which made the decision to exclude Nvidia from their devices and cloud services since 2015. Instead, Apple developed its own M-series chips with unified memory architecture. Apple’s foray into chip manufacturing began in 2008 when they acquired P.A. Semi, and they have become proficient in chip production after a decade of experience. As a result, Apple is now confident enough to utilize their own chips across all their devices, including AI cloud servers. Similarly, other tech giants such as Google, Amazon, Meta, Microsoft, and others have also been investing in building their own AI chips for the past 5-10 years. It may still take a few more years before they become completely independent of Nvidia’s offerings. This is not to say that Nvidia cannot out-innovate them, but it is undoubtedly more challenging for Nvidia to maintain its leading position.
https://about.fb.com/news/2024/04/introducing-our-next-generation-infrastructure-for-ai
https://aws.amazon.com/blogs/aws-insights/why-purpose-built-artificial-intelligence-chips-may-be-key-to-your-generative-ai-strategy/
https://azure.microsoft.com/en-us/blog/azure-maia-for-the-era-of-ai-from-silicon-to-software-to-systems/
https://cloud.google.com/blog/products/compute/introducing-trillium-6th-gen-tpus/
https://seekingalpha.com/news/4030251-alibaba-unveils-open-source-risc-v-tech-chip-for-use-in-ai-cloud-data-centers
Nvidia’s dominance is not only threatened by competitors in terms of brute force compute power. There are also specialized ASICs (Application-Specific Integrated Circuits) and FPGAs (Field Programmable Gate Arrays) designed for specific AI models like transformers, which are surpassing even the top H100 systems. Furthermore, there is always the possibility of new and innovative models emerging that may require a different kind of hardware architecture altogether.
https://www.bloomberg.com/news/articles/2024-06-25/ai-chip-startup-etched-raises-120-million-to-expand-production
https://www.ndtv.com/artificial-intelligence/chatgpt-ai-chip-sohu-new-ai-chip-in-making-claims-it-will-revolutionise-chatgpt-6116225
It will be fascinating to observe how everything unfolds. As we reflect on the early days of the internet when billions were invested in laying optical fiber networks, let us enjoy the free services while we can, as eventually, they will find ways to monetize them.
Leave a Comment