Typically, the most highly desired hardware consists of phones or gaming consoles that are constantly out of stock. However, this year, it appeared that everyone in the technology industry was willing to patiently wait for several months and spend a significant amount of money on a product that is unlikely to be available for purchase in stores or even physically seen: Nvidia Corp.'s H100 artificial intelligence accelerator.
Nvidia's chip has indisputably emerged as the most essential technological component driving the AI revolution. The H100, equipped with 80 billion transistors, is the preferred choice for training extensive language models that support applications such as OpenAI's ChatGPT. Furthermore, it has played a significant role in Nvidia's dominance in the AI chip market.
However, due to the strong demand for the H100 and the lack of competitive chips from Advanced Micro Devices Inc. and Intel Corp., major tech companies in 2023 were compelled to invest significant amounts of money in a competitive race to develop better processors. Currently, a solitary H100 is priced at $57,000 on the internet store of hardware provider CDW. Moreover, data centers are abundantly stocked with numerous units of this product.
In 2016, Nvidia's CEO Jensen Huang provided OpenAI with the company's initial AI server equipped with earlier generation graphic processing units. At that time, it was difficult to anticipate the significant impact these types of processors would have in the subsequent revolution triggered by ChatGPT. Nvidia's graphics cards were mostly associated with video games, rather than being recognized for their application in machine learning. However, Huang quickly realized that their distinctive architecture, proficient in parallel computing, would be more suitable for efficiently processing the vast and simultaneous data required by AI models, in contrast to conventional computer processors such as those made by Intel.
Microsoft Corp., as a stakeholder in OpenAI, constructed a supercomputer utilizing around 20,000 Nvidia A100 GPUs, which are the predecessors of the H100. Amazon.com Inc., Google (owned by Alphabet Inc.), Oracle Corp., and Meta Platforms Inc. have recently made significant orders for H100s to expand their cloud infrastructure and data centers, which Huang now refers to as "AI factories." Chinese enterprises actively sought to accumulate lower-performing versions of Nvidia's GPUs, as their capabilities were constrained by US export restrictions on semiconductors. Delivery times for chips may exceed six months. "Obtaining GPUs at this stage is significantly more challenging than acquiring drugs," Elon Musk humorously remarked earlier this spring.
Setting aside complaints, Nvidia's blockbuster product line propelled its worth to over $1 trillion and resulted in a significant increase in sales. The data center division's sales in the most recent quarter amounted to $14.5 billion, which is over four times higher than the sales during the corresponding period in the previous year.
However, the GPU bottleneck has also alerted the industry to the dangers of depending on a sole business for such a crucial element of their AI portfolios. Google has made substantial investments in its in-house TPU processors in an effort to reduce expenses and achieve performance improvements. Additionally, Amazon and Microsoft have lately showcased their own specialized AI accelerators. Intel is promoting its Gaudi 2 CPU as a substitute for the H100, while AMD has said that its new MI300 will allow it to enter the AI chip industry, which is estimated to be worth $400 billion in the future.
For certain prominent technology companies, the transition to using their own custom-designed microprocessors might potentially create a complex relationship where they are both friends and competitors, especially if their products become popular. Amazon and Google aim to strike a balance between avoiding excessive dependence on Nvidia while also preserving their partnership with the leading chip manufacturer, which could impact their future access to cutting-edge GPUs. Huang expressed to Bloomberg earlier this year that he is indifferent to the possibility of his largest clients potentially becoming his rivals, and he will not treat them any differently as a result.
Comentários