NVIDIA's Ambition: "AI Native" Completely Disrupts Data Centers

Investing millions of dollars in purchasing the previous generation of computing resources is simply foolish. As NVIDIA CEO Jensen Huang pointed out at Tuesday's press conference, it is much more cost-effective and energy-efficient to use the latest AI-specific development hardware, GH200.

Yesterday, NVIDIA unveiled the new generation GH200 Grace Hopper superchip platform, designed specifically for accelerating computation and the era of generative AI.

Huang Renxun emphasized that in order to meet the growing demand for generative AI, data centers need accelerated computing platforms tailored to these specific needs. The new GH200 chip platform offers exceptional memory technology and bandwidth, enhancing the ability to connect GPUs without loss of performance, and features server designs that can be easily deployed throughout the entire data center.

It is worth mentioning that with the advent of large-scale models, various AI-native applications have emerged, driving a surge in computing power demand and rapidly growing the market for data centers specifically designed for data-intensive AI applications.

New Transformations in Data Centers

According to The Wall Street Journal, analysts point out that as established cloud computing providers race to upgrade data centers with advanced chips and other enhancements to meet the demands of AI software, some emerging players see an opportunity to build new facilities from scratch.

Data centers are like large warehouses equipped with multiple servers, networks, and storage devices for storing and processing data. Compared to traditional data centers, AI data centers have more servers using high-performance chips, resulting in an average power consumption of 50 kilowatts or more per rack, compared to approximately 7 kilowatts per rack in traditional data centers.

This means that AI data centers need to expand their infrastructure to provide higher power capacity. Additionally, due to the increased power consumption, AI data centers require additional cooling methods, such as liquid cooling systems, to prevent overheating of the equipment.

Manju Naglapur, Senior Vice President at service and consulting company Unisys, pointed out:

Dedicated AI data centers can accommodate servers utilizing AI chips, such as NVIDIA's GPUs, allowing for simultaneous execution of multiple computations when processing vast amounts of data for AI applications. These data centers are also equipped with fiber optic networks and more efficient storage devices to support large-scale AI models.

AI data centers are highly specialized buildings that require substantial investment of funds and time. According to Data Bridge Market Research, global spending on AI infrastructure is projected to reach $422.55 billion by 2029, with a compound annual growth rate of 44% over the next six years. DataBank CEO Raul Martynek said that the deployment speed of artificial intelligence is likely to lead to a shortage of data center capacity in the next 12 to 24 months.

AI Computing Power Rising Star Raises $2.3 Billion in Financing

Currently, various industry giants are betting on AI data centers, such as "real estate benchmark" Blackstone selling houses to invest in AI data centers. Meta has also stated that it will build a new artificial intelligence data center.

As mentioned in a previous article, AI computing power rising star CoreWeave secured a $2.3 billion (approximately 16.5 billion RMB) debt financing by mortgaging NVIDIA H100.

CoreWeave stated that this funding will be used to accelerate the construction of artificial intelligence data centers, following the $221 million raised in April and $200 million raised in May this year. CoreWeave was established six years ago and currently has seven artificial intelligence data centers online, with plans to double that number by the end of this year.

CoreWeave is collaborating with NVIDIA and Inflection AI to build a super-large AI server cluster with a goal of running 22,000 NVIDIA H100s. If completed, it will become the world's largest AI server cluster.

It is worth mentioning that according to CoreWeave's official website, their services are 80% cheaper than traditional cloud computing providers. The latest NVIDIA HGX H100 server, which includes 8 H100s with 80GB of VRAM and 1TB of memory, starts at just $2.23 per hour (16 RMB).

Compared to the previous generation platform, the new GH200 Grace Hopper platform doubles the memory capacity and triples the bandwidth with its dual-chip configuration. Each server is equipped with 144 Arm Neoverse high-performance cores, 8 petaflops of AI performance, and the latest HBM3e memory technology with 282GB.

No wonder, in this era of the AI explosion, Huang Renxun boldly declares, "The more you buy, the more you save!"