The development of artificial intelligence raises many challenges in terms of infrastructure. Octave Klaba, CEO of OVHcloud, sheds valuable light on the stakes involved in datacenters dedicated to AI, particularly concerning plans to build 1GW datacenters in France and the issues this will raise.
Proportional logic Euros / Megawatt
Investment in AI infrastructure follows a clear proportional scale. To give you an idea, an investment of 50 million euros will buy 1,000 GPUs consuming 1MW. Following this logic, to reach a power of 1GW, the investment required amounts to 50 billion euros for one million GPUs. 80% of this investment is devoted to the purchase of GPUs, while 20% goes towards building the datacenter.
Superchips to boost GPUs staking
Today’s GPUs have a significant physical limitation: the distance between processors can slow down performance beyond 100,000 connected GPUs. To meet this challenge, a new generation of “superchips” is emerging, integrating several GPUs on a single board. While this innovation makes it possible to connect 10 times as many GPUs, it also raises new challenges in terms of power distribution and cooling of denser infrastructures.
Watercooling in every rack
Traditional data centers use standard air-cooled racks, consuming around 20KW. To meet new requirements, these infrastructures are evolving towards much more powerful racks, consuming between 120KW and 240KW. This significant increase in power requires the widespread adoption of watercooling and the development of new datacenter architectures.
For inference, requirements are more modest, with consumption varying between 100W and 10KW per system. This requires fewer GPUs to be connected, and enables a multiplication of systems in parallel. The main challenge lies in the geographical distribution of datacenters to optimize latency and ensure high availability.
OVHcloud’s position
OVHcloud is a major player in this field, with over 40 datacenters worldwide. Backed by 20 years’ expertise in watercooling, the French group cools over 500,000 physical servers using proprietary solutions that are far less expensive than those on the market. Its 40MW to 100MW datacenters are adapted to training, while its global infrastructure efficiently handles inference requirements.
This analysis by Octave Klaba highlights the considerable stakes involved in the development of AI infrastructures, in terms of both technology and energy, and underlines OVHcloud’s strategic position in this rapidly evolving field.
It’s certainly also a good way for OVH to gain market share, as it can carve out a niche for itself by remaining neutral on its positions and focusing on resources, not models. Traditional network requirements are also to be expected, with the amplification of network calls, models that easily weigh 500 GB, and the need for solid nodes to distribute all this.
Source : https://twitter.com/olesovhcom/status/1889231304543130043