Jensen Huang claims 19 days tops a year for others

NVIDIA CEO Jensen Huang stated that Elon Musk completed in 19 days what others would take a year, calling it incredible. The comparison is with xAI building the training cluster Colossus in Memphis.

Huang broke down the typical timeline on the BG2 podcast: supercomputers usually require three years of planning, followed by another year for installation, power-up, and debugging before the first training. He described the period from xAI's concept to liquid cooling facility approval, power-up, connecting NVIDIA equipment, and running the first training as 19 days, claiming that only Musk could achieve this.

The cluster consists of 100,000 NVIDIA H100s, which he referred to as the fastest supercomputer in a single cluster at that time. He also acknowledged xAI's engineering, software, networking, and infrastructure teams, noting that it was likely no one slept during those 19 days. NVIDIA provided network-side support, as its equipment is more challenging to install than traditional data center servers.

There is not just one timeline. Musk told Jordan Peterson that from hardware installation to starting training took 19 days, claiming it was the fastest known. In another public statement, he noted that the cluster took 122 days from going online to running successfully. The 19 days does not encompass the entire calendar from land acquisition to training.

xAI was established in 2023 and had already released Grok-2. Huang attributed the speed to an understanding of engineering, construction, large systems, and resource scheduling, rather than the specifications of a particular chip.

The buyer is xAI, which needs a trainable cluster within weeks, while the seller is NVIDIA, providing H100s and supporting networks. The event was driven by the speed of construction. Funding has shifted from annually planned cloud vendor data centers to teams that can simultaneously mobilize permits, power, liquid cooling, and chips, benefiting model companies and chip suppliers that can run training first, while putting pressure on data center projects still following a three-year plan and one-year debugging schedule.

Source: Public information