AI factories reshape data center economics

In the race to power the next generation of AI, traditional data centers are facing a radical transformation. The hardware that trains and serves trillion-parameter models demands entirely new approaches to power, cooling, and management systems – creating what NVIDIA now calls "AI factories."

Key Points:

• Power is the new currency in data centers. As Charlie Bole, VP of DGX systems at NVIDIA explains, "Every single data center in the future will be power limited. Your revenues are power limited." Companies now calculate potential revenue based on available power, not physical space.

• New hardware designs focus on extreme density. NVIDIA's latest DGX GB300 systems pack 72 GPUs into a single chassis – the same capacity that previously required an entire rack. This density creates new challenges for cooling, with systems now requiring liquid cooling rather than traditional air cooling.

• Software automation has become critical. NVIDIA's Mission Control software handles everything from commissioning servers to optimizing workloads, applying automatic checkpoints, and even controlling liquid cooling pumps. This automation enables non-specialists to run complex AI infrastructure.

Why AI Factories Matter Now

The most compelling insight from NVIDIA's approach is how they've reimagined the economics of data centers around AI workloads. Traditional data centers required provisioning for peak power usage, meaning companies paid for electricity they rarely used. The new generation of DGX systems includes energy storage that smooths out power peaks, allowing companies to run at 100% capacity without overprovisioning.

This shift fundamentally changes the economics of AI deployment. As Bole notes, when you save power, "I can run more systems, more GPUs, and get more work done." For inference workloads especially, this translates directly to revenue: more power efficiency means generating more tokens and serving more users.

The implications extend beyond just NVIDIA's hardware. We're witnessing the birth of specialized infrastructure built specifically for AI that differs radically from general-purpose computing. Companies that fail to upgrade their infrastructure approach will be at a competitive disadvantage when deploying large language models and other AI services.

Real-world Applications Beyond NVIDIA's Ecosystem

While NVIDIA dominates the AI hardware conversation, similar transformations are happening across the industry. Microsoft's Azure team recently redes