AI’s explosive growth isn’t magic; it’s math. And lots of it. The latest deep learning breakthroughs, from large language models to generative image tools, need staggering amounts of computing power.
Tech giants including Microsoft and Meta are spending big on superclusters in hopes of AI breakthroughs
Photo by Steve Johnson on Unsplash
AI’s explosive growth isn’t magic; it’s math. And lots of it. The latest deep learning breakthroughs, from large language models to generative image tools, need staggering amounts of computing power.
Training a modern AI model is like teaching an elephant to dance: You need raw muscle, perfect timing, and endless repetition. That’s where AI superclusters come in.
On the surface, a supercluster is kind of like a giant brain made of many smaller brains. It's thousands, or even hundreds of thousands, of GPUs or AI accelerators wired together with ultra‑fast networking, custom‑built to tackle massive computations. Unlike your garage gaming rig or even your typical cloud server, these machines are built from the ground up for one purpose: AI.
AI has introduced a new class entirely when it comes to data storage and processing. While traditional data centers are designed to handle a wide mix of digital workloads (think email servers, website hosting, and cloud storage), AI superclusters are built with a singular purpose: processing massive amounts of data at lightning speed.
To understand why they matter, it helps to first see how they diverge from the data center blueprints of the past.
AI breakthroughs require both scale and speed. Big models need massive datasets, and bigger data means longer training times, unless you throw significant computing power at the problem. Superclusters allow:
Photo by Levart_Photographer on Unsplash
AI superclusters don’t come cheap or small. Building one often requires custom hardware, specialized networking infrastructure, real estate the size of an airport, and enough energy to power a small nation. Naturally, only a handful of companies are positioned to make this kind of investment.
These tech giants aren't just participating in AI — they're reshaping its physical foundations. Each one is investing heavily in dedicated superclusters that make it possible to train state-of-the-art models on previously unmanageable datasets. While some clusters power internal research, others underpin the AI services now reaching millions of users.
Here’s a breakdown of who’s building what and why:
Meta $META – Project Prometheus / Hyperion: Meta’s supercluster push is huge. It’s crafting Prometheus, a 1 GW AI data center, and planning Hyperion, a 5 GW facility approaching the size of Manhattan. It’s a key part of Meta's AI strategy.
Microsoft $MSFT / OpenAI – Azure ND H200 v5 clusters: Microsoft’s Azure ND H200 v5 VMs pack eight NVIDIA H200 GPUs each and can scale to clusters with thousands of GPUs. These systems power OpenAI’s model training as well as Azure AI workloads.
Google – TPU v4 / Ironwood Pods: Google’s TPU v4 Pods use 4,096 specialized chips and custom optical networks. They achieve exaflop-level speeds (an exaflop equals one quintillion calculations per second) and power their large-model training. Its latest “Ironwood” TPU v7 platform boosts scale, delivering tens of exaflops for generative inference systems.
Amazon $AMZN Web Services – EC2 UltraClusters & Project Rainier: AWS's UltraClusters are massive GPU pools that include P4d (A100 GPUs) and Trn2/Trainium2 chips with ultrafast networking. Project Rainier, built with Anthropic, uses hundreds of thousands of Trainium2 chips, delivering more than 5 times the exaflops used for current leading AI models.
Nvidia $NVDA – Partner-built supercomputing efforts: Nvidia orchestrates AI superclusters globally, sponsoring multi-thousand-GPU builds like Foxconn’s and collaborating with Oracle $ORCL, HPE, and others to deploy specialized clusters.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.