Atlas · frontier
Chips and data centers
Models run on accelerators with fast memory, packed into buildings measured in megawatts.
See it working: Plate V
In plain words
The model you talk to runs on chips built for multiplying large tables of numbers, each with a pool of very fast memory.
How it works
Memory capacity limits how many weights fit on a chip, and memory bandwidth is one factor in how fast it runs. NVIDIA lists the H100 SXM at 80GB and 3.35TB/s, and the B200 at 180GB and up to 8TB/s. Google lists its TPU 8t at 216 GB and 6,528 GB/s. For comparison, Apple lists the M5 Pro Mac mini at 307GB/s.
Going deeper
Epoch AI estimates that the record IT power capacity of a single AI data center has doubled every ten months since mid-2024, reaching about 950 MW at xAI’s Colossus 2. IT power leaves out cooling, and these are Epoch’s estimates, not the companies’. The memory calculator on Plate V puts these chips beside a Mac.
Sources
- NVIDIA H100 Tensor Core GPU NVIDIA
80GB
3.35TB/s
3,958 teraFLOPS
With sparsity
Up to 700W (configurable)
- Components — NVIDIA HGX AI Factory (HGX H100, H200, B200) NVIDIA
180GB HBM3e
Up to 8TB/s
based on eight B200 GPUs, that delivers 144 petaFLOPs of AI performance.
- Inside the eighth-generation TPU: An architecture deep dive Google Cloud
216 GB
12.6
6,528 GB/s
- The record for power capacity in a single data center has doubled every 10 months Epoch AI
Since mid-2024, the record for the largest AI data center by IT power capacity has doubled every ten months. The current record-holder is xAI’s Colossus 2, at an estimated 950 MW.
IT power is the power capacity available to chips, servers, and other IT equipment, excluding cooling and other overhead.
- Mac mini (M6 or M5 Pro) - Tech Specs Apple
307GB/s memory bandwidth
24GB unified memory
48GB or 64GB
Quotes were checked against the live pages on 2026-11-14.