AI Cluster Scaling
Overview
AI cluster scaling refers to the empirical trend of exponentially increasing compute requirements for training frontier AI models. Cisco's analysis at OCP Global Summit 2025 quantified the trajectory of OpenAI's GPT series between 2019 and 2025:
| Model | Parameters | FLOPs | GPU Count | MMLU Score |
|---|---|---|---|---|
| GPT-2 (2019) | Baseline | 1x | ~10–100 | — |
| GPT-5 (2025) | ~3,000x | ~100,000x | ~50,000–100,000 | ~2.5x |
Despite 3,000x parameter scaling and 100,000x FLOP scaling, reasoning ability (measured by MMLU) improved only 2.5x — illustrating the diminishing returns at the frontier and the massive infrastructure cost imposed by continued scaling.
The practical implication is that cluster sizes now exceed the power capacity of single data centers. A two-tier network can connect 16,000 GPUs at 22 MW; a three-tier clos fabric using today's 51.2T switches could theoretically reach 33 million GPUs but would require 44 GW — physically and economically untenable. GPT-5 training is already in the regime requiring multi-datacenter clusters, which in turn requires the scale-across network tier.
Sign in to read the full article.
Sign In