Generative‑AI and agentic systems are hungry for compute and energy. Understanding the hardware landscape helps organisations plan for sustainability and scalability.
Hardware roadmaps – Keep an eye on NVIDIA’s H100/B100, AMD’s MI300 and new custom accelerators. Foundation models require memory bandwidth as much as compute; innovations in high‑bandwidth memory (HBM) and on‑package storage are critical.
AI supercomputers – MIT’s TX‑GAIN (TX‑Generative AI Next) is one of the most powerful university‑based AI supercomputers. It is powered by more than 600 NVIDIA GPU accelerators and delivers two AI exaflops—that’s two quintillion floating‑point operations per second. (news.mit.edu) Optimised for generative workloads, TX‑GAIN drives research in biodefense, materials discovery and cybersecurity. Its architecture allows pre‑training and fine‑tuning of foundation models, simulation of radar signatures, filling gaps in weather data and evaluating anomalies in supply chains. (news.mit.edu)
Compute per employee – At recent investor panels, start‑ups began reporting compute‑per‑employee and revenue‑per‑employee as key metrics. These reflect the capital intensity of running large models and encourage efficiency.
Energy and cooling – The price of AI is converging toward the price of energy. Data centers increasingly rely on direct‑liquid cooling and modular nuclear or renewable energy to power GPU clusters. Regulatory pressures (e.g., carbon taxes) will shape where and how AI models are hosted.
I’ll update this section on what I hear, and projects I am in collaboration within this domain.
therenaissancequant.com 