Mustafa Suleyman: AI development won’t hit a wall anytime soon—here’s why

2026-04-08 · Source: MIT Technology Review · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Emerging Technologies & Innovation · Depth: Intermediate, short

Summary

The "compute explosion" describes the exponential growth in computational power driving modern AI, a trend that has seen the training data for frontier AI models increase by a trillion-fold, from 10¹⁴ flops in 2010 to over 10²⁶ flops today. This growth is fueled by three converging advances: significantly faster chips like Nvidia's, which increased raw performance sevenfold in six years, and Microsoft's Maia 200; high bandwidth memory (HBM) like HBM3, which triples data bandwidth to keep processors busy; and the development of warehouse-scale supercomputers connecting hundreds of thousands of GPUs via technologies like NVLink and InfiniBand. These hardware improvements, combined with software advancements that halve compute required for fixed performance every eight months, have led to a 50x improvement in training times for language models since 2020, far exceeding Moore's Law predictions. Projections indicate another 1,000x increase in effective compute by late 2028, potentially enabling human-level AI agents and transforming industries built on cognitive work, despite significant energy demands which are being addressed by rapidly falling solar and battery costs.

Key takeaway

For AI Architects and Directors of AI/ML evaluating future infrastructure investments, recognize that the compute explosion is not slowing; your planning should account for a projected 1,000x increase in effective compute by 2028. Prioritize scalable, energy-efficient hardware and software solutions that can integrate into warehouse-scale supercomputing environments to support the development of increasingly complex, human-level AI agents.

Key insights

Exponential compute growth, driven by hardware and software advances, is rapidly accelerating AI capabilities.

Principles

• AI compute growth far outpaces Moore's Law.
• Hardware and software co-evolve for compute gains.
• Energy costs are offset by renewable tech advancements.

In practice

• Utilize HBM3 for faster data delivery to GPUs.
• Deploy warehouse-scale GPU clusters for complex AI.
• Optimize software to reduce compute for fixed performance.

Topics

• Compute Explosion
• AI Hardware Acceleration
• Software Optimization
• Human-Level Agents
• Energy Constraint Mitigation

Best for: AI Architect, Director of AI/ML, CTO

Related on AIssential

• Energy! Chips! ...and INSURANCE? (WTF) — Physical infrastructure, supply chain, and regulatory hurdles are the primary brakes on AI acceleration, not capital or research.
• 🔮 Exponential View #571: DeepSeek shows the future, again; drones on a learning curve; solar goes up, LLM pixels & tennis robots++ — Rapid iteration and cost-efficiency are critical for technological advantage in both warfare and AI development.
• Bloomberg Tech at CES — AI compute demand is rapidly accelerating across cloud, enterprise, and consumer sectors, driving innovation in chip design and autonomous systems.
• 🔮 Exponential View #573: Are the AI labs building for an intelligence explosion? — A frontier AI model training its successor by 2028 is a 60% possibility, shifting AI development to an industrial scaling problem.
• 😸 You're behind on AI. Here's the recap — The AI landscape is rapidly advancing across agents, benchmarks, and hardware, driving significant shifts in market dynamics.
• The Download: AstroTurf wars and exponential AI growth — Exponential AI growth is driven by hardware advancements, while environmental and social impacts of technology spark debate.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MIT Technology Review.