$769 billion in AI infrastructure investment expected by 2026

Summary

McKinsey's 2026 Technology Trends Report projects that global AI infrastructure spending will reach $769 billion. This increase is driven by three key factors: the shift from model training to daily inference processes, hyperscalers designing custom chips, and an accelerating cyber threat environment where AI-powered attacks exploit security vulnerabilities before they are publicly disclosed.

Why it matters

Massive capital expenditure, energy constraints, and a changing security landscape are converging. By 2030, AI inference is expected to account for 30-40% of data center electricity consumption. Additionally, the average exploitation time has dropped from 800 days to negative 16 hours. This indicates that the bottleneck is no longer model capability, but rather physical infrastructure, governance, and defense speed. The report warns companies that token processing costs are consuming one-third of digital development budgets, and one-third of adopting companies are experiencing efficiency losses due to weak governance.

Highlights

• Hyperscalers are designing custom ASICs with chip manufacturers to reduce energy consumption per token. This is reshaping data center cooling systems and physical architecture.
• AI-powered security vulnerability exploitation has dropped to a negative 16-hour timeframe. Over 75% of logged vulnerabilities are now zero-day in nature. In large enterprises, the number of non-human identities has reached 100 times that of human identities.
• Agentic software is projected to capture up to $1 trillion in enterprise value. However, one-third of adopting companies are experiencing efficiency declines due to unverified code and insufficient governance.

Critical take

This text is a subtle McKinsey press release disguised as news. Key figures, such as the $769 billion in spending, the $1 trillion value expansion, and the 55% efficiency increase, are based on McKinsey's own modeling rather than audited results. While the 'negative 16 hours' exploitation time cited from IBM is striking, it is a single data point. The lack of methodological transparency makes it difficult to distinguish the real signal from a narrative designed to sell the next consulting engagement.