Cognitive scientist Gary Marcus argued in a Financial Times opinion piece published around June 25, 2026, titled "How much compute does the world really need?", that merely adding GPUs, larger models and data centers cannot resolve AI's fundamental accuracy problem.
The Compute Arms Race · By 2030
A $6.7 Trillion Bet on Compute — and a Fight Over Whether It's Too Much
Data center investment could hit $6.7tn by 2030 as AI's hunger for compute soars. But power shortages, short-lived silicon and overbuild fears are sharpening the debate over how much the world actually needs.
$6.7tn
Cumulative data center investment by 2030 (base case)
156 GW
AI power demand in McKinsey's base case
+165%
Projected rise in data center power demand vs 2023
$300bn+
Combined Big Tech capital spending in 2025
How big could the spend get?
McKinsey's 2030 scenarios — total data center investment, drawn to scale.
Base case splits into AI $5.2tn + non-AI $1.5tn · Spend mix: tech/hardware 60% · power & cooling 25% · construction 15%
AI takes over the data center
By 2030, AI is projected to account for roughly 70% of global data center demand.
Inference is shifting to become the dominant workload, overtaking training.
The skeptics' math
At 100 GW scale, interest payments alone could demand ~$800bn a year .
5-year
amortization cycle
Short-lived
silicon raises overbuild risk
Optimists: demand is insatiable
NVIDIA's Jensen Huang calls the build-out the largest infrastructure project in human history. Falling training and inference costs only widen experimentation and adoption — pushing total demand higher.
Skeptics: beware the overbuild
IBM's Arvind Krishna and others warn against excessive expectations — echoing the dot-com fiber glut. Power-connection queues, unclear ROI and surging memory prices remain real obstacles.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…