← Archive

Compute Is Becoming a Commodity — Slowly, and Expensively

by ·July 24, 2026·1 min read·Technology & Platforms
इस निबंध का पूरा हिंदी अनुवाद अभी तैयार नहीं है — नीचे का लेख अंग्रेज़ी में है। चित्रों के लेबल और साइट का बाकी हिस्सा हिंदी में दिख रहा है।
AI Applications Model Training & Inference GPU Datacenters Capital & Power Supply

It's tempting to talk about compute the way we talk about software — as something that gets cheaper and more abundant automatically over time. Compute at the frontier doesn't behave like that. Every leap in model capability has been paired with a leap in capital expenditure, and the companies training frontier models are increasingly capital-intensive utilities wearing a software company's branding.

This matters because capital intensity changes who can compete. Software has historically rewarded small teams with good ideas; frontier compute rewards balance sheets. That's a structural shift in who gets to set the pace of an entire industry, and it's easy to miss if you're only reading model benchmarks instead of capital expenditure disclosures.

The interesting long-term question isn't which lab has the best model this quarter — it's whether compute costs fall fast enough, for long enough, to eventually commoditize the layer the way cloud computing commoditized servers. Until then, this is a capital game as much as a research one.

Dr Nadeem Khudboddin Shaikh
Dr Nadeem Khudboddin Shaikh
Ex–Wells Fargo · Ex–Goldman Sachs · Columbia University alumnus