Dylan Patel, founder of semiconductor research firm SemiAnalysis, laid out a striking prediction on a recent podcast: by 2028, OpenAI and Anthropic could account for 70 to 80 percent of all new AI compute capacity worldwide — a combined buildout of more than 100 gigawatts. The two labs already take up around 30 percent of annual new compute, Patel said, and that share is climbing fast.
Patel argues the business model for frontier AI labs is changing as compute output efficiency improves dramatically. Anthropic is already generating roughly $50 million in revenue per megawatt of compute, and that number could rise to $100 million. That's what allows OpenAI and Anthropic to pay premium prices — $25 million to $50 million per megawatt — to buy or lease compute.
More striking: Patel estimates global AI-related capital spending will reach about $11 trillion between 2024 and 2029, with more than $5 trillion of that financed through debt. Because AI infrastructure's potential returns are far higher than traditional industries, tech giants may accept higher financing costs, pushing up overall credit rates and squeezing valuations of traditional assets and heavily indebted economies.
Patel also believes future compute additions won't necessarily go primarily toward serving external model inference requests. Instead, more capacity may flow into internal R&D and model self-improvement.