Tomasz Tunguz (Theory Ventures) · 2025-05-29 · 371d

AI Demand Surge Driven by Reasoning Models: 1000x GPU Scaling Requirements

NVIDIA's latest earnings reveal a fundamental shift from simple one-shot AI inference to compute-intensive reasoning models, driving a 1000x increase in token generation and GPU deployment. Major hyperscalers are deploying 72,000 Blackwell GPUs weekly across nearly 100 AI factories, with over $300 billion in annual capex investment. While algorithmic improvements help manage demand, reasoning capabilities are outpacing efficiency gains.

7 metrics· Cited 0× in the knowledge base ·Open source ↗

Metrics in this report

AI Factory GPU Density Growth

2xmultiplier

year-over-year

Average GPUs per AI factory

AI Factory Growth

2xmultiplier

year-over-year

Number of NVIDIA-powered AI factories in deployment

Annual AI Infrastructure Capex

300$ billions

2025 estimated

Hyperscaler data center investment

Microsoft Blackwell GPU Deployment

100000+GPUs

expected range

GB200 ramp with OpenAI partnership

Token Inflation from Reasoning

100-1000xmultiplier

range

Tokens per task for reasoning vs one-shot inference

Token Processing Growth

5xmultiplier

year-over-year

Microsoft Q1 2025 total tokens processed

Weekly GPU Deployment Rate

72000GPUs

per week per hyperscaler

Blackwell GPU deployments across major cloud providers