Laura Katzman — Buoyant VC - Efficient Compute: Why Optimization Is Now Inevitable - October 2025
FreeEfficient Compute: Why Optimization Is Now Inevitable
FreeFree tier
About Laura Katzman — Buoyant VC - Efficient Compute: Why Optimization Is Now Inevitable - October 2025
This article by Laura Katzman of Buoyant Ventures argues that efficient compute is now inevitable due to the soaring electricity demands of AI, grid bottlenecks, and the fundamental distinction between training and inference workloads. Drawing on expert interviews from Microsoft, MIT, UC Berkeley, and others, it provides data-driven insights into how data center electricity consumption is projected to double by 2030, how infrastructure constraints are delaying projects, and why inference will outpace training as the dominant energy driver. The piece serves as both a market analysis and an investment thesis, highlighting the urgent need for optimization across the AI stack.
Key Features
Analysis of AI-driven electricity demand growth (4.4% to 12% of US electricity by 2028)
Explains the grid bottleneck: 7-year backlogs for new power connections
Distinguishes training vs. inference energy profiles for AI workloads
Provides investment perspective from a climate-tech VC firm
Cites expert sources from Microsoft, MIT, UC Berkeley, Salesforce, AWS, and data centers
Pros & Cons
Pros
- Grounds analysis in specific data points and projections (Goldman Sachs, expert interviews)
- Clearly articulates the underappreciated challenge of inference latency and always-on demand
- Connects climate-tech and AI in a timely, actionable framework
Cons
- Represents a single VC perspective (Buoyant Ventures) and may lack counterarguments
- Does not provide specific tool or technology recommendations for optimization
- Focuses primarily on U.S. and China markets, less on other regions
Best For
For investors evaluating climate-tech and AI infrastructure opportunitiesFor AI companies planning energy and deployment strategiesFor energy analysts and policymakers studying grid constraintsFor researchers exploring efficient compute solutions
FAQ
Why is efficient compute inevitable according to this article?
AI is driving unprecedented electricity demand, grid bottlenecks are delaying data center projects by 7+ years, and inference workloads (unlike training) are latency-sensitive and always-on, making optimization essential.
What is the difference between training and inference in terms of energy?
Training models like GPT-4 are resource-intensive but can be time-shifted and scheduled flexibly. Inference is always-on, latency-sensitive demand that is harder to optimize around grid constraints.
What are the projected growth rates for data center electricity consumption?
Data centers consumed 4.4% of U.S. electricity in 2023, projected to reach 6.7-12% by 2028. Globally, 1.5% in 2024 is expected to double by 2030, with 130% growth in the U.S. and 170% in China.