The AI community has been obsessed with bigger models and more data centers. But researchers at Stanford’s Hazy Research Lab are proposing we optimize for something entirely different.
They’ve introduced Intelligence per watt (IPW)—a new metric that fundamentally reframes how we should think about AI utilization in an era of exploding demand. Their paper breaks down the challenge and the opportunity, pointing us toward a compelling path forward for future research and innovation.
Efficiency is critical to meet ever-growing demand
Demand for AI computation is growing exponentially, with Google reporting an 8.1x increase in tokens processed per month from February 2024 to October 2025. However, the Hazy Research team also observes that internal ChatGPT telemetry data shows 77% of requests are practical tasks like writing emails or summarizing documents. In other words, for well over three fourths of real-world AI usage, we’re shipping routine queries–requests that could be answered accurately on the local device–to frontier-level models in datacenters.
History offers a better path. From 1946-2009, computing efficiency doubled every 1.5 years, shifting workloads from mainframes to PCs. PCs won not through raw performance, but because efficiency improvements made computing capable enough within personal device power constraints.
We’re at that same inflection point with AI inference. Can we get more of our needs met on the edge, where power efficiency is greater and the absolute maximum AI reasoning capabilities are unnecessary? Can the exponential growth in demand for AI be met more effectively through better leverage of the devices in our pockets and backpacks? The Hazy Research team says yes!
Hazy Research’s intelligence per watt
The Hazy Research team defined IPW elegantly:
IPW = (mean accuracy across tasks) / (mean power draw during inference)
Their empirical study—20+ local models, diverse hardware, 1 million real-world queries—reveals three key findings:
- Local LMs accurately respond to 88.7% of single-turn queries, with accuracy improving 3.1× from 2023-2025
- Local accelerators have significant efficiency headroom—the M4 Max achieves 1.5× lower IPW than NVIDIA B200 for the same model
- Intelligence efficiency has improved 5.3× over the past two years through combined model and hardware advances
Snorkel AI’s contribution to the IPW initiative
At Snorkel AI, we’ve built benchmarks to evaluate frontier LLMs across expert-level, domain-specific tasks using our Expert Data-as-a-Service—powered by a global network of specialists across thousands of domains.
We’re excited to contribute these specialized datasets to Hazy Research Lab’s Intelligence Per Watt initiative. While their foundational work focused on general chat and reasoning, real-world deployment demands domain-specific evaluation.
By combining Hazy Research’s IPW measurement framework with Snorkel’s industry-relevant benchmarks—spanning insurance underwriting, financial analysis, legal review, and PhD-level technical domains—we can drive an industry-wide shift in how we approach AI’s compute needs.
This partnership will answer critical questions: How efficiently can local models handle medical reasoning? What’s the IPW for regulatory compliance tasks? Can edge devices deliver expert-level performance within power budgets?
The path forward
Hazy Research’s Intelligence Per Watt metric should guide AI’s transition to the edge, just as performance-per-watt guided the mainframe-to-PC shift. They’re releasing a hardware-agnostic profiling harness to make IPW measurement systematic and accessible.
The future of AI isn’t just bigger models—it’s smarter systems delivering the right intelligence, in the right place, with the right efficiency. Snorkel AI is proud to support this vision with specialized datasets that ensure IPW becomes an important consideration for real-world enterprise deployment.
Read the full paper here and check out hazyresearch.stanford.edu for more information about Stanford University’s Hazy Research Lab, headed by Snorkel AI cofounder Chris Ré. Learn more about Snorkel AI’s data-centric approach at snorkel.ai.
Recommended articles
View all articles
The Standard for Agents You Can Trust: Lessons from the Federal Front Lines
In the first installment of Agentic in Action — a series about real AI deployments, not demos — Snorkel AI’s Kevin Olivieri sat down with three people who have spent their careers where trust isn’t optional: Chris Sniffen, Federal Applied AI Lead at Snorkel AI; John Hickey, President of August Schell; and Mike Baca, CIO of August Schell. The conversation focused on
June 5, 2026
•
Snorkel Team
Collaborative Gym: A Framework for Enabling and Evaluating Human-Agent Collaboration
At our latest Snorkel AI Reading Group, Yijia Shao (Stanford NLP) stopped by our San Francisco office to present Collaborative Gym: A Framework for Enabling and Evaluating Human-Agent Collaboration. As LLM agents get better at automating tasks on their own, a large class of real-world problems still needs a human in the loop – for their preferences, their domain expertise, or simply for control.
June 4, 2026
•
Benchtalks #2: The future of coding benchmarks
For our second Benchtalks, the series dedicated to the researchers building the measurement toolkits that frontier labs hill-climb on, Snorkel AI co-founder Vincent Sunn Chen sat down with John Yang, a Stanford PhD student and creator of the SWE-bench franchise, SWE-smith, CodeClash, and most recently ProgramBench. Highlights More on ProgramBench: See the benchmark and the upcoming leaderboard at programbench.com. More from John Yang: Publications and writing at john-b-yang.github.io. Snorkel
June 3, 2026
•
