My Q1 work used at least ~30% less energy with AI than without

I completed 4.5 game prototypes, 3 web tools, and 2 video editors in 73 active days. Based on prior experience, comparable output without AI would have taken 8–10 months. Equivalent quality—matching the actual architecture and test coverage—would likely exceed 12 months.

That estimate is grounded in specific projects. I've been working in web and video game development for the last 15 years. One of the Q1 projects was a complex game prototype that was half completed at the start of 2026. Based on my experience developing Axe Ghost, a game with some similarities, I estimate that unassisted, the full prototype would have taken about ten months; the half I finished in 2026 would have consumed five months on its own. Another project was a reimplementation of an old web experiment, FaceMaker, based on crowd-sourced artificial selection. The first time around it cost me about two months. If I had reimplemented it by hand, with everything I learned the first time, I might have been done in three weeks. The AI-assisted reimplementation took two days. These two projects alone would have exceeded Q1.

Energy Use

The most defensible with-AI figure is day-based: 73 active days at ~2 kWh/day local workstation energy, plus 73 days of model energy at Simon Couch's median daily rate of 1.3 kWh/day. This avoids token-level energy weights entirely.

Component	Energy
Local workstation (~2 kWh × 73 days)	~145 kWh
Model energy (~1.3 kWh × 73 days)	~95 kWh
Total, day-based estimate	~240 kWh

A token-weighted estimate using Couch's category-specific weights gives a lower total:

Token type	Count	Energy
Input	613,172	~0 kWh
Output	706,652	~1 kWh
Cache creation	34,434,928	~17 kWh
Cache reads	441,891,742	~17 kWh
Claude total		~35 kWh

With local energy: ~145 kWh + ~35 kWh = ~180 kWh. I treat this as a speculative lower bound, not a more precise figure.

For the without-AI case: 8–10 months at 5 days/week ≈ 175–215 workdays, or ~350–430 kWh.

Scenario	Energy
With AI, day-based estimate	~240 kWh
With AI, token-weighted lower bound	~180 kWh
Without AI, 8 months	~350 kWh
Without AI, 10 months	~430 kWh

Caveats

Model energy. Couch's cache weights are extrapolated from Anthropic's sparse disclosures. Cache operations drive roughly 96% of the token-weighted model total. Given current disclosure quality, the model total could plausibly range from ~10 kWh to ~100 kWh. The day-based estimate (~95 kWh) sits at the high end of that range and is the more reliable anchor because it does not depend on cache weights.

Counterfactual time. The 8–10 month estimate reflects superficially similar output. The true no-AI baseline would likely be higher, not lower, because the counterfactual has no AI assisting with architecture, refactoring, and test coverage.

Workstation intensity. The 2 kWh/day figure is a midpoint drawn from component power estimates (laptop ~90–180 W, monitor ~33–75 W in typical use). A longer solo project might land above or below that midpoint depending on build/test intensity and research overhead.

Active days. The 73 with-AI days are any day with at least one logged Claude request, not necessarily full 8-hour sessions. The 175–215 no-AI days are estimated standard workdays. Both figures are uncertain: the with-AI local total could be lower if many sessions were short, while the no-AI total could be higher if a longer timeline involved more overhead.

Scope

This comparison is about energy, not carbon. Grid intensity varies by location (~300–400 gCO₂/kWh in the Netherlands), and Anthropic's data center energy mix is not disclosed. A carbon comparison could flip depending on where inference runs.

Method Notes

Workstation sources: Samsung energy sheet (33W SDR, 104W HDR), Tom's Hardware (75W at 200 nits), Notebookcheck Legion measurements (~150–230W under load).