Evaluating the environmental impact of patent AI.

How long is a piece of string? Evaluating the environmental impact of patent AI.

The energy savings of AI - assisted patent drafting

I started with a basic observation from my own work.  If AI is used to assist patent application drafting, it takes less attorney time, something we pass on in savings with our Visser AI work.  I can easily quantify the IT energy saved.  How does that compare to the energy used by the AI system in its work?

Starting with the easy bit, from my experience, AI assisted patent drafting can save at least 5 hours.  It may well be more, but this shorter time makes it harder for the AI to justify itself.  Typical at wall power for a docking monitor with my laptop attached and drawing power from the monitor is about 70W.  Patent drafting work really benefits from a bigger screen, and I am nerdy enough that I’ve measured this! So, 5 hours of work is 350 Wh, or around 1,260 kJ.

That number can be hard to visualise. You may be surprised to know it’s as much energy as there is in 56g of Cadbury Dairy Milk. IT uses a lot of energy!

How much energy does patent drafting AI use?

The AI energy usage side of the equation is not as simple.  As many readers will know, AI uses tokens, both input and output to do its work.  The number of tokens depends on the context you give the AI to work with in the first place (e.g. the invention disclosure and related documents) and the overall length of the output.  Here I’m going to estimate 1,000,000 tokens for my example patent draft. It sounds a lot, when a consensus is that one English word is about 1.4 tokens.  This estimate is based on the reality of the AI patent drafting process with use in Visser AI.  That includes iterations with AI checking and refining multiple aspects to review and strengthen the draft in ways not previously possible within an acceptable time frame. Those iterations rapidly increase the number of tokens.

Estimating energy consumption per token

Next, we need the energy per token, which is harder to find than I expected.  It’s also a fast-moving target as hardware is evolving so rapidly in terms of energy efficiency.  In its blog post in March 2026, Nvidia highlighted how inference throughput per MW has improved by 1,000,000 times  in the six hardware generations from 2012 to 2026.  But trying to pin down an energy use per token is a challenge.  A widely cited paper from 2023 presented the energy use as around 4 J per token . That testing used Volta and Ampere GPUs, while the latest Blackwell and Rubin architectures are around 1000x more efficient at a hardware level according to that Nvidia blog.

Comparing AI energy use with traditional patent drafting

Pulling all this together, we get around 4,000 kJ assuming the models run on older Nvidia hardware, and 4 kJ if the 1,000x hardware efficiency of newer generations is fully realized (probably not realistic).  Recent Microsoft research published in Joule (April 2026) suggests that the whole “energy per token” approach is too simplified and “Frontier-scale inference” at the query level is perhaps 4-20x lower than public estimates .

So, with that 4J/token as the “public estimate” we end up with an AI energy usage of between 200 and 1,000 kJ for our patent draft using that, or between 9g and 45g of Dairy Milk.  A clear answer is basically impossible unless and until we switch to local models to measure at wall power consumption.  In the worst case, the patent draft uses slightly less energy than is saved by not using the laptop, but it could also be as much as 6 times lower!  

Of course, I haven’t considered the energy used to build the laptop / AI data centre or other factors such as water usage and training energy consumption.  Converting to a CO2e would be interesting too, but that varies by time of day with grid generation mixes.  But this gives us a point of comparison between two very different workflows.

Is patent AI more sustainable than you think?

To conclude, AI usage for patent drafting probably doesn’t use as much energy as you thought and is already likely using less energy than a non AI-assisted draft. Economic incentives will only reduce energy consumption over time, and the gains for AI hardware are much faster than that for relatively mature general computing hardware.  Or perhaps the drafting attorney could just cut down on their chocolate consumption to offset the use of AI.

Author: Laurence Brown - Head of Visser AI, Partner, UK & European Patent Attorney