Percy's energy use estimate

Supporting documentation for the carbon-footprint question in our FAQ. If you're new to Percy, percy.garden is a better starting place.

This document explains how we estimate Percy's energy use across a full growing season. It's the methodology behind the figures referenced in our FAQ ("Does using AI mean Percy has a big carbon footprint?").

We've published this because we'd rather show the math than ask people to trust a tidy number. AI energy estimates can be flattering or unflattering depending on which assumptions you pick; here are ours, and where they could be wrong.

This isn't an energy research paper. It's a small team's best attempt to answer a question we wanted to be able to answer ourselves. The figures are estimates assembled from public research on similar models; they're not measurements, and they're not audited. We've published the methodology so anyone curious can see where the numbers come from and judge them for themselves.

What we're estimating

Percy's per-user energy footprint over a single growing season (roughly April through October in the US), measured in watt-hours of electricity consumed across the cloud services that run Percy. We don't try to estimate the energy used by your phone running the Percy app, your home Wi-Fi, or the upstream cost of model training — those are real but are common to nearly every app and not specific to Percy.

We use optimistic (high) usage assumptions deliberately. It's better to overestimate Percy's energy footprint and be honest than to lowball it.

What "an interaction" means

Throughout this document, "interaction" means a single user-initiated action that triggers at least one AI call. That covers:

• Logging an observation by voice (transcription + parsing)
• Logging an observation by text (parsing only)
• Identifying a plant from a photo
• Asking Percy a gardening question
• Reviewing or correcting an entry

A single observation can trigger several backend calls (transcription, parsing, photo screening, optional plant ID), but we count it as one interaction for the user-facing total. Photo screening and the once-a-year report are counted as their own lines below, because they don't map one-to-one onto interactions.

Percy's model stack

Different tasks use different models, deliberately chosen to match the work:

The biggest lever on Percy's energy footprint is still the same: a compact model for everything that happens many times a day, and larger models only where the call is rare and the accuracy matters.

Per-interaction energy estimates

Cloud AI providers don't publish per-query energy for the models we use. The figures below are best-available estimates drawn from independent research on equivalent model classes, then applied to Percy's actual call patterns. We use the upper end of published ranges where the science is uncertain, to avoid undercounting.

Plant ID is still the heaviest thing you do day to day. A single plant ID is roughly 10–20× a voice observation. The plant-ID model was updated to a newer version in the same class this fall. Nobody publishes per-query energy for either version, so we've left the estimate unchanged rather than claim a saving we can't show.

The season report is the heaviest single event, but it happens once. Large models use several times the energy of mid-tier ones per call. Percy keeps those calls short (a few hundred words of output each, written from a summary of your season rather than every entry) and makes only three of them per person per year. Each passage is then checked against your actual journal by a compact model before it's used.

Typical seasonal usage assumption

A representative engaged gardener over a six-month growing season (using deliberately optimistic, high-end usage numbers):

Plus, per season: about 65 photos added to entries (each one screened), about 13 Garden Letters, and one season report.

For context, this assumes roughly one or two interactions per day on average across the growing season, with heavier days during planting and harvest. Light users do considerably less. The most engaged power users do more. We use the optimistic number to avoid undercounting.

Seasonal energy total

Applying the per-interaction figures to optimistic seasonal usage:

Plant identification is still the largest share: about half of Percy's AI energy, from 15% of interactions. If we used a smaller model for plant ID, the total would drop significantly, but at a real ecological cost in misidentifications. The season report is the second-largest line, about a sixth of the total. Both are deliberate tradeoffs.

What else does Percy run on?

The numbers above cover the AI services that do the heaviest work. Percy also runs on a small set of supporting cloud services for hosting, data storage, scheduled jobs, and operations. We estimated each one separately to make sure nothing meaningful was being hidden in a catch-all line:

Adding this to the ~80 Wh from AI brings Percy's total estimated seasonal energy footprint to roughly 100 Wh per user (we'd put the honest range at about 85–115 Wh). That's about 11 hours of a 9-watt LED bulb: one bulb left on for the day.

These figures are necessarily rougher than the AI estimates. Shared infrastructure like a managed database has a fixed baseline cost regardless of how many users it serves, so the per-user figure goes down as we grow. Today's per-user infrastructure cost at our current user count is higher than it would be at scale; the figures above reflect today's count, which is the more honest framing for now.

Converting to carbon

To translate watt-hours into greenhouse gas emissions, we apply the US national average grid carbon intensity from the EPA's eGRID dataset:

• US grid average: approximately 350 g CO₂-equivalent per kWh (EPA eGRID2023, the most recent year published as of this update)
• Percy seasonal energy: ~100 Wh = 0.100 kWh
• Percy seasonal carbon: ~35 g CO₂-equivalent per user per season

Regional variation matters: a user on a coal-heavy grid (some Midwest and Mountain West areas) might be at 600+ g/kWh, while a user on a hydro- or nuclear-heavy grid (Pacific Northwest, parts of New England) could be at 50–100 g/kWh. We use the national average for the public-facing figure.

The garden center comparison

We say a round trip to the garden center produces "more than a hundred times more carbon than Percy uses in an entire season." Here's that math:

• EPA estimates a typical passenger vehicle emits roughly 400 g CO₂ per mile
• A 10-mile round trip: 4,000 g CO₂
• Versus Percy's seasonal ~35 g CO₂
• Ratio: ~115× more carbon for one garden center trip than Percy's entire season

A 15-mile round trip is roughly 170× Percy's seasonal footprint. "More than a hundred times" holds for any round trip of about 10 miles or more; we don't say "hundreds of times," because for a typical trip it isn't.

Sources of uncertainty

Honest places where this estimate could be off:

1. Per-query AI energy is the largest unknown. The providers of the models Percy uses don't publish it, and independent estimates vary by an order of magnitude. We use figures from published research on equivalent model classes (MLPerf benchmarks; independent measurements of comparable LLMs and image classifiers).
2. Large models have the widest ranges of all. Independent estimates for the largest models run from about 1 Wh to well over 10 Wh per call, depending mostly on how much text goes in and comes out. Our 8–20 Wh for the season report assumes three short calls at the high end of that range, plus checks. If those estimates turn out to be low, this is the line that moves most.
3. Some providers now publish median per-prompt figures for their own chat assistants, generally a few tenths of a watt-hour including data-center overhead. Those are for general-purpose assistants, not the compact models Percy uses for routine work, but they suggest our compact-model figures are in the right range and not flattering.
4. Model efficiency is improving. Today's models are several times more energy-efficient than the same class of model two years ago. If our model providers ship efficiency gains, these numbers come down.
5. The 200-interactions assumption is deliberately high. A heavy power user might log 400+ interactions a season; a typical user logs closer to 100–150. Our copy uses the optimistic number, not the average. The season report is the exception: it costs about the same whether you logged 20 entries or 400, so for light users it's a larger share of their total.
6. Training cost is not included. The energy used to train the AI models Percy depends on is real, but it's spread across all users of those models worldwide (often billions of queries). The marginal training cost per Percy user is small but nonzero. We exclude it because we can't measure it cleanly. The same goes for the one-time work of creating the report's plant illustrations, which are drawn from a shared library rather than made fresh for each person.
7. Embodied energy of phones and infrastructure is excluded. This is standard practice for software carbon estimates, but it's worth knowing.

If any of these inputs are materially wrong in either direction, we'd want to know. Email us at hello@percy.garden if you spot an error.

Sources

• Grid carbon intensity: US Environmental Protection Agency, Emissions & Generation Resource Integrated Database (eGRID), 2023 data. https://www.epa.gov/egrid
• Vehicle emissions: US Environmental Protection Agency, Greenhouse Gas Emissions from a Typical Passenger Vehicle. https://www.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle
• Electricity reference data: US Energy Information Administration. https://www.eia.gov
• AI inference energy ranges: Independent academic measurements of small, mid-tier, and large LLMs and image-classification models, published in peer-reviewed venues (MLPerf benchmarks; published energy-per-inference studies). Provider-specific per-query figures for the models Percy uses are not publicly disclosed.

Treating this as a snapshot

These estimates reflect our best understanding as of the last updated date at the top of this page. Several of the underlying inputs — model efficiency, grid carbon intensity, our own typical usage — shift over time. We'll revise this when we get cause to (significant model changes, new published per-query figures from providers, material EPA updates, or a reader pointing out an error), but we make no promise about how often that will happen. If the date at the top is more than six months old, treat the numbers as directional rather than current.

Spotted an error or have access to better data than we used? Email us at hello@percy.garden — we read every one.

---

What changed since our last estimate in May

Our May 2026 estimate was 65–75 Wh per user per season. This one is about 100 Wh. Most of the difference is the energy that powers things Percy couldn't do in May:

• The end-of-season report. New this year. Once a year, each gardener gets a report on their season, with a few short passages written by a large, highly capable model. It runs once per person per season, not every day.
• Answers that know your garden. Gardening answers now draw on your own garden's history and local weather, and some go through short safety checks (edibility, invasive-plant advice) before you see them.
• The Garden Letter. New this year: a short letter every two weeks, with a few lines written by a compact model.

Two smaller changes are about counting, not new features:

• Photo screening is now in the total. Every photo you add has always been checked by a compact model before it's saved. The May version listed this but left it out of the total. It's in now.
• Plant ID moved to a newer model in the same class. We've held that estimate where it was.

---

Questions or corrections: hello@percy.garden