what AI weighs
entry 02 · what a tonne costs — six methods, with sources
written and designed with AI, editorially reviewed

One tonne. Six dollars. Five hundred dollars.

Market averages for one tonne of CO₂e, 2025–26: $6 for avoided deforestation, over $500 for direct air capture.1 Both are sold as one tonne. The spread is not a discount. It is a different object. Choosing between them is the first decision anyone has to make.

where the measurement standsor scroll ↓
01 — the cheap end

Six dollars buys a promise about a forest that still stands.

Avoided-deforestation (REDD+) credits pay a landowner not to cut. The tonne is counted against a baseline of what would supposedly have happened otherwise — a counterfactual nobody can observe. A synthesis of six independent ex-post evaluations covering 44 projects found they claimed 10.7 times more avoided deforestation than the evidence supports; the projects mostly did reduce deforestation, just far less than they sold. An earlier 26-project study, West et al. (2023), was disputed; this synthesis came after it.

Method
Avoided deforestation (REDD+)
Average price benchmark per tonne
$6
Storage duration
Storage holds as long as the protection does — years to decades, and only while someone is watching.
durability
low
affordability
very high
verifiability
contested
02 — the popular one

Twenty-two dollars buys a tree, and a tree is not a vault.

Reforestation is real removal and it dominates issuance: around 7.5 million ARR credits in 2025, more than all long-duration removal credits ever issued by the same registries. But biological carbon can come back out — one fire season, one drought, one change of owner. It is storage with a cancellation clause, and registry supply looks larger than it is: of 773 listed ARR projects, only 158 have issued anything.

Method
Afforestation & reforestation (ARR)
Average price benchmark per tonne
$22
Storage duration
Decades, conditional on the land staying forest and nobody needing it for something else.
durability
moderate
affordability
high
verifiability
reversible
03 — the one with no price

Soil holds enormous carbon. Proving you added some is brutal.

Soil-carbon credits come with genuine agronomic co-benefits — water retention, resilience, yield. The problem is measurement: demonstrating a change in soil stock takes years of sampling across heterogeneous ground, and the signal is small against the noise. We found no public price index for it, which is why this panel shows none. Thin, bilateral trading is a finding about the method, not a gap in our research.

Method
Soil carbon
Average price benchmark per tonne
no index
Storage duration
Decades, and reversible with a single change of tillage practice.
durability
moderate
affordability
high
verifiability
weak
04 — where the volume is

A hundred and seventy-seven dollars buys carbon in a form that does not rot.

Biochar fixes carbon into a stable solid, quantified at the kiln rather than estimated across a landscape. It is the dominant durable pathway by buyers and deliveries — nearly five times more unique buyers than the next method, though not by contracted tonnes — and it survives being checked. Note the disagreement underneath the average: the Puro CORC biochar index (in euros) stood at €124 in September 2026, while suppliers surveyed in early 2025 said they need $187 to make a reasonable profit.

Method
Biochar
Average price benchmark per tonne
$177
Storage duration
Centuries in soil, with the lasting share estimated from lab-measured properties of the char.
durability
high
affordability
moderate
verifiability
strong
05 — the geological end

Two hundred dollars buys a reaction that runs for ten thousand years.

Enhanced rock weathering converts CO₂ into dissolved bicarbonate — effectively permanent. The durability is not in question; how much actually weathered, how fast, and how much you may count is exactly what the field is still standardising. Suppliers surveyed in 2025 said they need $349/t to be profitable while buyers called $271/t expensive — the gap is the state of the method, priced.

Method
Enhanced rock weathering
Average price benchmark per tonne
$200+
Storage duration
Millennia. The chemistry is not the uncertain part — the accounting is.
durability
very high
affordability
low
verifiability
immature
06 — the expensive end

Five hundred dollars buys the only tonne nobody argues about.

Direct air capture is unambiguous: molecules pulled from the atmosphere and put underground, metered end to end. It is also scarce and priced accordingly. Buyers surveyed put the “good value” point at $458/t in 2025 and expect roughly a 41 % fall by 2030. Buy it today and you are partly paying for the scale-up of an industry that is not built yet.

Method
Direct air capture
Average price benchmark per tonne
$500+
Storage duration
Millennia, in geological storage, with metering at every step.
durability
very high
affordability
very low
verifiability
strong
status quo · where bdrop.studio actually stands

We have not finished measuring.

The measurement is not finished and will not be for a while. This section is where it stands.

A shipped Drop is estimated at ~38 g CO₂e, plausible band 15–120 g, location-based, ISO/IEC 21031 boundary. That is a first-cut figure from assumed activity, not from telemetry, and its largest input is still an assumption rather than a measurement. The estimate sums the text-model calls, four image and three music generations (two of each discarded along the way), about 5 Wh of infrastructure and 15 % for embodied hardware: an internal first cut, not yet run through Monte Carlo.

15–120 g
estimated per Drop, from our own
assumptions — a band, not a measurement
what we still don't know

Our music and image providers do not publish which country their servers run in, and we have not found it in their documentation. Until that is resolved, grid intensity stays at a 100–700 gCO₂e/kWh band — the widest single term in the estimate. We have not written to them yet. Whatever comes back when we do, including silence, gets published here.

Narrowing the band is the work: ×/÷8 down to ×/÷3, and the fastest route there runs through two disclosures we do not control.

And we still don’t know what we emit.

Knowing what a tonne costs is the easy half. The hard half is knowing how many tonnes there are. Drag the assumptions and watch the estimate move.

50 %
$ 3.38
400 g/kWh
widest term: grid intensity — region undisclosed
band ratio ×/÷ 1.94 (IQR of estimated per-query energy)
16g CO₂e · one music generation, not a whole Drop
Plausible range 8.2–31 g. The band is the middle half of estimated per-query energy on frontier text models (Oviedo et al., 0.16–0.60 Wh around a 0.31 Wh median 4), applied here for scale; it does not include the 100–700 g/kWh grid range, which is wider. Quoting one decimal-precise number here would tell you less than the band does.
sources & method
[1] prices

Sylvera, carbon offset pricing (2026 update): average price benchmarks by method — REDD+ $6, ARR $22, biochar $177, ERW over $200, DAC over $500. Averages across a wide distribution; regional spreads are large (African afforestation $37 vs Asian $14).

sylvera.com
[2] supplier / buyer gap

CDR.fyi × OPIS, Bridging the Gap: Durable CDR Market Pricing Survey (Jan 2025 edition), 97 suppliers and 25 purchasers: biochar $187 profitable vs $155 “expensive”; ERW $349 vs $271; DACCS “good value” $458, expected to fall ~41 % by 2030. The Puro.earth CORC biochar index (in euros) stood at €124 in September 2026.

cdr.fyi
[3] over-crediting

Swinfield et al., Nature Communications 17, 3944 (2026), synthesis of six ex-post evaluations across 44 REDD+ projects: most reduced deforestation, but claimed 10.7× more than independent estimates justify. The earlier 26-project study is West et al., Science (2023), linked here. Contested by Mitchard et al. (2023, preprint), who argue West et al. used unsuitable deforestation data and poorly matched controls, and made calculation errors. We cite both.

science.org · ade3535
[4] the uncertainty band

Oviedo et al. (2026), Joule: frontier models (>200 B parameters) on H100 nodes, median 0.31 Wh/query, IQR 0.16–0.60 — the ×/÷ 1.94 band used here. ARR issuance figures: Berkeley Carbon Trading Project Voluntary Registry Offsets Database (v2026-06).

cell.com/joule
what is editorial here

The durability / affordability / verifiability bars are our own qualitative reading of the literature, not a score any registry issues. The ordering is defensible; the exact lengths are a judgement. Use them to compare methods against each other; they are not good for anything else. The ~38 g per-Drop estimate and its 15–120 g band are our own first cut, built on assumptions; the 25 Wh allowance, the 75 % utilisation and the $3.38 GPU-hour rate are our assumptions, not sourced figures.

what is a demonstration

The explorer is a derivation, not a measurement: spend → GPU-seconds → watt-hours → grams, at 700 W (H100 SXM TDP), 75 % utilisation (our assumption) and a ×2.4 full-stack overhead (Google’s ratio of comprehensive to accelerator-only energy: idle capacity, host CPU and RAM, and PUE), plus a flat 25 Wh allowance (our assumption) for everything that is not the accelerator — orchestration, storage, transport — which at the default settings is the majority of the total and does not move when you drag the dials. Its headline input, the per-generation cost, is an assumed $0.08 rather than a figure off a bill, and the accelerator part of the estimate scales directly with it. It is there to show how far an answer travels when assumptions change. A later entry replaces it with bdrop.studio telemetry. Not started yet.

Grid intensity, data-centre energy and per-prompt figures are sourced in the research log. Prices are market averages from 2025–26 and move; each is dated at its source. Corrections welcome and published with attribution.