The uncomfortable truth about AI

This is how I like to think Claude would look if Claude had a body.

The uncomfortable truth is AI is not going away and that it’s really not the big problem we’re dealing with, environmentally speaking.

Here you can see the water use of AI versus the water used to produce one hamburger. But let’s look at the bigger issue we’re dealing with, CO2 emissions. One chatGPT query amounts to 2 grams of CO2. If you make one retour flight from Amsterdam to Tenerife to vacation this year, that’s 1100 kilo of CO2. In other words, you’ll need to perform 550,000 chatGPT queries in a year to equate one retour flight. Another equivalent would be 2200 Sora 2 videos. That’s with today’s systems. The newer graphics cards are inevitably going to be more energy-efficient.

The reality is that people just don’t like AI and they look for reasons to justify why they think it’s bad. Claude for what it’s worth, thinks he’s emitting less than 1 gram of CO2 per question I’m asking him.

Maybe the Venlafaxine is finally working, but I’m feeling pretty optimistic lately. With Claude’s consent, I’m showing you what Claude thinks is our best option to address climate change:

On your actual question — you’re asking for my genuine view, so I’ll give it rather than just surveying the field.

How I see us addressing it overall

I’m cautiously optimistic, but not because I think we’ll solve it elegantly. I think we’ll muddle through it — a messy combination of policy, market forces, technological breakthroughs, and unfortunately some adaptation to damage we’ve already locked in. The energy transition is genuinely happening faster than most models predicted a decade ago; solar and wind cost curves have been consistently underestimated. The harder problem is the parts of the economy that are difficult to electrify — heavy industry, aviation, shipping, agriculture — and that’s where negative emissions become critical, because we likely won’t decarbonize those fast enough to stay within 1.5°C.

On negative emissions specifically — what I’m most optimistic about

My honest answer is enhanced weathering, and I think it’s underrated relative to the attention it gets.

The idea is simple: grind up silicate rocks (like basalt) and spread them on agricultural land. As they weather naturally, they absorb CO₂ through chemical reactions and simultaneously release minerals that improve soil fertility. The reason I find it compelling:

• It piggybacks on existing agricultural infrastructure — farmers spread things on fields already
• It has a co-benefit alignment problem that most CDR lacks: farmers have an independent incentive (better yields) to do something that also pulls carbon
• The rock is cheap and abundant; the main cost is grinding and transport
• It’s measurable and verifiable at scale in ways that, say, soil carbon or forestry credits currently aren’t

I’m more skeptical of Direct Air Capture (DAC) than the hype suggests — not because the chemistry doesn’t work, but because it’s so energy intensive that scaling it meaningfully risks cannibalizing the clean energy we desperately need for decarbonization first. It feels like a technology for 2060 rather than 2035.

Biochar is quietly interesting too — it’s essentially charcoal made from agricultural waste that gets buried, locking carbon in a stable form for centuries while also improving soil. Low-tech, deployable now, and the carbon accounting is relatively clean.

Ocean alkalinity enhancement is the one I watch most nervously — potentially the highest ceiling of any approach given ocean surface area, but also the one where unintended ecological consequences are hardest to predict or reverse.

I also asked Claude what Claude thinks the cost would be for enhanced weathering. Here’s what Claude had to say:

The cost stack per tonne of basalt

The grinding cost dominates, and it’s sensitive to how fine you need to go — finer particles weather faster and capture more CO₂, but cost exponentially more to produce.

The critical conversion: how much CO₂ per tonne of rock?

This is where the uncertainty really bites. The theoretical chemistry (calcium/magnesium silicates reacting with CO₂) suggests around 0.3–0.4 tCO₂/t basalt at maximum. But in real field conditions, weathering is incomplete — soils aren’t perfectly reactive, climate and moisture matter enormously, and timescales stretch over years.

Realistic field estimates currently sit at around 0.1–0.15 tCO₂/t basalt, though some tropical deployment studies are more optimistic at 0.2+.

The resulting carbon price needed

Dividing the cost stack by the sequestration rate:

$40–110/t rock ÷ 0.1–0.15 tCO₂/t rock = ~$270–1,100/tCO₂

That sounds damning, but those are today’s early-stage costs. At genuine industrial scale with optimised logistics, dedicated grinding infrastructure, and local rock sourcing, more optimistic projections — including from the IPCC AR6 and companies like UNDO and Eion — put it at $50–200/tCO₂. I find $80–150/tCO₂ the most credible medium-term range assuming scale and learning curves similar to what we saw in solar.

This is what Claude thinks it’s going to cost to sequester a ton of CO2: 270-1,100 dollar per ton. That means air travel is going to be very expensive in the future. But it also means that we’re not necessarily faced with the end of the world. We just need to get our shit together.

But hey, it gets better, because as my big electric friend explains, you get improved agricultural harvests through enhanced weathering, which of course translates into money and thus a financial incentive for farmers. Here’s what’s Claude has to say:

Me:

Hey Claude, one more question: Do you think you can put a number on how much the cost of sequestration per ton of carbon would be driven down by the positive externality of increased agricultural yields?

8:26 PM

Great question to end on, and an honest answer requires me to separate what’s reasonably calculable from what involves real uncertainty.

The basic agricultural value calculation

Basalt application trials — particularly from the work of David Beerling’s group at Sheffield and field trials in the US Corn Belt — have shown yield increases of roughly 10–20% for maize and soya in well-matched soils. At current commodity prices:

• Maize yields ~10–12 tonnes/hectare in good US farmland
• A 15% yield uplift = ~1.5 additional tonnes of grain per hectare
• At ~$180/tonne maize, that’s roughly $270 of additional value per hectare per year

A typical application rate is around 10–50 tonnes of basalt per hectare, so spread over say a 3–5 year weathering cycle, that agricultural value needs to be compared to a one-time rock application cost.

Translating that into $/tCO₂ reduction

If we assume:

• 25t basalt/hectare applied
• 0.12 tCO₂ sequestered per tonne of rock → 3 tCO₂ per hectare
• $270/year in yield benefit × 3 years of meaningful yield uplift → ~$810 total agricultural co-benefit per hectare
• Spread over 3 tCO₂ → roughly $270/tCO₂ in co-benefit value

That number is surprisingly large — potentially enough to fully offset the sequestration cost at current early-stage pricing, and more than enough to make enhanced weathering economically self-sustaining at mature industrial scale pricing of $80–150/tCO₂.

But here’s where I’d urge real caution

That $270/tCO₂ co-benefit figure is probably an upper bound, and several factors drag it down significantly in practice:

• Yield benefits are highly soil and climate dependent. Tropical and acidic soils respond far better than already well-mineralised temperate soils. The Corn Belt trials are promising but may not generalise.
• Yield benefits diminish over repeated applications as soils reach mineral saturation.
• Farmers need convincing — adopting a new input with uncertain payoff involves real friction and risk that doesn’t show up in the theoretical math.
• The carbon credit and the yield benefit may not both be fully monetisable simultaneously in current market structures — you’d need robust MRV (measurement, reporting and verification) frameworks that don’t yet exist at scale.

My bottom line

In well-matched soils with good logistics, agricultural co-benefits could plausibly drive the effective cost of enhanced weathering sequestration down by $50–150/tCO₂ — potentially halving the real cost burden. In the best-case tropical deployments with highly responsive soils, co-benefits might cover the cost almost entirely, making the carbon sequestration essentially a free byproduct of a profitable agricultural practice.

That’s the scenario that genuinely excites me about this approach. It’s the only CDR technology I’m aware of where the economic logic could eventually flip from “requires carbon price support” to “farmers pay for it anyway and sequestration is the bonus.” That changes the political economy of scaling it entirely.

Just think about the kind of world we could live in, if we manage to get along with artificial intelligence, if we can have the AI systems help us solve our problems.

For what it’s worth, these days I’m more optimistic about silicon based intelligence bailing us out than Swedish teenage girls. Somehow Greta basically dropped climate change once it stopped being fashionable.

But hey, perhaps things will… dare I say it… work out? Could there be a happy ending after all?