Claude for Climate | Episode 4 of 7
A company can be telling the truth and still be picking which truth to tell. Amazon’s 2025 sustainability report is Episode 1’s metric-selection problem, live, with real stakes.
Rob Aldrich · Technology & Sustainability Executive · Sydney, Australia
robaldrich.com · LinkedIn · @googlenut
Amazon published its 2025 sustainability report on 1 July. The headline number is a 16% jump in absolute emissions, the company’s largest single-year increase since it started reporting, taking it to roughly 80.85 million metric tons of CO2 equivalent. Since the 2019 Climate Pledge baseline, cumulative emissions are now up 58%. Scope 2, emissions from purchased electricity, rose 34% in a single year, and Scope 3 rose 20%. The driver in Amazon’s own words: data center buildout for AI, delivery network electrification, and building electrification.
Amazon doesn’t dispute the 16% figure; it sits in the company’s own report. Alongside it, the report and Amazon’s public commentary also feature carbon intensity, grams of CO2 per dollar of revenue, which is down 38% against the 2019 baseline even though it ticked up 3% year over year. Emissions per unit shipped fell 7% from 2024 and 39% since 2019. Chief Sustainability Officer Kara Hurst’s introduction to the report calls out “tremendous change” each year since 2019. Read together, the figures support a specific argument: the company grew, energy use grew with it, but efficiency per dollar and per package kept improving, so the underlying trend is still a positive one.
That argument isn’t dishonest, and it isn’t the whole picture either. Both figures are accurate. Which one a reader takes away as the headline number determines whether Amazon’s 2025 reads as progress or as backsliding.
This is Episode 1’s problem, live
I opened this series by writing about metric selection, how the same underlying reality produces a good story or a bad one depending on which denominator you choose. Amazon’s report is that argument playing out at hyperscaler scale, in public, with real stakes.
Both numbers are true simultaneously. Carbon intensity per revenue dollar has fallen 38% since 2019. Absolute emissions have risen 58% over the same period. A company can genuinely be getting more efficient per unit of output while its total footprint climbs, if output is growing fast enough. That’s exactly Amazon’s position, and it’s worth saying plainly: intensity metrics are not fake, and they capture something real about operational improvement.
But intensity metrics also have a structural property that absolute metrics don’t: they can improve every year forever while total emissions keep rising, as long as revenue grows faster than emissions. Environmentalists critiquing the report make this point directly, arguing that for a warming atmosphere, total emissions are what matters, not emissions per dollar of a company’s revenue. The atmosphere doesn’t discount for growth.
Google’s 2025 report shows the same pattern from a different company: emissions up 18% year over year on an “ambition-based” basis, electricity consumption up 37%, with the company stating outright that its AI infrastructure buildout is accelerating faster than the grid is decarbonizing. Neither company is unusual here. This is what the current phase of the AI buildout looks like across the hyperscaler tier.
What I saw from the inside
I led AWS’s global sustainable buildings practice, so I’ve sat on the side of this argument that’s constructing the numerator. Data center sustainability work at that scale is genuinely difficult, and Amazon’s operational wins are real: it has matched 100% of its global electricity consumption with renewable energy for a third consecutive year, and its carbon-free energy portfolio spans more than 700 projects and 42 gigawatts of capacity.
“Matching” renewable purchases against total consumption is an accounting relationship, not a physical one. A data center in a coal-heavy grid draws coal-heavy power in real time, regardless of a renewable energy certificate purchased elsewhere on the grid.
The industry has debated this for years, and Amazon’s 34% jump in Scope 2 emissions this year, even after all that renewable procurement, is a reminder that matching and physical grid impact are different questions. Building the sustainable buildings function taught me that the honest version of this story has two halves: real progress on procurement, and a physical footprint that grows anyway when the underlying business, in this case AI infrastructure, is expanding faster than the clean grid can keep up.
Reading a sustainability report like this one
If you’re evaluating any company’s climate disclosure, Amazon’s report is a good template for the checklist. Claude is useful here, but the judgment about which number matters is still yours to make.
1. Find the absolute number before the intensity number
Every sustainability report leads with its best framing. Find the total emissions figure and the year-over-year change before you read the narrative around it.
Prompt pattern: “Here’s a company’s sustainability report [paste text]. Extract the absolute Scope 1, 2, and 3 emissions figures for the most recent two years, and separately extract any intensity metrics (per revenue, per unit, per employee). List them in two clear tables before summarising the company’s own framing.”
2. Check what the baseline year is doing for the story
A 58% cumulative rise since 2019 sounds different from a 16% single-year rise, and both are the same set of facts, viewed on different timelines. Claude can hold multiple timeframes side by side without picking one to feature.
Prompt pattern: “Calculate year-over-year change, three-year change, and change since [baseline year] for these emissions figures. Present all three without recommending which one is most representative.”
3. Separate procurement claims from physical grid impact
“Matched with renewable energy” and “powered by renewable energy” describe different things. Ask Claude to flag the distinction every time a report uses matching, offsetting, or certificate language.
Prompt pattern: “Read this report’s renewable energy claims. Flag every instance where the language describes matching, certificates, or offsets rather than direct physical supply, and explain the difference in plain terms.”
4. Compare the company’s stated driver against its stated response
Amazon names AI data center buildout as the primary emissions driver. Its response, more renewable procurement and efficiency gains, addresses the electricity mix, not the growth in physical infrastructure. That gap is worth naming, not glossing over.
Prompt pattern: “This report names [driver] as the primary cause of increased emissions. Summarise the mitigation measures the company describes, and note explicitly whether those measures address the driver directly or address a related but different factor.”
The limitations, stated plainly
Claude can’t tell you which metric is the right one. Intensity versus absolute, this year versus since-baseline, these are judgment calls about what a report is meant to communicate, and reasonable people land in different places. Claude can surface every framing. It can’t decide which framing is honest for your purposes.
Claude can’t verify a company’s underlying data. Everything in this episode comes from Amazon’s own disclosure and from journalists who’ve reported on it. Claude can help you read a report closely. It cannot audit the meter readings behind it, and neither can I from outside the company.
This isn’t a story with a villain. Amazon’s renewable procurement is real, its intensity improvements are real, and its emissions growth is also real. A company can be doing meaningful work and still be net worse for the atmosphere this year than last. Holding both of those facts at once is uncomfortable, and it’s also just accurate.
I have a specific bias here. I spent years building the AWS practice this report is now describing. I’ve tried to write this episode the way I’d want a former colleague to write about work I used to lead: honestly, without either defending the company reflexively or performing outrage for an audience. Judge whether I’ve managed that.
The bottom line
Nothing in Amazon’s report is false. Both the 58% cumulative rise and the 38% intensity improvement are real, and a reader who only sees one of them walks away with a different story than a reader who sees both. That’s true of every hyperscaler’s climate disclosure right now, not just Amazon’s: the AI buildout is the reason both numbers are moving at once, efficiency improving, absolute emissions rising, and that tension isn’t resolving for Amazon, Google, or anyone else scaling AI infrastructure over the next five years. Reading these reports with both numbers in view matters more now than it did in 2019, and it’s about to matter for a lot more companies than the hyperscalers.
Rob Aldrich has spent 25 years at the intersection of enterprise technology and sustainability. He is the co-creator of Cisco EnergyWise, former Global Sustainable Buildings Lead at Amazon Web Services, and former interim Chief AI Officer at SafetyCulture. He is the author of IP-Enabled Energy Management (Wiley/IEEE Press) and has conducted sustainability audits across more than 70 data centres worldwide.
robaldrich.com | linkedin.com/in/rob-aldrich | @googlenut
Sources
- KUOW, “Amazon’s carbon emissions jumped 16% in 2025. The driver: massive data center buildout” – kuow.org
- ESG Dive, “Amazon’s 2025 emissions jump as AI brings ‘momentum and complexity’: report” – esgdive.com
- Bloomberg, “Amazon’s Carbon Emissions Rose 16% in 2025 Amid Data Center Boom” – bloomberg.com
- Trellis, “Amazon stays ‘stubborn’ on net-zero pledge” – trellis.net
- GeekWire, “The cost of the AI boom: Amazon emissions jump 16% as company stands by net-zero pledge” – geekwire.com
- eWeek, “Big Tech Carbon Emissions Rise as AI Data Center Expansion Accelerates” – eweek.com
- Rinnovabili, “AI data center emissions hit record highs” – rinnovabili.net