Are we doing good?

Google's 2026 Environmental Report (covering 2024 data) disclosed total GHG emissions of ~14.3 Mt CO₂e — up ~13% year-on-year and ~48% above its 2019 baseline, driven by data-centre energy demand and supply-chain growth as AI infrastructure scales. In the same report, Google noted its global 24/7 carbon-free energy (CFE) average reached 90% in 2024, up from 64% in 2020, with some grids already at 100%. Microsoft's 2026 Environmental Sustainability Report tells the same structural story from the other side of the ledger: Scope 1 and 2 held near zero through large-scale power purchasing, while Scope 3 — chips, steel, concrete and supplier energy for datacentre build-out — remains well above its 2020 baseline (the 2024 edition put that gap at roughly 30%; the 2026 edition restates the series, so open the PDF before quoting a figure). Both still hold 2030 net-zero and carbon-negative targets. R4 R2 R10

Annual matching with certificates and hourly, local, additional clean power are not the same claim. Google's 24/7 CFE framework is the hardest public standard in the industry — a 90% global average in 2024 is genuine progress, but the remaining 10% gap, which concentrates in high-demand AI regions, is where the next accountability question lives. Microsoft's answer is shaped differently: long-dated power agreements plus supplier decarbonisation requirements. Fair question to both: which megawatts would not exist without you, and who will publish the hourly shortfall by grid region?

Microsoft runs an AI for Good Lab and its 2026 Environmental Sustainability Report devotes space to AI applied to environmental data, measurement and grid work; Google.org funds AI accelerators and social-impact cohorts; the Bezos Earth Fund launched an AI for Climate and Nature grand challenge with a $100M scale of ambition. Google's 2026 Environmental Report highlights AI applications for flood forecasting, wildfire detection and species identification. In neither company's report is the compute budget underpinning that work separately disclosed. R7 R8 R5 R4 R2

Put the two numbers side by side: grant dollars or reserved compute for AI-for-good versus annual capex on AI infrastructure. Any company can compute that ratio in an afternoon. None publish it. A single honest ratio would tell us more than a hundred case studies — and the 2026 reports from both Google and Microsoft are the closest any of the five has come to making the case for why it should exist.

Apple's 2026 Environmental Progress Report continues the Apple 2030 series: a carbon footprint cut by over 60% against the 2015 baseline on the way to the 75% reduction target for 2030, a Supplier Clean Energy Program bringing gigawatts of renewables onto suppliers' grids, and the Restore Fund (launched at $200M in 2021, expanded in 2023) for residual removals. Like its peers, Apple restates prior-year figures as its accounting improves — the linked PDF governs. Microsoft's 2026 report describes supplier clean-energy requirements and lower-carbon construction materials for datacentre build — the same lever, applied to concrete and steel rather than consumer devices. R1 · 2026 PDF R3 · 2025 PDF R2

Supplier clean power is genuinely leading practice. So: what happens to the materials — recycled aluminium and cobalt are reported, but recycled rare-earth and critical-mineral content, fab water use, and PFAS in chip manufacture stay mostly off the page. Same question for the cloud: how much of a new datacentre's embodied carbon is contractually required to be low-carbon, and how much is hoped for?

Amazon's Climate Pledge Fund: $2B venture capital, alongside The Climate Pledge signatory network, with the group-level picture set out in its 2025 Sustainability Report. Microsoft's Climate Innovation Fund: $1B, reported on again in its 2026 Environmental Sustainability Report alongside carbon-removal purchasing at unusual scale. The Bezos Earth Fund is a $10B grant commitment through 2030. Three very different instruments, one word: "investment". R9 R3 R2 R5

Venture-shaped money finds hardware with an exit. What share of each fund is first-loss, concessional, or grant capital aimed at adaptation, soils, coastal protection, cooling for informal settlements and the small-scale jobs that scale slowly? That's the unbankable middle — and it's where the system actually changes.

Grant announcements are published openly, often to large international NGOs and research institutions that then regrant onward. The first hop is visible; the second and third hops usually are not. R5 R8

Publish one metric: % of dollars where the final decision-maker is based in the country of impact. Also publish median time from application to first payment. Speed and proximity are the two things ground-level organisations actually feel.

Every one of the five runs youth-facing skilling, scholarship or fellowship work at meaningful scale. Almost none disclose a youth body with signing authority over a budget, a seat in an investment committee, or a published say in what gets funded. R1 R4 R2

Training is an input. Authority is an outcome. Ask: how many dollars were allocated last year by a decision body where under-30s held a majority — and what did they choose that the adults wouldn't have?

1. Water and energy per inference — Google's 2026 report flags water as a growing constraint and Microsoft's 2026 report reports water replenishment and datacentre cooling design, but neither goes to per-query disclosure. 2. Repairability and years-in-service per device. 3. Community consent and local benefit records for data-centre siting. 4. Recycled critical minerals. 5. Adaptation as a share of climate spend. 6. Open models in low-resource languages. R11 R12 R13 R4 R2

Frameworks lag reality by about five years. The organisations that will look good in 2035 are the ones publishing an imperfect number now on something no standard requires yet. Volunteering a bad first number is the highest-integrity move available.