Big tech promised net zero. Then AI arrived. Someone has to lose.

Four promises, one problem

Microsoft, Google, Amazon and Meta have each committed to net zero — three of the four by 2030, Amazon by 2040. In their most recently reported year, not one of those footprints fell. Microsoft’s rose 25%, Google’s 18%, Amazon’s 16%, Meta’s 11%. Every one of them now sits well above its own baseline, and the deadlines have not moved.

Episode 3 puts the four trajectories on one chart — the promise flat along the bottom, the reported emissions climbing away from it — and then does the thing the pledges avoid: it asks what it would actually take to close the gap, and whether the market that would have to close it is anywhere near big enough. The short answer, on today’s numbers, is no. Here is the long one.

The climb — AI’s electricity

The driver underneath all four numbers is the same: compute. The IEA’s 2025–2026 energy-and-AI work puts global data-centre electricity demand at roughly 485 TWh in 2025, rising to about 950 TWh by 2030 — on the order of Japan’s entire annual electricity use, and heading toward roughly 3% of world demand. Building and powering that fleet is what shows up as emissions: the concrete and steel of new campuses, the chips inside them, and the electricity — much of it still firmed by gas — that runs them.

So the pressure is structural, not a bad quarter. Every one of these companies is scaling the exact thing that lifts its footprint, at the same time as it promises to drive that footprint to zero.

The four companies

The episode walks each disclosure in the company’s own terms:

• Microsoft — 20.3 Mt in FY2025, up 25% in a single year and about 58% above the 12.9 Mt 2020 baseline of its carbon-negative-by-2030 pledge. Most of the footprint is Scope 3: the buildout itself.
• Google — up 18% in 2025 by its own preferred (market-based) count, and, in the words of its 2026 environmental report, risen every single year since 2020. Its 2030 net-zero goal is one the report itself describes as intentionally aspirational.
• Amazon — around 81 Mt in 2025, up 16%, and roughly 58% above its 2019 starting point. It gave itself the most time — net zero by 2040 — and carries the largest absolute number of the four.
• Meta — the one that looks best on paper: 8.2 Mt market-based in 2024, up 11%. Measured against the physical grid it plugs into, the location-based figure is close to twice that. The gap is contracts, not electrons.

The blank fifth card

Then the episode leaves a card deliberately blank — for the AI labs themselves. As of this recording we could find no public greenhouse-gas inventory, no environmental report, and no emissions target from OpenAI, even as its Stargate program describes roughly half a trillion dollars of compute and more than 9 GW of capacity by 2029. Anthropic states in its model documentation that it targets net-zero climate impact on an annual basis, achieved through offsets rather than a disclosed inventory, and in June 2026 became the first AI company to back the Frontier advance-market commitment — a buyers’ coalition it joined, not one it funds alone. The pure-play labs are scaling fastest and disclosing least.

Two doors out

If your emissions are climbing and your net-zero date is fixed, there are only two doors. Door one: move the target — push the date, or soften the ambition. Door two: keep the date and buy your way to zero with carbon removal. The episode’s middle act shows both doors are already being used.

The walk-backs are on the record: Amazon retired its Shipment Zero milestone in 2023; the Science Based Targets initiative removed both Amazon (2023) and Microsoft (2024) from its list of committed companies during a wider cull; Google ended its long-running carbon-neutral claim after 2023; and in 2026 Google’s 24/7 carbon-free-energy program was reported to be under review. Door one is well travelled. Which leaves door two — removals — and the question of whether that door is wide enough to walk through.

The new rulebook — SBTi V2

What removals have to do got stricter this summer. The Science Based Targets initiative’s Corporate Net-Zero Standard V2.0, finalised on June 11, 2026, is explicit that residual emissions at net zero must be neutralised with carbon removals, and it introduces durability expectations — the storage has to last on a timescale that matches the emission it offsets.

That last point is where the science bites, and it is the hinge of our model. The IPCC’s AR6 is clear that a molecule of fossil CO₂ perturbs the climate for centuries to millennia, while methane’s influence plays out over about a decade. If you emit something that lasts centuries, a store that might not, does not truly neutralise it.

So we modelled it

Here the episode opens our own model — disclosed as ours — and builds it in plain steps. Start with the four companies’ emissions and the IEA’s compute trajectory to size how much residual would remain at their net-zero dates. Then split that residual by what it actually is. Most of these emissions are fossil-derived and long-lived, so under the durability logic above they call for durable removal — direct air capture, mineralisation, durable biochar and the like — not a forest that can burn back. We simplify the matching rule to something anyone can hold: match a long-lived, fossil CO₂ tonne with long-lived, durable storage; match shorter-lived, nature-type emissions with nature-based removal. It is a simplification, and we say so — but it is the right direction, and it is what the new rulebook implies.

The collision — 61 vs 11

Run the residual through that rule and the supply side is where the promise breaks. In our VCM Outlook modelling, durable-removal demand on this trajectory reaches on the order of 61 Mt a year by 2036, while durable-removal supply on a realistic build-out sits near 11 Mt. The gap is not a rounding error; it is the majority of the requirement. Over the longer arc the removal share of the market climbs from roughly 10% today toward 40% and then above 80%, and the value of the durable segment grows from around a billion dollars to well over a hundred billion — but the tonnes have to physically exist first, and on today’s pipeline they do not.

The removal market today

The live market underlines it. Durable carbon-removal purchases had a record year — market trackers put roughly 30.4 Mt contracted in 2025, several times the year before — but contracted is not delivered: only around 3% of all durable tonnes ever contracted have actually been delivered, and a single buyer, Microsoft, accounts for the large majority of all-time volume. A market that leans on one buyer, and that has delivered a few percent of what it has sold, is not yet the machine that saves four net-zero pledges. And in 2026 even that buyer’s future purchasing was reported to have paused before trickling back.

The verdict — someone has to lose

Put it together and the episode’s title stops being a slogan. Emissions are rising because the thing driving them is being built as fast as capital allows. The deadlines are fixed. Door one — moving the target — is already being used. Door two — removals at the scale and durability the science now demands — is real but roughly an order of magnitude short of what the trajectory needs, and delivering only a few percent of what it has promised. On today’s numbers something has to give: the ambition, the date, or the definition of net zero itself. That is the collision, and it is arithmetic before it is opinion.

Sources

• Microsoft 2026 Environmental Sustainability Report & Environmental Data Fact Sheet — 20.3 Mt FY2025, +25%; 12.9 Mt FY2020 baseline.
• Google Environmental Report 2026 (pp. 4, 69–70) — +18% in 2025, footprint risen every year since 2020, 2030 goal described as “intentionally aspirational.”
• Amazon 2025 Sustainability Report — ~81 Mt in 2025, +16%; net zero by 2040; 51.2 Mt 2019 baseline.
• Meta 2025 Sustainability Report — 8.2 Mt market-based 2024, +11%; location-based roughly 2×.
• IEA, Energy and AI (2025) and Key Questions on Energy and AI (2026) — data-centre demand ~485 TWh (2025) to ~950 TWh by 2030.
• SBTi Corporate Net-Zero Standard V2.0 (final, June 11, 2026) — residuals neutralised with removals; storage-durability requirements (§§6.5–6.6).
• IPCC AR6 WG1, Ch. 5 & 7 — atmospheric lifetime of CO₂ (centuries–millennia) vs methane (~a decade).
• CDR.fyi market updates (2026) — ~30.4 Mt durable removal contracted in 2025; ~3% of all contracted tonnes delivered; concentration in a single buyer.
• Heatmap News & Bloomberg (April–May 2026) — reported pause on future Microsoft removal purchasing; 24/7 carbon-free-energy program under review.
• OpenAI Stargate program disclosures — ~$500B compute, 9+ GW by 2029; no public GHG inventory or target located as of recording.
• Anthropic model documentation — annual net-zero-impact claim via offsets; Frontier advance-market-commitment coalition (joined June 17, 2026).
• Demand-and-supply projections (durable ~61 vs ~11 Mt/yr by 2036; removal share rising past 80%): Climate Decode VCM Outlook 2026 — our own model; methodology, ranges and assumptions disclosed there.