Data center CO2 is quietly
surging upstream.
Google's supply-chain emissions rose 25% year-over-year and Amazon's 16%, new sustainability reports show — a hidden cost of AI scale that power, water and cooling debates never touched. With 2030 net-zero pledges on the table, the balance just got harder.
Behind the power and water
debates, an unseen source
Data center impact has long been discussed as an operations story — GPUs drawing electricity, cooling systems drinking down local water, grids groaning under the load. All of that framing focused on what happens while the servers are running. The new sustainability reports from hyperscalers shift the frame entirely: the fastest-growing emissions are landing before a single server boots, in the making of the chips and the building of the sites.
Google's supply-chain emissions climbed 25% year over year; Amazon's climbed 16%. This is not an efficiency regression. It is Scope 3 procurement emissions — GPU fabrication, steel and concrete, freight — inflating in step with the AI buildout.
| What the debate covered | What the numbers now expose |
|---|---|
| Operational electricity use | GPU fabrication & semiconductor supply |
| Water for cooling | Steel and concrete of new sites |
| PUE and renewable share | Visibility into upstream Scope 3 |
| Local grid strain | Alignment with 2030 net-zero pledges |
Not the power that runs the server,
but the CO2 of making the server.
The surge, in
three figures
Not an efficiency slip — this is buying more chips and building more sites, showing up as upstream emissions.
AI demand is pulling forward site construction and mass GPU procurement, and every ton of steel, every wafer, every truckload lands as procurement emissions on the books. Efficiency on the operations side — PUE gains, renewable PPAs — cannot outrun the upstream when it grows this fast. For hyperscalers holding 2030 net-zero pledges, balancing AI growth against climate targets is now the harder question.
Roughly 7-8 of every 10 tons
lives in the upstream
Emissions are reported in three scopes; for hyperscalers, Scope 3 is by far the dominant slice.
Scope 1 is direct emissions from owned operations (generators, on-site combustion). Scope 2 is the indirect emissions from purchased electricity. Scope 3 covers upstream inputs — raw materials, components, freight, capital goods. For hyperscalers Scope 3 sits at roughly 70-80% of the total, and no amount of renewable PPA or PUE tuning shifts the upstream directly.
The +25% and +16% are precisely the sound of the upstream not moving. Trying to hit a climate pledge from the operations side alone is now visibly, structurally short.
Three levers
operators still hold
Moving the upstream needs contracts, evaluation and disclosure to shift in parallel — no single lever gets there alone.
Additional-renewable PPAs
Sign long-term PPAs tied to newly built low-carbon generation, not just paper transfers on existing capacity. Move Scope 2 and push new supply onto the grid.
Supplier-level emissions scoring
Require measured Scope 1 & 2 data from GPU and component vendors and fold it into procurement decisions, alongside price and performance benchmarks.
Per-site build & cooling disclosure
Publish embodied-CO2 and PUE per data center, not just group-level averages. Give reviewers a way in to the black box of the buildout.
Can 2030 net-zero
still hold?
The 2030 net-zero pledges from Google and Amazon carry more weight once the assumptions behind AI scale visibly shift. If the operations side keeps improving while upstream grows at 25% a year, the pledge and the numbers run in opposite directions. Debating power, water and cooling alone will not close that gap.
Conversely, if upstream visibility hardens into an industry norm — down to sub-suppliers and capital goods — then AI expansion and climate targets can begin to move together again. The next front is upstream transparency: how deep into the supply chain the hyperscalers are willing to reach.