共有:
Supply-Chain Emissions

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.

AI Navigate Editorial·2026.07.06·6 min read
UPSTREAM DATA CENTER GPU fab Steel & cement Freight & build Scope 3 Ops & cooling (Scope 1 & 2) GOOGLE AMAZON +25% +16%
01
The Overlooked Focus

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 coveredWhat the numbers now expose
Operational electricity useGPU fabrication & semiconductor supply
Water for coolingSteel and concrete of new sites
PUE and renewable shareVisibility into upstream Scope 3
Local grid strainAlignment with 2030 net-zero pledges

Not the power that runs the server,
but the CO2 of making the server.


02
The Numbers

The surge, in
three figures

Not an efficiency slip — this is buying more chips and building more sites, showing up as upstream emissions.

+25%
Google supply-chain CO2 (YoY)
+16%
Amazon supply-chain CO2 (YoY)
2030
Target year for net-zero pledges

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.

03
Where It Comes From

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 BREAKDOWN — HYPERSCALERS SCOPE 1 Direct emissions (on-site combustion) SCOPE 2 Purchased electricity (indirect) SCOPE 3 ~70-80% of total This is the slice that grew +25% / +16%
FIG. Scope 3 dominates the total, and it is the slice that is now swelling with AI buildout.

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.

04
What Can Be Done

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.


05
Frontier

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.

AI Navigate — Daily Update · 2026.07.06 / Source: Power & data center limits — the physical constraints on AI scaling