Carbon Intensity of GPU Compute

Grid carbon intensity varies from 20g CO2/kWh (Iceland, 99% hydro/geothermal) to 400+ g CO2/kWh (Poland, Germany in peak coal generation). This is a factor-of-20 decision. A single H100 at 700W continuous adds 5,040 kWh/year.

In Iceland: 100kg CO2/year. In coal-dependent Poland: 2,016kg CO2/year. A 1,000-GPU cluster in Poland generates 2 million tonnes CO2/year; identical cluster in Iceland generates 100 tonnes.

Operator location is a carbon efficiency mechanism. Crusoe Energy explicitly optimises for stranded power (flaring capture, data centre cooling with waste heat recovery) and low-carbon grids. Lambda has deployed in Iowa and Texas (wind-heavy grids). CoreWeave has European capacity in cold, hydro-rich regions.

Many operators claim 'renewable energy', but the mechanics matter enormously. Power Purchase Agreements (PPAs) represent physical renewable energy generation tied to your consumption. RECs (Renewable Energy Certificates) are accounting abstractions.

A cluster on PPA with an operating solar farm is genuinely low-carbon. A cluster buying RECs whilst connected to a coal grid is not. RECs do nothing to increase renewable generation; they're a claim on existing supply.

EU regulations increasingly distinguish between these (Scope 2 Emissions accounting), creating competitive advantage for operators with actual PPAs. Crusoe's waste-heat recovery and flaring-gas energy is genuine carbon reduction. Lambda's Iowa capacity benefits from 55%+ wind generation. Most other operators claiming 'renewable energy' use RECs and are greenwashing.

Liquid cooling uses 10-15% less energy than air cooling in most climates, translating to 8-12% reduction in operational emissions. However, liquid cooling increases manufacturing emissions (more complex hardware, cooling fluid production) and disposal complexity.

Lifecycle carbon comparisons favour liquid cooling in warm climates (southern US, Middle East). They're neutral or favour air cooling in cold climates where air cooling is negligible cost.

Full lifecycle carbon (manufacturing plus operational plus disposal) for a B200 is roughly 5-8 tonnes CO2 equivalent. Operational emissions over a 3-year lifespan in a typical US grid (300g CO2/kWh) represent 8-12 tonnes CO2. Manufacturing is 40-50% of total carbon footprint; operational is 50-60%. Grid carbon intensity matters more than cooling technology choice, but cooling still influences 10-15% of operational emissions.

Carbon-aware scheduling (routing compute to low-carbon hours or regions) is emerging as a differentiation point. Google and some hyperscalers schedule batch workloads (training, model serving) when grid carbon intensity is lowest.

For a GPU cluster: queue training jobs during low-carbon hours (night, peak wind/solar periods), serve inference during high-demand times. This reduces carbon per unit of compute by 15-25% with zero hardware changes.

Operators offering 'green tier' pricing (discount for carbon-flexible workloads) can reduce per-unit carbon whilst maintaining margins. EU regulations (Fit for 55, corporate carbon accounting) are creating demand for this. First-mover operators (Crusoe, some European operators) are positioning this as a moat.

EU carbon regulations (Scope 1 and 2 emissions accounting, carbon border adjustment mechanism) are creating structural advantage for Nordic operators. A B200 cluster in Norway reports 50kg CO2/month (grid intensity advantage) versus same cluster in Germany reporting 1,500kg CO2/month. EU enterprises increasingly need to report carbon footprint, creating demand for low-carbon compute.

This tilts the competitive field. Nordic operators (Crusoe in Norway/Iceland, some European entrants) can legitimately claim 20-40x lower carbon intensity than US-based competitors. This is not a niche concern (ESG posturing) but a competitive fact.

Large enterprises (automotive, finance, energy) are mandated to report and reduce carbon. Low-carbon compute becomes a feature, not a nice-to-have. Hyperscalers are responding: AWS is pushing renewable regions; Azure is deploying in Scandinavia. Neocloud operators who don't prioritise low-carbon locations are at long-term disadvantage.

Grid carbon intensity (20-400g CO2/kWh) drives 20-40x variance in operational carbon; location choice outweighs any hardware efficiency

PPAs (physical renewable energy) differ fundamentally from RECs (accounting abstractions); operators with genuine PPAs have genuine carbon advantage

Liquid cooling saves 10-15% operational energy but is 50-60% of lifecycle carbon; grid intensity matters more than cooling technology

Carbon-aware scheduling (batch workloads during low-carbon hours) reduces per-unit carbon 15-25% with zero hardware investment

EU regulations are creating structural advantage for Nordic operators; carbon intensity becomes a competitive feature for enterprise customers