Carbon, Clean Power Procurement & 24/7 Carbon-Free Energy

Carbon is where a data center's sustainability story is either real or rhetorical, and the line between the two is an accounting choice almost no executive deck makes explicit. Two facilities on the same grid, drawing the same megawatt-hours, can each claim "100% renewable" — and one of them is genuinely running on clean power most hours of the year while the other is running on the marginal gas unit at 2 a.m. and papering over it with a renewable energy certificate (REC) bought from a wind farm a thousand miles away that was generating at noon six months ago. Both claims are defensible under today's GHG Protocol. Only one is decarbonization. The fork in this chapter is the matching granularity you commit to — annual volumetric matching versus 24/7 hourly carbon-free energy — and every downstream cost (procurement strategy, on-site generation, scheduling, disclosure exposure) cascades from it.

This matters more in 2026 than at any prior point because two clocks are running against each other. The load clock: AI is adding gigawatts of new, high-load-factor demand faster than clean supply can be built, so the marginal megawatt-hour a new data center pulls is disproportionately fossil. The standards clock: the GHG Protocol Scope 2 revision, with publication targeted for late 2027, is moving toward hourly granularity and deliverability requirements for market-based claims. A facility scoped in 2026 against an annual-REC mental model may find its headline carbon claim quietly invalidated by the time it commissions.

Before the procurement strategy there is the boundary question: what counts as your emissions at all. The GHG Protocol's three scopes map onto a data center cleanly, and getting the mapping right is what makes a disclosure auditable rather than aspirational.

Scope 1 — direct combustion you own. For most grid-connected facilities this is small: diesel genset test runs, backup generator hours during outages, refrigerant leakage (a high-GWP line item that is easy to forget), and any on-site gas. But the moment you put a behind-the-meter gas plant on site to beat the interconnection queue, Scope 1 stops being a rounding error and becomes the dominant term — a campus burning gas for primary power can emit more in Scope 1 than a grid-tied peer emits across all three scopes. That inversion is the carbon cost of speed-to-power, and it is the central tension of the on-site section below. → Chapter 3.2.

Scope 2 — purchased electricity. For a grid-connected AI facility this is the overwhelming majority of operational carbon, and it is the scope where the matching decision lives. It is reported two ways (location-based and market-based, below), and the divergence between them is precisely the space in which REC accounting operates.

Scope 3 — everything upstream and downstream. For data centers the dominant Scope 3 category is embodied carbon: the concrete and steel of the shell, and increasingly the manufacturing footprint of the GPUs, networking, and power/cooling plant. On a clean grid, embodied carbon is now 50-80% of lifecycle emissions — the operational story you optimize so hard becomes the minority of the total. That reversal is large enough to warrant its own chapter. → Chapter 15.6.

Annual matching is the incumbent regime: over a calendar year, buy enough renewable generation (directly via PPA or as unbundled certificates) to equal your total consumption, and you may claim a carbon-free market-based Scope 2. It is cheap, liquid, and — on its own terms — easy to hit 100%. Its flaw is temporal and geographic blindness. A 100%-matched facility on an annual basis is still drawing fossil power during every hour that its contracted wind and solar are not generating, which on most grids is a large fraction of the year. The certificate says noon-in-Texas; the load is 2 a.m.-in-Virginia. Annual matching can be fully satisfied while the physical grid serving the load barely decarbonizes — and at the margin, a high-load-factor AI campus that draws flat power 24 hours a day is the worst-case consumer for annual matching, because so much of its draw lands in hours when intermittent renewables are absent.

24/7 carbon-free energy (CFE) closes the temporal and geographic gap. The discipline is to match consumption with carbon-free generation every hour, on the same grid. Google's CFE Score formalizes this: the share of each hour's consumption met by carbon-free energy on the local grid, averaged across the year. The difference in difficulty is enormous. Google's fleet sits at roughly 66% CFE globally (2024 data, 2025 reporting), with wide regional spread — Latin America near 92% and Europe in the 80s, but Asia-Pacific around 12% where the grid is coal-heavy and clean PPAs are scarce. The last 20-40 percentage points of hourly CFE are the hard part, and they cannot be bought with more solar — they require clean-firm supply that generates at night and in still air.

The table is a strategic posture, not a checkbox. Annual matching is the cheap, defensible-today path that is becoming a stranded claim. 24/7 CFE is the expensive, hard, but durable path that aligns with where the standard is going and where regulators and sophisticated customers are already looking. The decision is not purely virtue: it is a hedge against the standards clock. If you scope a facility today against annual RECs and the revised Scope 2 standard lands in late 2027 with hourly + deliverability teeth, you may have to re-procure under worse market conditions, with a load that is already energized and a claim that has already been published. → disclosure mechanics in Chapter 15.7.

The arithmetic of 24/7 forces a conclusion: you cannot reach high hourly CFE on intermittent supply alone without absurd overbuild. Firming ideal-site solar to a flat 24/7 profile takes roughly a 7x overbuild plus storage (a contested, single-source figure that swings with site and storage assumptions), and the economics break down before you reach the last few percent of hours. The answer is clean-firm power — carbon-free generation that runs regardless of weather or time of day — and the procurement question is which clean-firm lever you can actually contract on your timeline. Each carries a different bet on technology maturity, cost, and schedule.

Nuclear — restarts and uprates first, SMRs later. The fastest clean-firm megawatts in 2026 come from existing reactors: restarts and uprates of operating fleet. Microsoft's 20-year, 835 MW offtake from the restarted Three Mile Island Unit 1 (rebranded Crane Clean Energy Center, targeting ~2028) and Amazon's ~2 GW arrangement with Talen at Susquehanna are the template — large, firm, long-tenor, and behind a known reactor. Small modular reactors (SMRs) are the next wave but a slower one: Google's Kairos offtake (~500 MW, first units ~2030+) anchored the first corporate SMR PPA in 2025, and the data-center SMR offtake pipeline has grown to roughly 45 GW by 2026 — but FOAK SMR capex is $6,400-12,700/kW and realistic at-scale deployment is 2032-2035. SMRs are a 2030s clean-firm bet you contract now and wait on. → Chapter 3.2.

Enhanced geothermal — the dark-horse 24/7 baseload. Enhanced geothermal systems (EGS) moved from pilot to commercial in 2025. Google's 115 MW 24/7 EGS PPA with Fervo Energy (delivered via NV Energy) is a clean-firm baseload product that generates around the clock with no fuel and no combustion — exactly the profile that fills the hours solar cannot. Fervo's 500 MW Cape Station (Utah) is the scale-up. Geothermal's appeal is that it is firm and carbon-free without the licensing tail of new nuclear, though it is geographically constrained to favorable resource basins.

Gas-plus-CCS — clean-firm in name, conditional in practice. Natural gas with carbon capture and storage is marketed as clean-firm and can be sited where nuclear and geothermal cannot. The catch is capture rate and lifecycle accounting: a 90%-capture plant still emits, upstream methane leakage is not captured at all, and the economics depend on storage geology and incentives that are not universal. Treat gas+CCS as a partial-decarbonization firm option, not a zero-carbon one — and disclose the capture rate, not just the nameplate.

The interconnection queue is the binding constraint of the power-bound era — 3 to 7+ years end-to-end on most ISOs — and the dominant workaround is to generate on site, behind the meter, with gas. This is the cleanest example in the entire sustainability domain of a decision where the operational win and the carbon loss point in opposite directions, and you cannot have both. Behind-the-meter gas can deliver primary power in 18-36 months versus a half-decade in the queue; roughly 82 GW of it has been announced since January 2025. Every one of those megawatts is Scope 1 combustion that a grid-tied facility would have reported (largely cleaner) as Scope 2.

The fork is not gas-or-nothing; it is a spectrum of on-site options with very different carbon profiles. The decision turns on three questions: is this bridge power (run gas now, transition to grid/clean-firm when the interconnection clears) or permanent primary generation; what is the fuel and capture posture; and what does the local community and air permit allow. The carbon and social-license consequences compound — a permanent gas campus is a 20-year emissions liability and an air-quality fight, while a temporary-turbine bridge under the EPA's <850 MMBtu/h, ≤24-month subcategory is a defensible interim posture.

The supply side (what power you procure) has a demand-side complement: when and where you run the work. Carbon-aware scheduling shifts flexible compute toward hours and regions where the grid's marginal emissions are lowest. It comes in two flavors. Temporal shifting moves interruptible work — batch inference, eval sweeps, embeddings generation, non-urgent fine-tuning — into clean hours (midday solar, windy nights), throttling or pausing during dirty peaks. Spatial shifting routes flexible work to the cleanest available region. The mechanism is the same one that makes a workload curtailable for the grid, and the carbon lever and the flexibility lever are two readings of the same capability.

The discipline here is that scheduling for carbon costs goodput, and goodput is the metric that pays for the building. Every hour you defer a job to chase clean power is an hour that expensive accelerators sit idle against a depreciation clock. The math only works for genuinely interruptible, latency-insensitive work — and that excludes the revenue workload for most operators. Online inference has a hard latency SLO and cannot be time-shifted; synchronous pre-training runs continuously and is geographically pinned to its cluster; the deferrable surface is the batch and offline tier. The decision is to identify exactly which fraction of your fleet is carbon-flexible and cap the goodput you are willing to trade for it — not to declare the whole facility carbon-aware and quietly miss SLAs. The same scheduling primitive, pointed at price and grid-stress signals instead of carbon, becomes a grid-services revenue lever. → flexibility economics in Chapter 15.8.

Carbon offsets — paying for emissions reductions elsewhere to compensate for your own — sit at the bottom of the hierarchy for a reason. The mitigation order is unambiguous: avoid (efficiency, reduce the draw), then match with clean supply (PPAs, 24/7 CFE, clean-firm), then — only for the genuinely irreducible residual — offset. The failure mode the industry keeps repeating is using cheap offsets as a substitute for the harder, more expensive work of hourly-matched procurement, because an offset retired today is cheaper than a clean-firm PPA contracted for a decade. That substitution is exactly what the GHG Protocol revision and serious customers are moving to disallow in market-based claims.

The decision discipline: offsets may appear in your inventory only against residual emissions you have already minimized through avoidance and matching, they must be high-integrity (additional, permanent, verified — durable carbon removal preferred over avoidance credits), and they must be disclosed separately, never netted silently against gross emissions to manufacture a "net zero" headline. An offset-heavy carbon claim on a facility that never seriously pursued 24/7 matching is the disclosure equivalent of the annual-REC gap: technically reportable, increasingly indefensible. → reporting frameworks and audit posture in Chapter 15.7.