Syncing your local AI training data shouldn't spike the grid. Here's how to use solar production curves and scheduling APIs to automate data egress — and why it matters more than ever.
The computing industry is facing an uncomfortable arithmetic problem. Data centres consumed approximately 460 TWh of electricity globally in 2024, a figure the International Energy Agency (IEA) projects will more than double to around 945 TWh by 2030 — roughly equivalent to Japan's entire current electricity demand. AI is the primary driver. Electricity demand from AI-focused data centres is expected to triple over the same period, and in 2025 alone, data centre electricity use surged by 17%, far outpacing global electricity demand growth of 3%.
This is the context in which "net-zero infrastructure" has moved from aspirational to urgent. Early sustainability efforts focused on hardware power efficiency — measured by Power Usage Effectiveness (PUE) — and those gains have mostly been captured. Google's global fleet, one of the world's most efficient, averaged a PUE of 1.09 in 2024, compared to the industry average of 1.56. The hardware optimisation story is largely written. The next frontier is about when and where energy is consumed — not just how much of it.
At the forefront of this shift is Solar-Scheduled Egress: a methodology that dynamically aligns heavy data transfers with local renewable energy surpluses, turning the network fabric itself into a carbon-aware participant in the sustainability stack.
For most of the last decade, organisations tried to neutralise their carbon footprint by purchasing Renewable Energy Certificates (RECs) to match total annual energy consumption. The logic seemed sound: buy a certificate for every MWh consumed, and your operations are "renewable." The reality is more complicated.
RECs don't require geographic or temporal matching between generation and consumption. A company can purchase RECs from a wind farm in one region and apply them against electricity consumed from a coal-heavy grid elsewhere, at any time of year. On an hourly basis, a data centre may still be drawing heavily carbon-intensive power when solar and wind are offline, while the REC accounting shows clean energy use. This gap has drawn sustained scrutiny.
In October 2025, the GHG Protocol released proposals to move away from annual REC matching toward hourly and regional matching requirements — a fundamental shift in how electricity-related emissions are calculated. Final standards are expected in 2027. The direction of travel is clear: annual averaging is no longer considered sufficient for credible net-zero claims.
Leading companies are already moving. Google has pioneered 24/7 Carbon-Free Energy (CFE) matching, which requires renewable energy generation to match consumption in every hour, in every region. As of 2024, Google achieved a global average of 66% CFE, with nine out of twenty grid regions reaching at least 80% hourly matching. The company signed contracts for over 8 GW of additional clean energy capacity in 2024 alone — the largest annual total in its history. But even at Google's scale and resources, full 24/7 matching has proven difficult, particularly in Asia-Pacific where structural grid constraints keep hourly matching rates in the low double-digits.
The lesson is that purchasing clean energy is necessary but not sufficient. Physical reality demands that clean energy be available at the moment of consumption. This is what makes demand-side flexibility — scheduling operations to coincide with renewable generation — so strategically important.
To understand solar-scheduled egress, it helps to understand carbon intensity as a time-varying signal. The carbon intensity of electricity — measured in grams of CO₂ equivalent per kilowatt-hour (gCO₂eq/kWh) — fluctuates dramatically throughout the day based on which generation sources are active. During daylight hours in solar-rich regions, carbon intensity can drop to near zero. At night or during overcast periods, fossil-fuel peakers and gas plants often fill the gap, pushing intensity sharply higher.
Hyperscalers have built sophisticated systems to exploit this variability on the compute side. Google's Carbon-Intelligent Computing System uses carbon intensity forecasts from providers such as Electricity Maps to reshape intraday CPU usage across its global data centre fleet — throttling or accelerating compute to align with cleaner grid periods, without impacting service quality. Microsoft and others have implemented similar demand-shifting frameworks.
What has received less attention is the network fabric itself. Moving massive datasets — AI model checkpoints, high-fidelity 3D assets, years of historical telemetry — across wide-area networks requires routers, switches, and optical amplifiers that draw significant power. The operational carbon cost of a terabyte transfer is not just a function of the data centre's energy mix; it's also a function of the energy mix powering every network device along the path, at the moment of transmission. This realisation is what motivates renewable-aware networking.
The foundational concept is spatiotemporal flexibility. A workload has temporal flexibility if it can be delayed until the local grid's carbon intensity drops. It has spatial flexibility if it can be rerouted to a different geographic region with a greener energy mix. Research published in 2025 reviewing carbon-aware scheduling across both edge and cloud environments found a clear trend: while earlier studies tended to treat temporal and spatial shifting separately, the most recent literature increasingly treats them as a combined, integrated approach — because the biggest gains come from doing both simultaneously.
Data egress operations are naturally suited to this kind of flexible scheduling. Consider a modern industrial environment relying on the Industrial Internet of Things (IIoT). Physical machinery is outfitted with thousands of sensors tunnelling telemetry to a cloud-based digital twin. Not all of this data requires instantaneous high-bandwidth transmission. It can be logically partitioned into two streams:
State-critical telemetry carries lightweight payloads — operational state, fault flags, safety overrides. This requires ultra-low latency but consumes negligible bandwidth. It must flow continuously regardless of grid conditions.
Bulk analytical logs and spatial assets carry heavyweight payloads — high-resolution historical sensor logs, dense point-cloud scans, 3D geometric updates. A multi-gigabyte update to a digital twin's texture library or historical analytics database does not need to hit the cloud the second it is generated. It has massive temporal flexibility.
Recognising this distinction allows network engineers to implement a bifurcated tunnel architecture: a permanent low-bandwidth trickle connection for critical telemetry, and a scheduled high-bandwidth pipe that only opens when renewable energy is available.
The implementation follows a clear pattern. Under normal conditions, when local solar production is low — at night or under heavy cloud cover — the egress tunnel operates in restricted trickle mode. State-critical IIoT telemetry flows uninterrupted. Network interfaces and edge routers run at lower power states. Heavy analytical logs and 3D asset updates accumulate in a local edge queue, tagged with priority levels.
The scheduling algorithm continuously polls local energy generation infrastructure — smart inverter APIs, building energy management systems, or external carbon-intensity oracles such as Electricity Maps or WattTime. The trigger condition is straightforward: open the high-bandwidth pipe only when local renewable generation exceeds baseline operational demand. When the solar array produces a surplus, the network controller receives the signal and dynamically provisions a high-bandwidth connection, spinning up parallel TCP or QUIC streams to maximise throughput. Queued data flushes to the cloud repository.
Because the energy powering routers, switches, and edge servers during this burst phase is 100% locally generated solar surplus, the operational carbon cost of the transfer approaches zero. When solar output drops — a passing cloud, a late-afternoon dip — the tunnel throttles back to trickle state, pausing the bulk transfer without dropping the critical telemetry connection.
The elegance of this design is that it requires no changes to the applications generating or consuming the data. Carbon-awareness is enforced at the network scheduling layer.
Local solar scheduling addresses the source side of the carbon equation. But a comprehensive approach must also account for the destination data centre. Delaying a transfer to use local solar power, only to push that data into a facility running on coal, achieves little.
This is where carbon-aware spatial routing becomes essential. Modern multi-cloud delivery controllers monitor real-time carbon intensity across geographically dispersed availability zones. When the local solar trigger opens the high-bandwidth pipe, the routing logic evaluates destination options based on their current grid conditions.
Consider an enterprise maintaining digital twin repositories in both AWS eu-central-1 (Frankfurt) and AWS eu-north-1 (Stockholm). Even if the local edge facility has a solar surplus, the destination facility will consume energy to ingest, process, and write the incoming data. If Stockholm is currently drawing on wind power surplus while Frankfurt is relying on natural gas, the router directs the bulk egress to Stockholm. This dual-axis optimisation — temporal shifting at the source, spatial shifting at the destination — is what creates a genuinely net-zero data pipeline.
Solar energy is inherently variable. Diurnal cycles and unpredictable weather mean that solar-scheduled pipelines need resilience mechanisms to prevent edge storage from overflowing during extended low-generation periods.
Predictive analytics are central to this. By integrating weather forecasting models and historical solar production data, scheduling algorithms can estimate expected yield over a 48-hour horizon. If forecasts predict sustained heavy cloud cover, the system can calculate whether the edge queue will exceed capacity before the next solar surplus and activate fallback strategies — such as scheduling egress during the lowest-carbon hours of the public grid (often late night, when regional wind power dominates), rather than waiting for strict local zero-carbon solar.
Advanced scheduling implementations in Kubernetes-based orchestration systems can formalise this as a multi-objective optimisation problem, enforcing Service Level Objectives (SLOs) alongside carbon minimisation. Latency constraints are respected even as bulk transfers are delayed or rerouted.
There is also a broader structural backstop developing in the market. The IEA reports that the tech sector accounted for roughly 40% of all corporate power purchase agreements for renewables signed in 2025. The pipeline of conditional agreements between data centre operators and small modular reactor (SMR) nuclear projects has grown from 25 GW at the end of 2024 to 45 GW by early 2026 — dispatchable clean energy that, unlike solar and wind, is not subject to intermittency. As these energy sources come online, the fallback conditions that solar-scheduled systems need to handle will become progressively less carbon-intensive.
Solar-scheduled egress sits within a larger sustainability architecture that the industry is building in parallel. On the accounting side, the GHG Protocol's move toward hourly and regional REC matching will force organisations to measure and report carbon at a granularity that makes time-shifting directly legible in their sustainability disclosures — creating direct regulatory and reputational incentives for adopting approaches like solar-scheduled egress.
On the energy side, renewables currently supply about 27% of electricity consumed by data centres globally, according to the IEA. Natural gas remains the largest single source for US data centres at over 40%. The gap between renewable procurement claims and real-time renewable consumption is the terrain that carbon-aware scheduling directly addresses.
On the infrastructure side, the edge data centre is increasingly integrated into the physical environments it serves — industrial floors, rooftop solar microgrids, ambient heat exchangers — rather than existing as a remote, pristine facility. This architectural evolution makes local solar surplus a practical and observable signal, not an abstraction.
The transition to carbon-neutral data logistics requires more than offset accounting. It demands that the actual moment of energy-intensive operations be aligned with the actual availability of clean energy.
Solar-scheduled egress provides a practical, deployable path toward this goal. By separating state-critical telemetry from bulk data transfers, network engineers can maintain real-time connectivity while sequestering energy-intensive operations for periods of renewable abundance. When combined with carbon-aware spatial routing that selects the greenest available cloud destination, the approach creates a data pipeline whose operational carbon footprint can approach zero.
As AI model sizes grow, IIoT sensor density increases, and the GHG Protocol tightens its emissions accounting standards, the pressure to implement these patterns will intensify. The tools exist today. The data on grid carbon intensity is available in real time. The scheduling logic is well understood and increasingly standardised.
The heaviest digital payloads can ride the lightest environmental footprint. The remaining question is how quickly organisations choose to build that capability into their infrastructure.
Acun, B., Lee, B., Kazhamiaka, F., et al. (2023). Carbon Explorer: A Holistic Framework for Designing Carbon Aware Datacenters. Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, 118–132. https://doi.org/10.1145/3575693.3575754
IEA. (2025). Energy and AI. International Energy Agency. https://www.iea.org/reports/energy-and-ai
IEA. (2026). Key Questions on Energy and AI. International Energy Agency. https://www.iea.org/reports/key-questions-on-energy-and-ai
Jayaprakash, B., Eagon, M., Yang, M., Northrop, W., & Shekhar, S. (2023). Towards Carbon-Aware Spatial Computing: Challenges and Opportunities. I-GUIDE Forum 2023. https://doi.org/10.5703/1288284317678
Maji, D., Pfaff, B., P R, V., et al. (2023). Bringing Carbon Awareness to Multi-cloud Application Delivery. Proceedings of the 2nd Workshop on Sustainable Computer Systems, 1–6. https://doi.org/10.1145/3604930.3605711
Radovanovic, A., Koningstein, R., Schneider, I., et al. (2021). Carbon-Aware Computing for Datacenters. arXiv. https://doi.org/10.48550/arxiv.2106.11750
Rocha, P., et al. (2025). Carbon-Aware Spatio-Temporal Workload Shifting in Edge–Cloud Environments: A Review and Novel Algorithm. Sustainability, 17(14), 6433. https://doi.org/10.3390/su17146433
Souza, A., Jasoria, S., Chakrabarty, B., et al. (2023). CASPER: Carbon-Aware Scheduling and Provisioning for Distributed Web Services. Proceedings of the 14th International Green and Sustainable Computing Conference, 67–73. https://doi.org/10.1145/3634769.3634812