Net Benefit AI, Measurement Integrity, and the Architecture of OCIS

Public debate about artificial intelligence and climate tends to settle into one of two positions: AI as an unqualified climate solution, or AI as an unaccountable cost to the grid. The Climate Action Coalition's new report, Net Benefit AI: Scaling Solutions, Opening Opportunities, co-chaired by Ambassador Patricia Espinosa and the Rt Hon Chris Skidmore OBE, takes neither position. It asks a narrower and more useful question: does a given application of AI deliver more decarbonisation, efficiency, and system-level value than it consumes in energy, water, and material resources across its own lifecycle? The report deliberately confines its case studies to energy systems, noting that AI's role in cutting emissions across nature and climate is significant enough to deserve a report of its own. That is, in effect, the gap the Ozeaon Climate Intelligence System, OCIS, is already building toward, organised around four functions where the same net-benefit test applies: project integrity scoring, dMRV for carbon and blue carbon, biodiversity and nature-market readiness, and AI footprint transparency.

The case for taking that question seriously starts with scale. The report puts global data-centre electricity consumption at roughly 448 terawatt-hours in 2025, more than Saudi Arabia's annual total, and enough to rank data centres as the world's eleventh-largest electricity consumer if they were counted as a single country. It cites a UN projection that this could climb toward 945 terawatt-hours annually by 2030, roughly three times the electricity that Pakistan, Bangladesh, and Nigeria together consume today, a group of countries home to more than 650 million people. The International Energy Agency estimates that emissions from generating that power will peak near 320 million tonnes of CO2 around 2030 and ease only slightly, to about 300 million tonnes, by 2035, as cleaner generation slowly displaces fossil fuel in the mix. None of this is offered as a reason to retreat from AI. It is offered as the baseline against which any claimed benefit has to be measured.

Measured against that baseline, the opportunity the report describes is large enough to matter. It leans on Lord Nicholas Stern's 2025 analysis, Green and Intelligent, to make the case: done well, AI could take as much as 5.4 billion tonnes of emissions out of the global total by 2035, comfortably more than AI's own infrastructure is projected to add over the same period. A separate IEA estimate suggests that AI tools already in everyday use across energy, industry, transport, and buildings could strip out 1,400 megatonnes of CO2 by 2035 under its Widespread Adoption scenario, three to four times more than the emissions those same data centres are expected to produce. The mechanism the report proposes for capturing that gap deliberately, rather than assuming it, is the World Economic Forum's Net-Positive AI Energy Framework. It rests on three levers: design for efficiency, deploy for impact, and shape demand wisely. Under this framework, an AI system only counts as net positive when the savings it documents across the real economy exceed the full lifecycle cost of running it. That is a higher bar than most AI deployment decisions are held to, and it is the bar each of OCIS's four functions is designed to clear.

The report's strongest argument, read carefully, is not really about energy at all. It is about information. The obstacle it keeps returning to is not a shortage of good intentions but a shortage of consistent, verifiable, comparable data: data centres typically report aggregated corporate or regional figures that collapse training, inference, storage, cooling, and networking into a single number; much of the carbon accounting in use still relies on broad annual or monthly regional averages, despite the fact that the actual carbon intensity of a grid can swing meaningfully within a single day; and the emissions embedded in hardware manufacturing, construction, and water use, the report's Scope 3 category, are seldom reported with enough granularity to act on. AI governance, in other words, has an evidence problem before it has an energy problem. That is the same problem Project Integrity Scoring exists to solve inside OCIS, applied to a different asset class. Carbon, blue carbon, and biodiversity projects have long suffered from a version of the same pattern: narrative-led pitches, inconsistent baselines, and claims that are hard to verify until capital has already moved. The report offers a working illustration of the alternative, even though its own case study sits in corporate energy transition rather than nature finance. Working with the AI platform Vyzrd, a global investment fund managing more than $300 billion in assets screened more than thirty portfolio companies against a consistent framework: emissions impact, financial return, capital intensity, implementation complexity, and scalability. The exercise turned an assessment that would normally run twelve to eighteen months into one that took three to five weeks, with a projected portfolio-wide emissions reduction of around 30 percent and more than $750 million in identified value. The sector is different, but the method, a structured, criteria-based screen that surfaces strengths, weaknesses, and risk before commitments are made rather than after, is exactly what Project Integrity Scoring is built to do for nature and climate project pipelines.

If Project Integrity Scoring screens projects before capital commits, dMRV is concerned with what happens afterwards: whether the claims a project makes can be checked, continuously, against reality. The report treats this as one of the unresolved problems in AI governance, and the regulatory response it documents is telling, because it amounts to a decision that self-reported, after-the-fact figures are no longer good enough. A new ITU recommendation finalised in February 2026, L.1801, lays out a single framework for tracing the direct and indirect environmental footprint of an AI system end to end; a companion IEEE standard pushes operators to report efficiency against the workload actually running rather than as one facility-wide average; and the EU's AI Act adds binding disclosure obligations for systems classified as high-risk. UNESCO-IOC's Blue Carbon Finance Toolbox makes the equivalent point for the ocean economy directly: reliable, consistent data are not a nicety for blue carbon finance, they are the precondition for credible markets, institutional capacity, sound policy, and durable ecosystem valuation. The Net Benefit AI report shows what it looks like to build that evidence layer where it did not previously exist. In Rajasthan, the Global Energy Alliance's Grids of the Future programme used AI-driven load-flow analysis and digital twins to turn a distribution grid that had been run largely on guesswork, with technical and commercial losses of 16 percent and metering rates as low as 20 percent, into one with location-aware, continuously verified operational intelligence, digitising 2.5 million grid assets and unlocking an estimated $53 million in annual savings. The underlying problem there, an un-gauged system in which nobody can say with confidence where value is leaking or whether an intervention worked, is structurally the same problem blue carbon and nature markets face today. OCIS's dMRV function is, in effect, building the equivalent evidence layer for ocean and nature projects.

The report is also candid about what happens when growth outruns the systems meant to govern it, and that caution carries directly across to nature markets. Developers, it notes, have often built data-centre capacity faster than grids, regulators, and host communities could absorb, and the fallout has been concrete: a cap on new Dublin-area connections in Ireland, a wave of paused hyperscale approvals across several Dutch provinces, and an estimated $77 billion in delayed or contested US capacity as of 2025. The lesson is not to stop building, but to sequence growth to the pace at which infrastructure, regulation, and public trust can actually absorb it. The European Commission's 2026 roadmap on nature credits in coastal and marine ecosystems draws the equivalent line for blue and nature markets, citing site-specific and multidimensional biodiversity outcomes, complex baselines and counterfactuals, expensive underwater monitoring, and real risks of leakage and non-permanence as reasons to favour evidence-led pilots over premature scaling. The Ocean Panel's Blue Carbon Handbook adds the other half of the discipline: that high-integrity blue carbon projects have to integrate mitigation, adaptation, resilience, biodiversity, livelihoods, local participation, and safeguards, not reduce a coastal ecosystem to a single carbon number. OCIS's biodiversity and nature-market readiness function is built around both constraints together, assessing projects on ecological condition, biodiversity significance, community relevance, and delivery feasibility rather than carbon potential alone, and treating evidence-led pacing as a feature of the system rather than a delay to be engineered around.

The report's final argument, and in some ways its most demanding, is that the standard being asked of AI broadly has to apply to the organisations building and running it, not only to the systems they sell. It points to specific commitments as evidence that this is achievable rather than aspirational: AWS has set a 2040 net-zero target and says it is already covering, region by region, all the electricity many of its data centres use through renewable purchases; Google has shifted from matching renewables annually to matching them hour by hour and location by location, aiming for fully carbon-free operation, AI infrastructure included, by 2030; Microsoft is aiming to go carbon negative by 2030, erase everything it has emitted since the company was founded by 2050, and is using an internal price on carbon to push that timeline along; and Meta has pledged net-zero across its entire value chain, not just its own operations, by the same year. The report's efficiency and demand-shaping principles translate this into something more specific still: train and run models that are sized to the job rather than defaulting to the largest option available, on the logic that plenty of work does not need a frontier-scale model to get done well, and shift workloads toward the hours when clean power is actually abundant rather than running them on demand regardless of grid conditions. That is the standard OCIS's own AI Footprint function exists to meet. If OCIS is going to ask carbon, biodiversity, and blue carbon projects to evidence their claims rather than assert them, the AI used to do that evidencing has to be held to the same test: measured, bounded, and justified by the value it adds rather than deployed simply because it is available. AI Footprint is not a compliance layer sitting alongside OCIS. It is the condition on which OCIS gets to ask anyone else to be transparent.

None of OCIS's four functions were designed by working backwards from this report. The convergence is still worth noting, because it suggests the underlying logic is more durable than any single document. The Coalition itself is now pushing for a global pledge on net benefit AI, pointing to the second Global Stocktake, due to open at CMA8 in November 2026 and close at CMA10 in November 2028, as the moment for it, much as COP28 produced its own pledge to treble renewable capacity. Whatever shape that multilateral architecture eventually takes, Ozeaon does not need to wait for it to apply the same discipline at the scale that matters most to ocean and nature finance: judge AI by what it demonstrably improves, build the evidence layer before the capital follows, sequence growth to the pace the science can support, and hold the system's own footprint to the standard it asks of everyone else. That, across project integrity scoring, dMRV, biodiversity and nature-market readiness, and AI footprint, is what OCIS is for.

References

Climate Action Coalition. (2026). Net Benefit AI: Scaling Solutions, Opening Opportunities: Defining the responsible role for AI in the energy transition.

European Commission. (2026, January 30). Roadmap towards nature credits – feedback and next steps: Nature credits in the EU's coastal and marine ecosystems.

International Energy Agency. (2025). Energy and AI. IEA, Paris.

International Telecommunication Union. (2026). ITU-T L.1801: Environmental sustainability of artificial intelligence and emerging ICT systems.

Schindler Murray, L., Milligan, B., et al. (2023). The blue carbon handbook: Blue carbon as a nature-based solution for climate action and sustainable development. High Level Panel for a Sustainable Ocean Economy.

Stern, N. (2025). Green and Intelligent, as discussed in Climate Action Coalition (2026).

UNESCO-IOC. (2025). Blue Carbon Finance Toolbox. UNESCO.

World Economic Forum. (2025). Net-Positive AI Energy Framework, as discussed in Climate Action Coalition (2026).