Market-based vs location-based emissions: the NVIDIA case

NVIDIA’s technical blog said in March 2026 that token consumption across the industry now exceeds 10 quadrillion tokens a year. The company that sells the accelerators those tokens run on publishes a sustainability report every June, and the 2026 edition is a useful document for anyone who reports AI usage, because it shows in one table how the accounting works and where it stops. This post reads that table, explains the mechanism behind its biggest number, answers whether NVIDIA is setting an example or moving dirty electrons around, and ends with the wording we use for our own product and why.

The two numbers in plain words

Picture the grid as a punch bowl at a party. Six guests pour in juice and four pour in cheap syrup, and everyone drinks the same mix. Now suppose the guests who brought juice can sell receipts, each saying that a jug of juice went in on your behalf. Buy six and your paperwork says you drank pure juice, although your cup held the same mix as everyone else’s.

Location-based accounting measures what was in the bowl. Market-based accounting counts the receipts. A renewable energy certificate is one of those receipts, a record that one megawatt-hour of clean electricity was generated somewhere, which the holder can claim as their own. Buying one changes nothing about the power that reaches the buyer’s building, since electricity on a shared grid cannot be sent to one customer.

Market-based vs location-based: two figures for the same electricity

The GHG Protocol’s Scope 2 Guidance asks a company to report the emissions behind its purchased electricity twice. The location-based figure applies the average emissions intensity of the grid where the power was used. The market-based figure applies the contracts the company holds, so electricity covered by an energy attribute certificate, a power purchase agreement or a utility’s renewable tariff is counted at the emissions of that generator, which for wind and solar is zero. A company that buys enough certificates to cover its consumption reports a market-based figure near zero while its location-based figure carries on describing the grid.

NVIDIA’s table for the fiscal year that ended in January 2026 shows what that looks like at scale.

The location-based figure grew 73% in two years as the company took on more offices, labs and leased data centre space. The market-based figure went from 40,555 to zero to 568 over the same period. The report explains the mechanism in one sentence: in FY26 NVIDIA purchased or generated enough clean electricity to match 100% of its global electricity usage, through on-site solar, long-term power purchase agreements, utility renewable electricity tariffs and energy attribute certificates. Under the market-based method, that is a 99.8% reduction. Under the location-based method, nothing was reduced.

Both figures are correct, and NVIDIA reports both, which is what the protocol requires. PricewaterhouseCoopers reviewed them and, in a report dated 5 June 2026, said it was not aware of any material changes they needed. That is limited assurance, a lighter check than a full examination, and the methods it reviewed are NVIDIA’s own. The question for a reader is what each one is a measurement of.

What a renewable energy certificate does, and what it leaves alone

Electricity cannot be traced once it is on the grid. Every generator feeds the same pool, every consumer draws from the mixture, and no one receives the wind farm’s electrons in particular. Certificates exist to fill that gap with paperwork. One certificate (the formal name is an energy attribute certificate, and in the United States a renewable energy certificate) records that one megawatt-hour of clean electricity was generated somewhere, and buying it gives the buyer the claim to that megawatt-hour. The buyer still draws the same mixture as everyone else on the grid.

A grid’s emissions factor is its total emissions divided by its total generation, and both are physical quantities. Buying a certificate does not start a turbine or switch off a gas plant, so neither number changes. What changes is who writes which figure on their report. The guidance deals with the other side of that trade through a residual mix, the grid’s mixture with the claimed clean megawatt-hours taken out. It is dirtier than the average, and everyone on that grid without certificates is supposed to report against it. Residual mix factors are published in Europe and almost nowhere else, so most reporters use plain grid averages that still contain the renewables somebody has already claimed. Summed across every company, market-based Scope 2 comes in below what the grids emitted, and the difference belongs to nobody’s inventory.

Two more limits sit inside the word matched. The match is annual and global by volume: a certificate from a wind farm generating at three in the morning in one market can cover a GPU drawing gas-fired power at six in the evening in another. And some of the consumption being matched was never metered. NVIDIA’s management assertion, which PwC reviewed, says its data centre electricity was estimated from capacity and equipment utilisation assumptions, and that estimates make up about 16% of the location-based Scope 2 figure and all of the market-based one. Every one of the 568 tonnes left after matching comes from estimated consumption.

The exception worth taking seriously is whether the money caused new clean generation to be built, which accountants call additionality. If the revenue from a contract is what got a new solar farm built, the grid genuinely has more clean generation than it would have had, and the factor improves for everyone. Long-term power purchase agreements can carry that claim. Unbundled certificates trade cheaply enough that they rarely decide whether a project goes ahead. NVIDIA’s report lists both instruments in its sourcing mix without saying how the total splits between them.

The 200,000 tonnes that stay on the books

Certificates cover electricity the company itself buys or generates. Two Scope 3 lines in the same table describe electricity-related emissions that no certificate reaches.

Together they are about 203,000 tonnes, roughly two thirds of the location-based Scope 2 figure, sitting in categories a reader has to scroll past the headline to find. The clean electricity percentage that reaches 100% is defined against electricity NVIDIA purchases or generates, and the report describes it as a point-of-generation measure that does not address full life cycle impacts. The landlord’s cooling load and the losses on the wires are outside that denominator. The 100% is true as defined. It does not describe all of the energy it takes to run a GPU for a year.

One more absence matters for anyone reading the table as an AI buyer. The categories stop at 8. Category 11, use of sold products, is where the electricity consumed by the GPUs NVIDIA sells would sit, and it is not in the inventory. The company does publish a target for it: a 75% cut in Scope 3 emissions intensity from the use of sold GPU products per petaFLOP by FY30, from an FY23 base. An intensity target without a tonnage line means the largest emissions the company influences appear nowhere in its reported total.

How a zero reaches you

This section is reasoning about standard practice rather than something NVIDIA discloses, and it is the part that matters most for a company reporting its own AI usage.

A company that buys hosted inference reports it under Scope 3, Category 1, purchased goods and services, which our guide to Scope 3 for AI walks through. The protocol’s calculation guidance ranks four methods for that category, and most companies use the last one, spend-based, which multiplies dollars by an industry factor and has nothing to do with electricity at all. The better method is a supplier-specific factor. Ask a provider for one and the usual practice is to derive it from the provider’s own market-based Scope 2, which is near zero after matching. The customer inherits the zero.

Follow one request through a modern stack and the chain gets longer: a routing layer, a model provider, a hyperscaler. The operator matches with certificates and reports near zero. The provider inherits it. The router inherits it. The end user inherits it. Each of the four layers is compliant on its own, and the physical megawatt-hours have no home after the second one. There is no clearing house in corporate emissions accounting that forces the residual onto anyone’s books.

CrbonFree does not derive its factors from suppliers’ market-based figures. It meters tokens per model and per provider and applies location-based grid intensity through the provider’s published power usage effectiveness, so the number describes what the grid was burning where the model ran. The methodology lists every factor, and the GHG Protocol guide shows where the figure sits in a corporate inventory. A customer who wants the market-based view can ask their provider for it and report both, which is exactly what the protocol asks of the provider.

Where the standard is heading

The GHG Protocol opened a public consultation on the first major revision of its Scope 2 Guidance on 20 October 2025, and it closed on 31 January 2026. The central proposal is an hourly matching and deliverability requirement for the market-based method: certificates would have to be matched to the hours in which the electricity was consumed and come from a grid region that could plausibly have delivered it. Dual reporting stays. A second consultation follows in 2026, and the new Scope 2 Standard is expected in 2027. Because the EU’s reporting standards, California’s SB 253 and the investor frameworks all reference the protocol, the change propagates into mandatory reporting once it lands.

The main objection raised against hourly matching is that the data is hard to collect. That objection is an argument for metering, which is the whole of what a usage-level measurement tool does.

What beyond value chain mitigation means, and why we use the words

CrbonFree measures a customer’s AI usage per token and, on paid plans, retires verified carbon credits in a matching quantity. Every part of that sentence has a specific meaning, and it is worth being precise about what it does not mean.

Retiring credits does not reduce the customer’s Scope 1, 2 or 3. The Science Based Targets initiative’s term for it is beyond value chain mitigation, defined as mitigation action or investment that falls outside a company’s value chain, and the initiative places it below reductions in its mitigation hierarchy: set and meet the targets first, then invest outside the value chain in addition. The measured footprint is reported unchanged. The retirement is reported next to it, with the registry, the project, the serial numbers and the vintage, so an auditor can check each line on its own.

The wording also has a legal edge. The EU’s Directive 2024/825 applies from 27 September 2026 and adds to the list of practices banned in all circumstances any claim, based on mitigation outside the product’s value chain, that a product has a neutral, reduced or positive impact on the environment in terms of greenhouse gases. Its recital names the phrases: climate neutral, CO2 neutral certified, carbon positive, climate net zero, climate compensated, reduced climate impact, limited CO2 footprint. Companies may still describe their investments in credit projects, as long as the description is not misleading. A product page that says the tokens have been cancelled out crosses the line. One that says this usage measured 1.2 tonnes, and 1.2 tonnes of verified credits were retired against it on these serial numbers, does not.

That is the language we use, and it is the language a sustainability officer can put in a disclosure without a footnote.

Is NVIDIA setting an example, or moving dirty electrons around?

Certificates don’t move electrons in either direction, so NVIDIA hasn’t pushed dirty power onto anyone else’s grid. What it bought is the right to report its own electricity as clean, which the GHG Protocol allows, and it published the location-based figure alongside the market-based one. Part of its sourcing is long-term power purchase agreements, the kind of contract that can get a new wind or solar farm built. Part is unbundled certificates, which rarely decide whether anything gets built. The report doesn’t give the split, so from outside there is no way to tell how much new clean generation the 100% paid for.

The report is worth copying for publishing both figures and for defining exactly what its percentage covers. The 568 tonnes are the wrong number to give a board or an auditor who asks how much the company’s AI emitted, because that question is about the grid. An AI company can answer it more directly than NVIDIA’s report does, by counting its tokens, applying the intensity of the grid where they ran, and reporting the market-based figure next to that result without calling either one neutral.

What to ask a provider

1. Both Scope 2 figures, location-based and market-based, not one or the other.
2. Whether the matching is hourly and from a deliverable region, or annual and global.
3. How much of the sourcing is long-term power purchase agreements and how much is unbundled certificates.
4. What sits in Categories 3 and 8, and whether the facility overhead is metered or assumed from an industry-average PUE.
5. The token count. NVIDIA meters its own internal AI factory and reports its output as years of human work rather than tokens or megawatt-hours. A provider that will not give a token count cannot give a defensible per-token factor either.

The methodology page has every factor and formula behind our per-token figures, the case study works a year of gateway traffic through them, and the sample account opens without a login.