The Emissions Are Recorded in the Wrong Place

Commentary on "AI Giants Are Quiet on Climate in Sign of Post-ESG Wall Street" · Bloomberg

The most consequential fact about AI's environmental footprint is not how large it is. It is that the fastest-growing part of it does not appear in anyone's accounts.

Bloomberg reports that OpenAI and Anthropic, both approaching public offerings, have not disclosed greenhouse gas emissions, made net-zero commitments, or published sustainability reports. The hyperscalers building the infrastructure underneath them, Google, Meta, Amazon and Microsoft, all report, and their numbers are climbing sharply with the data centre buildout. Amazon's emissions rose 16 per cent in 2025, to roughly 81 million tonnes. The regulatory pressure that would have forced the question has receded: the SEC has moved to scrap climate disclosure rules, and the political consensus that sustained voluntary reporting has broken down. Bloomberg + 3

The story is framed as a governance failure, and at one level it is. But the more durable problem is structural, and it will survive any change in political weather.

Where an AI workload's emissions actually land

A lab that trains models on rented infrastructure does not consume electricity in any accounting sense. Its landlord does.

Under the standard framework, the power drawn by that training run appears in the operator's Scope 2, because the operator holds the meter and the utility contract. For the lab, the same electricity is Scope 3, category eight, one of the least reported categories in the least reported scope. So a single megawatt hour is recorded once, in the accounts of the party that supplied the building, and attributed to a business whose commercial interest in the workload is a lease payment.

This is not evasion. It is the framework working as designed for an era when the tenant and the beneficiary were usually the same firm. Applied to a sector where compute is rented at enormous scale by companies that own no facilities, it produces a systematic misattribution: the emissions of artificial intelligence show up on the balance sheets of real estate and cloud businesses, and the companies whose demand created them report nothing at all.

The consequence is visible in the Bloomberg figures. Hyperscaler emissions are spiking, which reads as a Big Tech problem. A significant portion of that increase is demand from customers who face no obligation to account for it.

Which makes the disclosure that does exist misleading

An investor comparing a hyperscaler to a frontier lab today sees one company with a large and worsening carbon number and another with none. The natural inference is the wrong one.

That distortion has consequences beyond investor confusion. It shapes which companies attract shareholder resolutions, which get local opposition, and which face regulatory attention. Activist proposals have gone to Amazon, Alphabet and Meta, because those are the companies with published figures to argue about. Nobody can file a resolution against a number that does not exist.

It also lets an entire industry describe itself as efficient while the measurement runs through someone else's meter.

Voluntary disclosure is worth more now, not less

The conventional reading of a receding regulatory environment is that disclosure becomes optional and therefore pointless. The opposite is closer to true.

When everyone reports because a regulation requires it, a report signals nothing. When almost nobody reports and the requirement has been withdrawn, a report is information. It tells a customer, a lender, and a host community something they cannot obtain any other way, and it tells them the operator was willing to be measured before anyone made them.

For infrastructure operators specifically, the case is stronger than for the labs. A data centre is a physical asset in a specific community, drawing on a specific grid, competing for specific resources. Its Scope 1 and 2 are knowable with reasonable precision from metered consumption and published grid intensity factors. There is no technical obstacle. There is only the question of whether an operator would rather have the number in public or be asked for it later.

Everything this series has argued about municipal permitting applies here in a different register. Projects that will not disclose get stopped. Companies that will not disclose eventually get legislated at, usually badly and by people working from worst-case assumptions.

What a serious disclosure baseline looks like

Four things, none of them exotic.

Scope 1 and Scope 2 stated annually against a defined boundary, with the grid intensity factor and its source named, because the same facility on the same load produces wildly different figures depending on which factor is used and which year it comes from.

Water withdrawal and consumption stated separately. They are not the same measurement and the difference is where most of the argument sits.

Energy efficiency stated as a measured operating figure rather than a design specification.

And Scope 3 acknowledged honestly, including the parts that cannot yet be quantified. Embodied carbon in compute hardware is substantial and poorly characterised across the industry. Saying so is more credible than an estimate nobody can defend.

What this means going forward

The regulatory reprieve is temporary. California's threshold already reaches some of these companies, European rules reach others, and the political cycle that removed the SEC requirement will eventually turn.

The operators and labs that spend this window building a defensible baseline will be answering questions in 2029 from a position of having already answered them. The ones treating the reprieve as permission will be assembling their first inventory under deadline, in public, with a number they have never seen before and no ability to explain how it changed.

Nobody is making them say. That is precisely why saying is worth something.

VOLTEDGE

Reference: "AI Giants Are Quiet on Climate in Sign of Post-ESG Wall Street" · Bloomberg · read the article