Anthropic Stood Up to the Pentagon. Now It Needs to Stand Up for the Planet.
Anthropic's public identity rests on a claim no other major AI lab makes as loudly: that building powerful technology responsibly means taking the consequences of what you build seriously, not just its capabilities. It's the premise behind Constitutional AI, behind the company's public safety research, behind Dario Amodei's own willingness to warn the world about the risks of the systems he is racing to build.
On 10 June 2026, Bloomberg's Katrina Manson and Emily Chang published “Inside Anthropic, the $965 Billion AI Titan” — close to fifty minutes of rare, in-depth access to Dario and Daniela Amodei. The conversation covered the company's origin story, its battles with the Pentagon, the race against OpenAI, and Claude's role in modern warfare, including an admission that Dario Amodei did not know what role Claude had played in a missile strike that killed an estimated 120 children at a school in Iran. It was a wide-ranging, candid hour, built around exactly the theme Anthropic has made its own: taking responsibility for consequences.
Not once in that conversation was Anthropic's carbon footprint mentioned.
This isn't a claim that Bloomberg avoided the subject, or that the Amodeis dodged it. Fifty minutes cannot cover everything, and the Pentagon story was, understandably, the news of the day. But the absence is worth sitting with, because of what the interview was actually about: a company that has built its reputation on looking squarely at the downstream effects of what it makes.
What Ethics Is Made Of
Some companies forget their raison d'être. Google went from “Don't Be Evil” — written into its code of conduct in the early 2000s — to “Don't Mention It”: the motto quietly dropped once Alphabet formed and scale met commercial reality. It's a reminder of what ethics actually requires to remain real: not a stated intention, but a consistent account of what a company does, how it does it, and what results from it. All three. Anthropic has built its identity on the first two. This piece is about the third.
The Footprint Gap
The facts are straightforward. Anthropic has not published Scope 1, 2, or 3 emissions figures, and has no publicly documented reduction target aligned to the Science Based Targets initiative. In May 2026, it signed a deal with SpaceX — which now owns xAI — for compute capacity at the Colossus data centres in Memphis, initially reported at 300 megawatts. SpaceX's amended IPO filing with the SEC, submitted 1 June 2026, disclosed the true scope: access to roughly 325,000 Nvidia GPUs across both Colossus and Colossus II, at $1.25 billion a month through May 2029. That same amended filing added water scarcity as a formal risk factor for the first time — a disclosure change tied directly to the infrastructure now running Anthropic's workloads. Colossus has already drawn criticism over the gas turbines used to power it, and over environmental hardship for nearby Memphis communities.
In June 2026, Anthropic joined Frontier, the carbon removal purchasing coalition founded by Stripe, Google and others, pledging $915 million toward direct air capture, bio-oil sequestration and related technologies — the first AI-pure-play company to do so. It is a genuinely notable step, and it deserves credit as one. Anthropic has also committed to absorbing electricity price increases its data centres cause for nearby consumers, and to cutting its own consumption by up to 30 per cent during peak grid demand.
All of this is real. None of it is measurement. A carbon removal purchase without a published baseline is an offset with no ledger to offset against — a company can announce it is buying back carbon without ever stating how much it produced in the first place.
Measurement Is Not a Courtesy
It would be easy to frame emissions reporting as a nice-to-have — a good corporate citizenship gesture sitting alongside the ethics, rather than inside it. That framing is wrong, and worth naming as wrong. A company whose differentiator is that it takes the consequences of its technology seriously cannot ring-fence that seriousness to model outputs while treating the electricity, water and emissions behind those outputs as someone else's business. Disclosure isn't a value Anthropic would be adopting from outside. It's the value it already claims, applied to where its own operations actually sit.
Location-agnostic abatement and undisclosed emissions are not the same thing, and the difference between them is the entire argument.
Abatement and removal, done well, don't need to happen where the emission occurs — the atmosphere carries no passport, and a tonne of methane destroyed in one country is worth exactly the same as a tonne destroyed anywhere else. That principle is sound. But it only holds together if the emission was measured somewhere in the first place. Location-agnostic abatement and undisclosed emissions are not the same thing, and the difference between them is the entire argument.
A Model Built to Scale Further, Faster, With Less Scrutiny
This is not a problem contained to one company. In February 2026, India's government announced a twenty-year tax holiday, running to 2047, for foreign cloud companies operating data centres on Indian soil — zero tax on revenue from services sold outside India, so long as the workloads run through Indian infrastructure. Several Indian states have layered their own incentives on top. Google, Microsoft and Amazon have already committed a combined $67.5 billion to data centre build-out in the country by decade's end. Globally, data centres accounted for more than a fifth of greenfield investment value in 2025, north of $270 billion, according to UNCTAD — and the incentive model India has adopted is the kind of framework more developing economies are likely to follow, precisely because it works as an investment magnet.
None of these frameworks carry any requirement to measure or disclose the resource cost of what they're attracting. That silence is being written into the incentive structures of countries with far less regulatory capacity than the US or EU to impose disclosure later, once the capital, the jobs and the political relationships are already locked in. The Scope 1, 2 and 3 ask isn't a complaint about one company's public relations. It's a request to set the norm now, while it can still be set voluntarily by companies with the credibility to make it stick — before the model scales past the point where anyone can require it.
What Auditable Already Looks Like
At Buchanan #1, a coal mine in southwestern Virginia, four companies operate under four distinct commercial roles in the site's methane abatement work. Coronado Global Resources owns and runs the mine. CNX, through its subsidiary Pocahontas Gas LLC, holds the underlying gas rights. Biothermica supplied the regenerative thermal oxidation units abating ventilation air methane on two shafts. NextEra Energy Marketing is the registered offset project proponent, with credits issued through the American Carbon Registry and California's compliance market.
None of these companies has built a public identity on ethical leadership. Yet the structure is fully legible: publicly filed, cross-referenced across four separate corporate filings and a state regulatory pooling order. When CNX's reported Scope 1 methane jumped from 3,303 tonnes in 2024 to 17,717 tonnes in 2025 — a five-fold increase — the company's own reporting explained why, rather than obscuring it. A coal mine's methane accounting is more auditable, today, than a foundation model company's total carbon footprint. That is not a technical limitation on Anthropic's part. It is a choice not yet made.
The Ask
Publish Scope 1, 2 and 3 figures, including emissions from compute leased through third parties — Colossus and Colossus II specifically. Set a public reduction and offset target with a stated timeline. Let the Frontier purchases sit against a disclosed baseline, rather than a private one nobody outside the company can check. None of this requires Anthropic to build its own data centres, or choose a different infrastructure partner. It requires measuring what it already runs — including, now, the water risk SpaceX itself has put on the record.
Standing Up for the Planet
Anthropic has shown, in public, that it is willing to sit across from a journalist and admit what it doesn't yet know about the consequences of its own technology in war. That is a harder thing to do than most companies manage. The same seriousness, turned toward the physical cost of the infrastructure running that technology, is not a new demand on Anthropic. It's the completion of a claim the company has already made. The ethics that define its identity were never supposed to stop at the server room door.