Is Claude Sustainable? What Anthropic Discloses, What It Doesn’t, and Why It Matters

I use Claude every day. Most of my consulting work runs through it, and the workflow patterns I’ve been writing about, like using Claude Code for sustainability research and cutting per-query energy with skills and caching, are real and useful. None of that changes the question I keep getting from clients, peers, and people I meet at sustainability events.

Is Claude sustainable? Is Anthropic sustainable as a company? Can I credibly advise a client to build an AI workflow on a tool whose maker doesn’t publish an environmental report?

The honest answer takes longer than a yes or a no. Anthropic’s per-query efficiency is genuinely strong, the best of any major lab on the most rigorous independent benchmark I’ve seen. At the same time, Anthropic discloses less environmental information than its peers, and in May 2026 the company made an infrastructure choice that materially worsens its footprint. These two things sit alongside each other. Neither cancels the other. The work, for anyone advising clients on AI adoption, is to hold both in view.

This article is my attempt to lay out what’s actually known, what’s missing, and where I’ve landed personally. It’s not a defence of the tool I use, and it’s not a takedown of the lab I work with. It’s the kind of read I wish had been available when I was first asked the question.

Where Anthropic Actually Sits on Disclosure

The first thing a sustainability advisor wants to do, faced with a vendor question, is compare disclosure. Not because disclosure equals performance, but because disclosure is the precondition for any honest assessment. A company that publishes verified Scope 1, 2, and 3 emissions, an infrastructure mix breakdown, and a third-party-assured climate transition plan is signing up for scrutiny. A company that publishes none of those is asking to be trusted.

Here’s how the five companies most relevant to AI workloads compare, based on what each has published as of May 2026.

The pattern is uncomfortable to look at if you use Claude. Google and Microsoft, the hyperscalers Anthropic and OpenAI depend on, both publish detailed environmental reports with verified emissions data, named climate targets, and third-party assurance. The two leading AI labs, on the other hand, publish almost nothing on a standalone basis. They rely on their cloud partners’ disclosures to imply a footprint, and that implication is partial at best.

It’s worth being honest about where this disclosure gap comes from, because the framing matters for what to do about it. Google and Microsoft are publicly listed companies. They sit under shareholder pressure, EU CSRD reporting requirements for their European operations, and a wave of state-level rules in the US. Anthropic, OpenAI, and xAI are still privately held. Most of the standard mandatory climate disclosure obligations haven’t bitten them yet.

“Yet” is the operative word. California’s SB 253 and SB 261 already apply to private companies. Any US-based company with over $1 billion in annual revenue doing business in California has to report Scope 1 and 2 emissions starting in 2026, with Scope 3 from 2027. Anthropic’s revenue, by its own statements and recent reporting, comfortably clears that threshold, as does OpenAI’s. Both companies are headquartered in California. The legal obligation is months away, not years.

Which means the current silence is mostly a choice about timing, not a question of capacity. Anthropic could publish a verified environmental report today and front-run the obligation. It hasn’t. The federal SEC climate disclosure rule that would have forced US-listed companies to report consistently was terminated by the SEC in 2025 after the change in administration, which weakens the comparison to public peers somewhat. But CSRD applies to anyone with significant EU operations, California’s rules apply to anyone doing business in the state at the revenue thresholds, and 35 other countries are developing similar regimes. The regulatory floor is rising, slowly but steadily, and the gap between voluntary and mandatory disclosure will close over the next two years.

Even the Leaders Aren’t Disclosing Enough

Holding up Google and Microsoft as the disclosure standard is fair as a relative comparison. It’s also incomplete. Academic and advocacy research has flagged that even the hyperscalers’ reporting has significant gaps once you look closely, particularly on the AI-specific environmental impact.

A February 2026 analysis by Beyond Fossil Fuels and Stand.earth reviewed Big Tech’s claims about AI’s climate benefits and found that 74% of them were unproven. The reviewers couldn’t identify a single example where consumer-facing generative AI products like ChatGPT, Gemini, or Copilot were producing material, verifiable, substantial emissions reductions in line with what was being claimed. The narrative of “AI as climate solution” turns out to be largely aspirational at this stage.

Specific reporting gaps include:

• Methodology shifts that obscure trends. Microsoft now estimates water withdrawals using water-use-efficiency metrics rather than direct measurement, which makes year-on-year comparisons harder to interpret cleanly.
• Aggregation that hides growth. Google’s water consumption rose 28% in a single year, with 28% of total withdrawals coming from regions with medium or high water stress. The headline number is easy to miss in a 70-page report.
• Blurred categories. Sustainability reports often bundle generative AI’s footprint with classical AI applications, which makes the genuinely energy-intensive part of the business look smaller in the aggregate.
• Missing model-level data. The 2025 Patterns analysis of thirteen AI companies (including Google’s AI divisions) found that ten disclosed none of the key model-specific environmental metrics. Energy, carbon, or water on a per-model basis, the level a procurement question actually needs, is still absent.

The point isn’t to dismiss Google and Microsoft’s reporting. The methodology is real, the targets are credible, the third-party assurance is meaningful, and the trajectory is broadly in the right direction. The point is that “they disclose, the AI labs don’t” is the floor of the argument, not the ceiling. What the field actually needs is per-model, third-party-assured, standardised AI energy and water reporting. Nothing currently in production at any major lab meets that bar.

The Two Pictures of AI Sustainability

There are two ways to measure the environmental cost of an AI tool, and they tell different stories about Anthropic. Both stories are true. Most of the public conversation conflates them, which is why the question of whether Claude is sustainable feels harder to answer than it should be.

The first picture is per-query efficiency. How much energy, water, and carbon does a single inference cost? This is what the ml.energy benchmark measures. The benchmark integrates API performance, infrastructure carbon multipliers, and statistical inference to assess real-world environmental cost across thirty-plus models. On this benchmark, Claude 3.7 Sonnet scored 0.886, the highest in the field. Claude 3.7 with extended thinking consumes about 17.045 Wh for long-form input, less than half the energy of OpenAI’s o3 model.

That’s a real and durable advantage. Anthropic has prioritised inference efficiency, and the choice shows up in independent measurement. For a consultant who runs hundreds of Claude calls a week, the per-query story matters.

The second picture is infrastructure-level reality. Where does the inference physically run, what powers that facility, and what’s the environmental record of that infrastructure? This is the story that ml.energy can’t fully capture, because the benchmark uses aggregate cloud carbon multipliers rather than facility-specific data, and because new infrastructure decisions can change the picture faster than any benchmark can refresh.

On this second picture, Anthropic’s story in May 2026 got materially worse. To understand why, you have to look at the Colossus 1 deal in detail.

What Changed in May 2026: The Colossus 1 Deal

On 6 May 2026, Anthropic and SpaceX announced a compute partnership giving Anthropic access to all of xAI’s Colossus 1 data centre capacity in Memphis. The deal covers more than 300 megawatts of compute, distributed across more than 220,000 Nvidia GPUs including H100, H200, and GB200 accelerators. Anthropic explicitly stated that the new capacity will “directly improve capacity for Claude Pro and Claude Max subscribers.” The lease is projected to generate $5 to $6 billion in annual revenue for SpaceX/xAI.

This matters environmentally because of where Colossus 1 sits and how it’s powered. Before the Anthropic deal, the facility was best known for one of the most contested environmental records in current US AI infrastructure.

According to the Southern Environmental Law Center, when Colossus 1 began operations in mid-2024 it ran on as many as 35 methane gas turbines that operated without Clean Air Act permits, classified as “temporary” installations via a regulatory loophole. After the SELC filed a notice of intent to sue on behalf of the NAACP, xAI removed all but 15 of those turbines and obtained permits for the remaining units from the Shelby County Health Department. The site now receives 150 MW from MLGW (Memphis Light, Gas and Water), with the remainder from on-site gas generation. Given Colossus 1’s proximity to a 1.1 GW Allen Combined Cycle plant, the facility is, in the Southern Alliance for Clean Energy’s words, “powered almost exclusively by combined cycle natural gas.”

That’s the facility Anthropic is now occupying in full.

The picture gets worse if you look at xAI’s adjacent facility. Colossus 2, in Southaven, Mississippi, currently operates 27 methane gas turbines (with 46 portable units on the site overall) generating up to 495 MW of power. None of them have Clean Air Act permits. The NAACP, SELC, and Earthjustice are actively suing xAI for Clean Air Act violations, alleging the facility has the potential to emit more than 1,700 tons of nitrogen oxides per year, which would make it the largest single industrial source of smog-forming pollution in the greater Memphis area. The area already fails to meet national air quality standards. Inside Climate News and local advocacy groups have linked the cumulative facility footprint to elevated pollution exposure in predominantly Black neighbourhoods of South Memphis.

To be precise about what Anthropic is and isn’t responsible for here, Colossus 2 is xAI’s own facility, not Anthropic’s. The unpermitted turbines at Colossus 1 were removed before Anthropic moved in. Anthropic is not directly operating illegal pollution.

What Anthropic chose to do, with full visibility into the public record, is scale a substantial slice of Claude’s inference onto a Memphis-area natural gas infrastructure with a documented permit-violation history, owned by a company being actively sued for ongoing illegal emissions at the adjacent site. The announcement made no reference to the facility’s environmental record, no commitment to renewable PPAs, no carbon accounting for the new capacity, and no climate impact assessment. Simon Willison, in his notes on the deal, captures the awkwardness by quoting Andy Masley, a data centre defender who generally rejects environmental criticism of AI compute, saying he himself “would simply not run my computing out of this specific data center.”

Why Per-Query Efficiency Is the Small Story

When a sustainability advisor or a procurement lead asks whether Claude is sustainable, the per-query story is what they’re usually shown. Anthropic’s ml.energy result. A microwave-for-five-seconds analogy. Some chart comparing inference energy to ChatGPT or Gemini.

That framing makes Anthropic look good, and the underlying numbers aren’t wrong. They’re just not the part of the equation that scales.

The total environmental footprint of an AI vendor is roughly the product of four things. Per-query energy. Volume of inference. Infrastructure mix (grid carbon intensity, water source, PPAs in place). Growth rate of compute capacity. Improving one factor while another grows can leave total impact flat or rising. Microsoft is the clearest published example. Microsoft’s 2025 Environmental Sustainability Report documents a 29.9% reduction in Scope 1 and Scope 2 emissions from a 2020 baseline, driven by 34 GW of contracted carbon-free electricity. Despite that, total emissions across Scope 1, 2, and 3 rose 23.4% over the same period, because the build-out of AI data centres added Scope 3 emissions faster than operational improvements could offset.

Google shows a similar pattern. Google’s 2025 Environmental Report records a 12% year-on-year reduction in data centre emissions despite a 27% increase in electricity consumption from AI. The data centre footprint per unit of compute genuinely improved. Total company emissions still rose 11%, because Scope 3 grew 22%, driven again by AI-related infrastructure.

The lesson for vendor selection is that per-query benchmarks are necessary but not sufficient. A model that’s 30% more efficient per inference, scaled into a 300 MW gas-powered facility that didn’t previously exist, produces more absolute emissions than the same workload running on an older facility powered by a wind PPA. The math doesn’t bend that way for any plausible efficiency gain.

Per-query efficiency is the part of AI sustainability that’s easy to measure and easy to optimise. It’s also where the smallest share of the total environmental harm sits. The infrastructure choices, the growth rate, and the disclosure quality matter more.

The Position I’ve Landed On

I’m going to write this section in first person because the question of whether to keep using Claude, given everything above, isn’t an abstract one for me. It’s a daily working decision.

I’ve landed on the following. I keep using Claude. The reasoning quality is real, the per-query efficiency is genuinely better than the alternatives I’ve tested, and the workflow patterns I’ve been writing about produce better consulting outputs than I could without them. I think the case for using AI well in sustainability work is stronger than the case for refusing to use it on environmental grounds, given the marginal energy cost relative to the alternatives most consultants substitute for AI in practice.

And I think Anthropic’s environmental position right now is the worst among the major AI labs by disclosure standards, and that the Colossus 1 deal makes it worse. I don’t think those two beliefs cancel. They sit alongside each other, and the tension is real.

What I do about it, in practice:

• Use less. The fastest way to reduce the environmental cost of a Claude session is to have a shorter session. Tighter prompts, clearer goals, fewer rounds of revision when the first output is good enough. This sounds obvious. It’s also the most-ignored lever.
• Batch and cache. The patterns from the energy article still apply. Prompt caching, skills, sub-agents, and programmatic image generation cut compute by roughly half on sustained work. Per-query efficiency at the workflow level is something I control, and I should.
• Choose smaller models for routine tasks. Haiku for triage and routing, Sonnet for synthesis, Opus only when reasoning genuinely needs it. This is the cleanest model-selection discipline and the one most consultants skip.
• Push for vendor disclosure in client conversations. When I’m advising on an AI tool selection, the disclosure question goes on the procurement checklist. Not as a deal-breaker, as a data point. If Anthropic wants to be the preferred vendor for sustainability work, the price of that should be disclosure.
• Stay willing to switch. If a peer lab publishes a verified environmental report, names a credible climate target, and matches Anthropic’s quality, I would change my default. The switching cost is non-zero but it’s not infinite, and the comparison should be live, not foreclosed.

For consulting clients in sustainability-adjacent sectors, I’d say roughly the same thing with one addition. The vendor disclosure gap is a risk worth naming explicitly in any AI procurement memo. Not because Anthropic is uniquely bad, but because the gap is real, and pretending otherwise sets up an awkward conversation when the question eventually comes from procurement, sustainability, or a board sub-committee.

Anthropic Is Transparent About Almost Everything Else

There’s a specific awkwardness to Anthropic’s environmental silence that’s worth naming directly. Anthropic positions itself as the safety-focused, responsibility-led AI lab. It’s the company that publishes more detailed material on the risks, capabilities, and societal impact of its models than any of its peers. Against that backdrop, the absence of environmental disclosure looks less like an oversight and more like a category they have decided not to cover.

A non-exhaustive list of what Anthropic does publish:

• The Anthropic Economic Index, now in its third edition, analysing millions of anonymised Claude conversations to map how AI is being used across occupations and tasks. This is original research, made publicly available with the underlying dataset on Hugging Face. I’ve spent enough time with the data to say this is real, useful work.
• The Responsible Scaling Policy, now at version 3, documenting how Anthropic assesses catastrophic-risk capabilities of new models before release. This includes the AI Safety Level framework and the evaluation methodology used for each model release.
• Detailed system cards with each new model, covering capability evaluations, safety testing, and known limitations. Claude Opus 4.5 launched under ASL-3, and the reasoning for that decision is public.
• A compliance framework for California’s SB 53 (the Transparency in Frontier AI Act), published proactively to show how Anthropic plans to meet new state-level frontier AI safety reporting obligations.
• A voluntary commitments hub covering responsible deployment, security and privacy, and other transparency areas.

That’s a lot of disclosure. It’s specifically the disclosure that supports Anthropic’s brand position as a safety-first, transparency-led AI lab. The environmental gap stands out against the rest of the picture because it’s the only major category where the same company has chosen not to publish.

This matters for two reasons. First, capability-wise, a company that runs an Economic Index research program and a Responsible Scaling evaluation pipeline clearly has the internal data-gathering apparatus to produce an environmental report. The infrastructure is there. The decision not to publish is strategic, not logistical.

Second, and more uncomfortably for anyone who buys the “ethical AI” framing, the disclosure gap sits in tension with the brand. A company whose market position is built on responsibility is being asked, fairly, why responsibility apparently stops at the boundaries of environmental impact. The honest answer is probably some combination of “we haven’t gotten to it yet” and “it’s a less favourable story than safety alignment.” Neither is unreasonable. But neither is publishable, which may be why the question doesn’t get answered.

For context, the AI Energy Score initiative is worth noting. Hugging Face, Salesforce, Cohere, and Carnegie Mellon launched it in February 2025 as a standardised framework for benchmarking inference energy across both open-source and proprietary models. Proprietary labs can submit their models via a containerised, privacy-preserving submission process that publishes only aggregate metrics. Anthropic and OpenAI have not appeared on the leaderboard as of May 2026. They could, on terms that would protect their internal infrastructure detail. The barrier is willingness, not infrastructure.

Five Questions to Ask Any AI Vendor

Most enterprise AI procurement processes don’t yet have a sustainability section. The categories of question they need are easy to write down, harder to find good answers to. Here’s the short version I use, which I’d encourage any sustainability advisor or procurement lead to adapt.

If you ask these five (or six) questions of the major AI labs today, the answers are mostly “no” or “not yet.” That’s a useful baseline. The point isn’t to use the answers to reject vendors. It’s to make the gap legible to the client, document the trade-off, and put pressure on the vendor to close it over time.

What Better Disclosure Would Look Like

It’s easy to criticise disclosure gaps without naming what good disclosure would actually contain. For AI labs specifically, the bar I’d hold them to has five elements.

• Verified Scope 1, 2, and 3 emissions on an annual basis, with third-party assurance, against a recognised framework (GHG Protocol at minimum, ideally SBTi-aligned).
• Per-model environmental data for both inference and training. Energy, water, and carbon, on a per-1,000-tokens or per-image basis, refreshed when models change. This is the gap independent academic work has been flagging since at least 2023.
• Infrastructure mix disclosure. Which clouds, which facilities, what grid carbon intensity in each, what fraction is covered by direct PPAs versus RECs versus residual grid mix. The aggregation hides too much.
• A time-bound climate target at the company level, covering operational and value-chain emissions, with an explicit infrastructure transition plan as compute capacity scales.
• Material event disclosure. When the vendor enters a compute partnership that materially changes the environmental footprint, that change should be disclosed alongside the announcement, not left for journalists and law-centre press releases to surface.

None of this is technically difficult. Google and Microsoft already publish most of it. The question for Anthropic, OpenAI, and the rest of the AI lab category is whether sustained pressure from enterprise buyers, regulators, and the sustainability advisory community will be enough to make disclosure the norm. I think it will, eventually. The question is how quickly, and whether it happens before or after the next round of high-emissions infrastructure decisions.

Where This Is Heading

The conversation about AI sustainability is shifting. For the past two years, most of the public attention has been on per-query energy, often via the misleading framing of “one ChatGPT query equals X smartphone charges.” That framing is useful for getting the topic onto front pages, but it’s not where most of the actual environmental harm sits. The harm sits at the infrastructure level. In facility decisions, in PPA quality, in growth rates that outpace clean energy procurement, in regulatory enforcement that hasn’t caught up with the build-out.

That’s where vendor scrutiny is moving, and it’s where AI labs are most exposed. Per-query efficiency is a story Anthropic can tell well. Infrastructure disclosure is a story Anthropic, currently, cannot tell at all.

For sustainability consultants advising clients on AI adoption, this shift is the one to prepare for. The next twelve to twenty-four months will be when “how efficient is your model?” gets replaced by “where does the inference physically run, what powers it, and what does your annual environmental report look like?” The vendors who can answer those questions credibly will have a real procurement advantage. The vendors who can’t will be increasingly visible.

For Anthropic specifically, the path forward isn’t complicated. Publish a Scope 1, 2, and 3 environmental report. Set a credible climate target. Disclose infrastructure mix. Be transparent about the Colossus 1 footprint and what’s planned to offset it. The technical capacity to do this exists. The internal will to do it is the open question.

In the meantime, the honest answer to “is Claude sustainable?” is that it depends on what you mean. Per-query, it’s better than most. As a vendor, with the current disclosure gap and the current infrastructure choices, it isn’t. The work for those of us who use it is to hold both pictures at once and to keep asking for the disclosure that would let us evaluate it more cleanly.

If you found this useful, the related pieces on Claude Code for sustainability research workflows and cutting AI energy use through smarter workflow patterns sit alongside this one. Together they’re the closest thing I have to an honest position on how to work with AI in sustainability consulting without pretending the trade-offs don’t exist.