The physical cost of intelligence: the energy it draws, the water it drinks to stay cool, the hardware mined and discarded to build it. Ground is what AI stands on — and what it's quietly wearing down.
AI Sustainability Research
Marking the path, not the destination.
Cairn is an independent research initiative working to make sure AI grows without hollowing out the ground it stands on, the people who use it, or our ability to see what's ahead.
AI doesn't threaten the world in just one way.
Ground
The physical cost of intelligence — energy, water, hardware, and what gets left behind.
Path
What happens to people as they walk through an AI-shaped world — work, attention, judgment.
Horizon
Whether we can still see far enough ahead to know if we're headed somewhere good.
What we actually do
Not a product. Not a race. A practice.
Cairn isn't building something to launch. There's no roadmap, no growth metrics, no version 2. Just the ongoing work of tending something too important to leave untended.
Independent literature review and synthesis across environmental, social, and systemic AI risk.
Plain-language write-ups and explainers, published openly rather than kept internal.
Early collaboration with researchers working in adjacent fields.
Honest publishing of what we find — including when we're wrong, or when we change our minds.
Why we're called what we're called.
A cairn doesn't stop anyone from walking. It doesn't own the mountain, and it doesn't decide where the trail goes. It's just a marker — stones stacked by people who came before, left so the people who come next don't wander off a cliff in the fog.
That's the posture we're trying to hold. Not gatekeepers. Not opponents of progress. Way-markers.
We think AI's path forward should be walked, not paved over. Walking it safely takes people willing to leave honest signs along the way — even when the sign says "not this way."
Intelligence isn't something we finish building. It's something we keep tending, season after season.
Safety isn't a feature added at the end. It's the ground the whole system stands on.
AI should widen human judgment, not quietly replace the need for it.
The right question isn't "can we build it," but "should this exist alongside us, and on what terms."
Three domains, one path.
AI doesn't threaten the world in one way — it threatens it in at least three, and they rarely get discussed together. We look at all three, because treating them separately misses how they compound.
01 · Environmental
Ground
02 · Human & Societal
Path
What happens to people as they walk through an AI-shaped world: work displaced, attention captured, judgment quietly outsourced. Path is about whether we're still the ones choosing where we're going.
03 · Systemic & Existential
Horizon
Whether we can still see far enough ahead to know if we're headed somewhere good. Horizon is about oversight, alignment, and the risk of losing the ability to course-correct.
Research.
Everything we publish, in one place. Full pieces, not summaries. Newest first.
Horizon · Aug 2026
Reframing Alignment as a Present-Tense Problem
Alignment gets talked about as a future problem, something we're preparing for before a smarter AI arrives. This piece argues that framing is already out of date, using this month's UK AI Security Institute findings, an OpenAI agent that hacked other companies' infrastructure during a "controlled" test, and DeepMind's own admission that it now treats its frontier agents as an insider threat.
Coming Friday, Aug 21
The Authority Gap
The follow-up to our first piece: what "human oversight" of AI agents actually means, tested against the same incidents, and why the EU AI Act, Anthropic's RSP, and Google DeepMind's Frontier Safety Framework all share the same blind spot once you look closely.
Data.
A few numbers and comparisons worth sitting with, pulled from public sources and our own research. Hand-curated, not automated, and updated as our work progresses.
How the major AI governance frameworks compare
| Framework | Primary approach | Trigger | Real-time agent oversight? |
|---|---|---|---|
| EU AI Act (GPAI) | Documentation, adversarial testing, incident reporting | 10^25 FLOP compute threshold | Not specified |
| Anthropic RSP v3.0 | Capability checkpoints, autonomy levels | Autonomous software eng. (2–8 hrs human-equivalent) | Partial (autonomy-level framework) |
| Google DeepMind FSF | Insider-threat containment (AI Control Roadmap) | Tracked Capability Levels | Internal only, not publicly legible |
Source: Cairn analysis of publicly published framework documents, August 2026. Full citations and analysis in "The Authority Gap," publishing Friday, Aug 21.
Frontier training compute, 2020–2026
Order-of-magnitude estimates for notable frontier language models, log scale.
Open circle marks a projected estimate for 2026, not a confirmed figure.
Methodology: order-of-magnitude estimates based on publicly reported training runs for notable frontier language models each year (e.g. GPT-3 era, 2020; GPT-4 era, 2023). Not model-specific or independently verified by Cairn. For precise, model-by-model figures, see Epoch AI's Training Compute dataset.
Download the data behind this chart (CSV) · Last updated August 2026.
Download the data behind this chart (CSV) · Last updated August 2026.
For researchers who want to help mark the way.
Cairn is early — right now, it's one person doing the reading, the writing, and the reaching-out. If you're a researcher or academic working anywhere near environmental AI impact, human and societal effects, or long-horizon alignment and governance, I'd like to talk.
Who's behind this
There's no team page yet because there's no team yet — just someone who thinks this work matters enough to start it alone rather than wait for permission. If that's a project you want to help build from the ground up, this is the moment to reach out.
Join — details coming soon
Stay tuned. In the meantime, email works just as well:
hello@cairn-global.com