The sustainability mirage at the heart of neocloud

Three neoclouds, three sustainability stories yet only one stands up

Article three in our AI cloud sustainability research series examines the neocloud challengers which present a more values-driven message than their hyperscale competitors. Do the sustainability credentials match the ambition?

Neoclouds are distinct from the hyperscale giants because they are not merely selling scale, speed, or geographic reach. Each brings to market a different vision about what cloud infrastructure should look like in the AI era.

Where hyperscalers increasingly resemble utilities, neoclouds market themselves as purpose-driven alternatives. Sustainability, sovereignty, energy innovation, or infrastructure efficiency are presented not as secondary concerns but as defining characteristics of the business model itself.

That distinction matters because sustainability claims at hyperscale level often function as compliance exercises. Neoclouds, by contrast, are still constructing both their infrastructure and their identities. Their sustainability narratives sit much closer to the core of their commercial proposition.

Why these neoclouds?

Headquartered in France, OVHcloud was top of the Computing cloud ranking for sustainability in 2025 because its ‘frugal by design’ vertical integration and hardware ownership model produces more sustainable infrastructure than the hyperscale alternative.

In addition to its sustainability narrative, OVHCloud is increasingly framing a message that it offers greater opportunity for sovereign AI than hyperscale competitors. Geopolitical turmoil and cyber risk analysis has led many IT leaders in the UK to question long held dependence on US owned and/or located infrastructure so are more receptive to sovereignty-focused messaging.

CrusoeAI is US-based and leads with a message that the demand for AI compute is booming and therefore a sustainable energy model is needed. Crusoe has developed such a model which uses the methane by-product of il extraction to power compute infrastructure in a way that (it claims) reduces emissions. That model – and the fact that Crusoe is developing the Stargate Abilene site for Oracle is why Crusoe earns a place in the analysis.

Amsterdam-based Nebius leads with sustainability messaging too but is also increasingly articulating an offering on data sovereignty, pitching itself as an alternative to US hyperscaler dependency. It’s an interesting angle given that it announced a five-year contract last September with Microsoft worth $17.4 billion to provide "access to dedicated GPU infrastructure capacity.” This deal, and the vast quantity of capital Nebius has attracted despite having only existed in its present form for two years is why we chose it.

Summary for IT leaders

• OVHcloud is the only neocloud in this analysis whose sustainability claims consistently withstand scrutiny. Its reporting is mature, transparent and supported by operational decisions designed to reduce resource consumption rather than simply offset emissions.
• Crusoe’s energy model is innovative, but its carbon accounting is highly questionable. The company makes credible arguments that its methane capture technology is preferable to flaring but stretches credibility by claiming “net negative” emissions through avoided-emissions calculations built on multiple hypothetical assumptions.
• Nebius’s sustainability claims are unsubstantiated. Reported emissions are extraordinarily low relative to the scale of infrastructure expansion being announced, while key metrics such as Scope 3 emissions and water consumption are absent entirely.
• Nebius also has a much smaller physical footprint than its marketing, media coverage or ability to raise money suggests.
• Transparency is the key dividing line. OVHcloud provides detailed operational data and explanations for changes in emissions performance. Crusoe and Nebius rely far more heavily on selective disclosure, framing and vibes.

Comparative issues

There is a substantial gap in ESG reporting maturity between the three companies.

OVHcloud, founded more than 25 years ago, provides detailed reporting across scope 1, 2, and 3 emissions, carbon intensity, water withdrawal, and waste management. Importantly, the company also tends to respond constructively when clarification is requested.

Crusoe’s transparency is much more uneven. Large sections of its reporting focus on “avoided emissions” generated through its Digital Flare Mitigation ® (DFM) technology. This introduces major methodological questions because avoided emissions depend on assumptions about what would otherwise have happened.

As the hyperscalers tend to do, Crusoe also frequently substitutes narrative for measurable disclosure. Its reporting contains extensive descriptions of cooling technology, procurement optimisation, and sustainability ambition, but often omits the underlying operational metrics needed to evaluate those claims.

Nebius provides even less visibility.

Scope 3 emissions are not reported at all. Water consumption is absent. Reported Scope 1 and 2 emissions are extremely low despite the company simultaneously announcing substantial GPU infrastructure expansion across Finland, France, and a commencement of a build-out in Kansas City.

It is very difficult to reconcile such low emissions and zero waste to landfill with this scaling up of capacity. Which does of course suggest that either the data is wrong, or the scaling up was more a set of announcements as opposed to the genuine addition of capacity. Either way, transparency is lacking.

Carbon emissions

OVHcloud performs best in this category because its reporting is both comprehensive and internally consistent. Emissions increased during 2024 as you can see:

However, OVHcloud accounted for increases in emissions, explaining that the increase in Scope 3 emissions was largely due to the purchase of more hardware.

OVHcloud is also transparent when it acknowledges that a recorded reduction in its Scope 2 location-based emissions despite an increase in fuel consumption, was due to progress in further decarbonising the French electricity grid. Its market-based emissions rose because the company doesn’t have access to renewable sources of electricity in every new region.

In terms of carbon intensity, OVHcloud provides a metric of 0.13kg CO2e per kilowatt hour.

CrusoeAI’s reported carbon data is set out below. The increase in Scope 1 emissions is explained by way of DFM expansion into new sites. These should reduce when the reporting period catches up with Crusoe’s divestment of its Bitcoin mining operation.

Low Scope 2 emissions are to be expected from a company that generates most of the energy for its operations. The company uses third-party datacentres for its Crusoe Cloud business, and these datacentres use a mix of PPAs and EACs which are controversial as we’ve explained before on these pages because they may or may not reflect units of additional renewable energy generation. Nonetheless, Crusoe deserves some credit for reporting these emissions within their own Scope 2 data because they don’t have to.

Crusoe limits its reporting of Scope 3 emissions data. Only 6 categories of emission are reported, and no distinction is made between location and market data. Growth came mainly from growth in the Fuel and Energy Related Activities (FERA) Scope 3 category because of the amount of gas the company burns.

Creative carbon accounting

Crusoe has invented a new emissions category for those it claims to have avoided – Scope 4. The company argues that by avoiding a total of 1,304,000 mtCO2e it makes it a net negative emitter to the tune of -550,900 mtCO2e.

Let's interrogate that claim.

The first part of the calculation is probably the most defensible. If the methane gas burned by the Crusoe generators were either flared or vented, more CO2e would be emitted. You still have to make some assumptions about the likelihood of oil companies to flare/vent excess methane and avoid questioning whether maybe in some parallel universe some more efficient ways of dealing with the by-products of fossil fuel extraction haven’t been invented but, if you squint a bit, this stacks up.

Where Crusoe’s carbon accounting starts to look really iffy is in the claim that the emissions that were avoided by not flaring the methane also displace emissions that would otherwise have been generated.

This claim stacks multiple counterfactual assumptions against the same unit of methane, chiefly one that the unit of energy Crusoe has produced by burning methane has also displaced the need for a unit of energy that would have been generated and consumed in a way that produces more carbon. This is unprovable.

Crusoe’s activity as a company - building AI cloud datacentres - creates a carbon footprint. It can argue that the way it produces its energy reduces that footprint, relative to what it would be if it drew energy from the grid.

But to argue that its existence as a company not only removes its own footprint but also part or all someone else’s footprint seems a bit of a reach.

Nebius presents a different problem

Nebius’s carbon emissions look strangely low for a company which attracted $700 million investment in the same year to fuel the ‘insatiable demand’ for AI compute.

• Scope 1: 98 mtCO2e
• Scope 2 LBM 2525 mtCO2e
• Scope 3 MBM 1938 mtCO2e

A tiny footnote in the 2024 Impact report reads as follows: “As of the end of 2024, two facilities — Finland and France — were operational, and are therefore in the scope of the current report.”

This neocloud is a whole lot smaller than it looks. Nonetheless, even if Nebius only had two datacentres operational at the time, and the company provides a PUE of 1.3 (bear in mind Google average PUE of 1.08 so these two datacentres aren’t as efficient as the text of the Nebius Impact report implies) those emissions look (and I mean this literally) unbelievably low.

Also, Nebius does not report Scope 3 emissions. Unlike Crusoe, Nebius doesn’t seem to report emissions from third-party datacentres within its own Scope 2 data, so any emissions generated up or downstream of third-party datacentres are kept helpfully off the books.

Nebius’s carbon intensity metrics also look strange. Tonnage of CO2e generated per MWh is given at 0.05. Bear in mind that this number is partly a function of PUE which measures how much of a datacentre’s overall power use is consumed by the compute as opposed to cooling, networking, lighting etc. so the closer to one it is, the more efficient with power that datacentre is.

Google has an average PUE of 1.08 and a carbon intensity per MWh of 0.0957. But Nebius reports a much less efficient PUE of 1.3 but a lower carbon intensity of 0.05.

It makes no sense.

The non-reporting of Scope 3 emissions also hides the indirect emissions from the work which had commenced in Kansas on the site of a former print factory. The first phase of that development became operational in April 2025, which must have meant some serious construction work took place in the last quarter of 2024. We can’t see any Scope 3 emissions associated with this build-out, and it seems unlikely to the point of impossibility that the Scope 1 and 2 emissions generated from it are also accounted for.

The most generous interpretation of these inconsistencies relates to the years long gap between announcing datacentres and them becoming fully operational. Maybe, like lots of the datacentres that Nebius announced between 2024 and now, missing emissions data will appear in the future.

Water & waste

OVHcloud is the only company of the three which reports comprehensive water withdrawal data. It’s total water consumption of 125.7ML remains a fraction of those of those of the hyperscale operators.

Crusoe doesn’t report its waste volumes and did not respond to a request for the information. OVH Cloud generated total waste volumes of 845 tonnes, and 12% of that went to landfill.

Nebius reported waste volumes of 731 tonnes of waste but states that none went to landfill. This sounds unlikely but does fit the profile of a business which is very good at announcing datacentres, but less so at building them.

Marketing meets reality

The neocloud tier promised a more sustainable and values-driven alternative to the scale-at-any-cost hyperscaler model.

OVHCloud largely delivers on that promise. Its reporting is mature, its methodology is transparent, its explanations of year-on-year changes are credible, and its water consumption looks genuinely modest against hyperscale comparators. Whilst emissions increased in 2024, the increases are explained, contextualised and traceable. This is what good ESG reporting looks like.

Crusoe is more complicated. Its DFM technology allows it to make a defensible argument that burning methane at high efficiency is better than the alternative. But the leap from "we reduce emissions relative to flaring" to "we are net negative emitters" relies on a chain of counterfactual assumptions that does not survive scrutiny. This is not sustainability leadership. This is creative carbon accounting at its most cynical.

Nebius is the most difficult to assess, not because the picture is complicated, but because there is so little picture to assess. By its own admission, just two facilities were operational at the end of 2024. Yet even for a company of that scale, emissions figures that are almost non-existent whilst hundreds of millions of dollars are (presumably) being deployed just doesn’t seem to reflect a fully reported operation. Carbon intensity figures that outperform Google despite a significantly worse PUE suggest that what is being reported is not the complete picture.

In summary, OVHCloud has earned its sustainability credentials through two decades of considered infrastructure decisions. Crusoe has a genuinely interesting energy thesis but is undermined by its carbon accounting and selective reporting.

Nebius has operationalised a great sustainability narrative - but very few datacentres.