AI-native clouds are built on borrowed time and money

Neither CoreWeave nor Nscale publishes ESG data but that’s not what you should worry about

Part three of Computing’s sustainability research series examines AI-native clouds CoreWeave and Nscale. Both are young companies; both have a crypto-mining heritage; and both are very shy about publishing sustainability data.

CoreWeave is still less than a decade old, having started up as an Ethereum mining outfit in 2017 in a New Jersey garage. When the crypto crash of 2018 occurred, the company went on a shopping spree for distressed assets and began providing cloud services to organisations running GPU-intensive workloads.

In 2022, CoreWeave invested in Nvidia H100 GPUs and pivoted hard to AI-cloud services. CoreWeave went public in March 2025 with an IPO of $1.5 billion, and earlier this year attracted further investment from Nvidia.

CoreWeave has more GPUs than any other AI-cloud company, bar hyperscalers, but its customers are also hyperscalers – primarily Microsoft but also Oracle – and other AI companies like OpenAI and also Meta.

Nscale shares the crypto mining origin story with CoreWeave but the similarities end there. Nscale is barely two years old, spun out of a company called Arkon Energy which also specialised in crypto and hydro-powered datacentres.

NScale was officially incorporated in May 2024 so has much less history than CoreWeave. We included it in our analysis largely because of its centrality to the UK government’s digital and AI growth strategy and the hype and capital the company has managed to generate.

The initial goal of this exercise was to examine the sustainability data published by both companies, but we had to revise this goal, due to neither company publishing any data.

Summary for IT leaders

What does CoreWeave publish?

CoreWeave published an annual report at the end of 2025 and contained within that report is a small section on sustainability. There are some interesting claims such as:

“Several of the existing data centers in the United States and Europe in which we are a tenant are powered by non-emitting sources of energy”.

Whether “non-emitting” means 100% renewable or nuclear energy is not made clear and no further detail is available in either the report or on CoreWeave’s website. “Non-emitting” seems like a deliberate choice of words. If these datacentres were powered 100% by renewables, you would expect that to be specified.

The report also says:

“We are currently evaluating and measuring Scopes 1 and 2 GHG emissions from our facilities and data center operations and will soon begin to evaluate our emissions resulting from our supply chain (Scope 3).”

So, work in progress. For a company which says it “operates” 43 datacentres across the US and Europe, it seems odd that not even Scope 1 or Scope 2 emissions are reported.

It might be related to the fact that CoreWeave doesn’t own any infrastructure. The S-1 filing of March 2025 says: “We lease all of our data centers and certain equipment under lease agreements.”

CoreWeave owns chips, servers, racks etc. What it doesn’t own is datacentres, power infrastructure or land. As of the end of 2025, it had a massive capex - $14.9 billion and $8.2 billion in operating lease liabilities. Its revenue backlog (coming mainly from Microsoft) stood at a dizzying $66.8 billion.

The revenue backlog is worrying an increasing number of investors and commentators. That, and the debt. CoreWeave's total debt has grown from under $8 billion in 2024 to over $21 billion, with variable rate debt averaging approximately 11% interest. This means, according to DCPulse, that nearly a third of CoreWeave quarterly revenue is consumed by interest payments alone.

A company spending a third of its revenue on debt servicing has challenges, and its bandwidth to invest in building out ESG reporting infrastructure, let alone meaningful sustainability programmes, is minimal.

It’s also troubling that during the IPO process investors don’t appear to have demanded sustainability disclosure as a condition.

Page 2

And Nscale?

Nscale publishes even less than CoreWeave because it’s still a private company, and a young one at that. The only information about sustainability on the company website is given in information about each of the datacentres the company operates – or will operate.

This is a crucial distinction. Nscale’s website currently lists four Nscale operated datacentres. Two of these are in Norway (the hydro-powered ones), one in Loughton in Essex and one in Texas.

Update 02/06 5pm: The following two paragraphs have been updated to reflect further information provided by an Nscale spokesperson since publication earlier today.

An Nscale spokesperson confirmed that the company owns and operates land, datacentres, GPU infrastructure and energy assets in these locations. A further seven co-located, partner-run datacentres are listed across Norway, Iceland, the UK and Portugal with an additional “available” site in West Virginia.

An Nscale spokeperson also clarified that the operational Glomfjord site in Norway is 100% owned by Nscale. The company holds a certified Guarantees of Origin confirming 100% renewable hydropower consumption for this site issued under the Norwegian Guarantees of Origin system.

But the other sites listed on the website as Nscale operated are not all finished and the website doesn’t make this clear. An example is the web page with more information on the Narvik site in Norway. At the time of writing, this simply links to a press release dated July 2025 of the announcement of Stargate Norway – a joint venture between Nscale, Aker and OpenAI.

According to the press release the “facility will target to deliver 100,000 NVIDIA GPUs by the end of 2026.”

OpenAI abandoned plans to rent that capacity in April this year (it also abandoned a UK development and scaled back Stargate Abilene in Texas) and Microsoft is taking up the contract. The date for delivery has slipped and earlier this month Nscale secured another $790 million in financing for the site.

Another extract from the press release still up on the website reads:

“The consortium behind Stargate Norway has made long-term sustainability a core principle of the project. The facility will run entirely on renewable power and is expected to incorporate closed-loop, direct-to-chip liquid cooling to ensure maximum cooling efficiency. Additionally, excess heat from the GPU systems will be made available to support low-carbon enterprises in the region.”

This sounds great, but what happens now the Stargate Norway consortium is no longer the customer? Also reported was the switch to Nvidia’s Vera Rubin chips at these and other Nscale sites. These are far more power intensive chips than their predecessors and have only just gone into full production. Will this affect the power and water needs of the datacentres? Will there be enough renewable energy to power racks of Vera Rubins? What fills the gap if there isn't?

Update 02/06 5pm: After this article was originally published, Nscale told us that the chip upgrade won’t affect power and water consumption. Because whilst Vera Rubin chips draw more power per unit, fewer are needed for greater compute output.

The Narvik site was always planned as a large-scale, high-density AI datacentre drawing on renewable hydropower. The change of customer hasn’t changed the plan to use closed-loop, direct-to-chip liquid cooling which is non-evaporative and does not consume or evaporate local water resources.

Another flagship site in Loughton still appears to be, according to reporting in The Guardian in March this year, a builder’s yard and a pile of scaffolding. The picture on the Nscale website of the Loughton site is an artist’s impression.

An Nscale spokesperson said:

“Site investigation and permit work is under way on the Loughton site. While the schedule was recently updated to accommodate the installation of the latest Vera Rubin 200 technology, we expect the site to be operational in 2027.”

Do AI-native clouds even exist?

It’s not a philosophical question. The absence of sustainability data means that we must ask why it isn’t published. To be clear, neither CoreWeave nor Nscale are obliged to publish sustainability metrics, but their nearest peers in the neocloud category such as Nebius and Crusoe do publish at least some sustainability data.

Nebius and Crusoe are both young companies, so the idea that CoreWeave and Nscale can’t possibly have built any kind of reporting infrastructure doesn’t stand up. It certainly doesn’t in CoreWeave’s case because it has been through the IPO process.

More worrying is the explanation that there is not much to report on because the capacity doesn’t exist yet or, worse, the companies are not being fully transparent about the years long gap between deal announcement and being fully operational.

Page 3

Why does it matter?

The lack of transparency matters for many reasons, the first of which is basic governance. Nscale has raised $3.7 billion, secured contracts with Microsoft, ByteDance and OpenAI, and seems to be the lynchpin of UK AI strategy, whilst remaining vague about some important operational details.

It matters for sustainability too. The CoreWeave and Nscales of this world exist, we are told, because of the explosive growth in the market for GPU clouds. Certainly, the seemingly endless funding rounds and investments from Nvidia which have pumped these companies' valuations (at the time of writing CoreWeave is valued at around $60 billion and Nscale at $14.6 billion) so high would suggest that demand is off the scale.

But the demand (in both cases) appears to be coming almost entirely from a tiny number of hyperscalers and OpenAI.

Having so few customers removes the incentive for CoreWeave and Nscale to bother collecting and reporting this data beyond what they are compelled to do as suppliers. Who else is asking them for it?

And whilst the emissions data of CoreWeave and Nscale will probably be reported within the Scope 3 emissions data of their hyperscale customers, there’s no direct visibility, and there are lots of other sustainability metrics such as water consumption which would not be covered.

Also, if these companies have the reporting infrastructure to be able to report their direct CO2e emissions to their hyperscale tenants why not include the data within their own reporting?

Datacentre delays

Even if we accept that most of the demand for AI compute is coming from about three companies, there’s a problem on the supply side of the ledger. A report from earlier this year found that around half of the promised datacentre capacity for 2026 globally is likely to be delayed, in some cases permanently.

It is a lot harder to build datacentres than anyone in the industry likes to admit. Anti-AI sentiment has crystalised in staunch opposition to datacentre development proposals which currently enjoy approval ratings somewhere between those of Keir Starmer and the black death. The subject has even caught the attention of none other than Erin Brockovich.

CoreWeave is being sued in the US in a class action by investors arguing that the company overstated its ability to meet customer demand pre-IPO and concealed significant construction delays at its datacentres.

Returning to the over reliance on Microsoft, TD Cowen analysts reported in early 2025 that Microsoft scrapped leases for "a couple of hundred megawatts" of US datacentre capacity, suggesting that the hyperscaler might be rectifying an oversupply of AI infrastructure.

The company also paused converting statements of qualifications into formal leases. It was reported later that Microsoft had let more than a gigawatt of agreements on larger sites expire and walked away from multiple deals involving about 100 megawatts each.

Microsoft and OpenAI have also ended their exclusive agreement, meaning that Microsoft no longer needs to build out all the datacentre capacity for OpenAI – an obligation seemingly picked up by Oracle.

If Microsoft materially reduces its CoreWeave commitments, given the company’s acute dependence on Microsoft and very high levels of debt, CoreWeave is in a lot of trouble and could quite feasibly leave a trail of partially built datacentre campuses and significant physical infrastructure with no guaranteed tenant.

The environmental cost of stranded assets

Nscale carries less debt than CoreWeave but as a less mature company has no IPO process to have forced at least some stress-testing of its business model. The company has disclosed debt of more than $2 billion and has raised three very chunky equity rounds. But because it’s a private company, we have no idea how it’s structured its accounting.

The $1.4 billion loan that Nscale announced in February was partly funded by Blue Owl, a private credit provider which is increasingly under scrutiny for risky lending practices, and which was forced earlier this year to block customers withdrawing cash from certain funds.

Nscale’s debt-fuelled growth has been founded on the back of relationships with a tiny number of very large customers, one of which is already showing signs of rethinking how much datacentre capacity it will need in the long term.

Update 02/06 5pm: The following paragraph has been updated to reflect information supplied by Nscale.

Nscale does not always specify how long contracts with anchor tenants are in place. A spokesperson shared a press release stating five-year contract terms for the Narvik site with Microsoft but this is just one site. Five years is only a fraction of a datacentre lifetime.

In an interview with the New York Times earlier this year, Nscale Chief Executive Josh Payne expressed confidence that new tenants would always be found.

Let’s hope he’s right, because the environmental cost of building datacentres that aren’t fully operational - the embodied carbon in steel, concrete and hardware, the land use, the grid connections, the cooling and the energy infrastructure (which are huge infrastructure projects in themselves and increasingly involve building off grid gas turbines) is a sunk cost that no ESG report can ever account for.

Over the last year, the “insatiable” demand for AI compute that justified the entire AI-native cloud buildout has begun to show signs of being satiable. If the usual laws of economics still apply, the sharp increase in LLM costs will reduce demand – we're already seeing companies like Uber and Amazon tell coders to use LLMs with greater care. Tokenmaxxing is out.

But the carbon emissions trapped in underutilised datacentres don't disappear from the atmosphere along with the business case for their existence.

For IT leaders curious about the long-term environmental cost of AI infrastructure, it might be best to swerve questions about sustainability for now and ask yourself instead when the contracted infrastructure will exist and what financial shape the companies building it are likely to be in.

The absence of sustainability data is a problem, but right now the sustainability of the AI-native business model itself looks more pressing.

Both companies named in this article were offered the opportunity to respond prior to publication. CoreWeave did not respond. Nscale provided one comment for attribution, included above, and engaged on a background basis. Information provided during those conversations has informed this reporting both before and after initial publication.