The Neoclouds

A neocloud owns accelerators and sells time on them. It is not a general cloud with an AI product bolted on; the accelerators are the whole asset base. That makes these firms the marginal buyer for most of what the rest of this index describes — racks, cold plates, substrates, power contracts — and it makes their published accounts a useful independent read on how much AI demand is contracted rather than merely anticipated.

Two of them file. For the quarter ended 30 June 2026, CoreWeave reported revenue of $2,575 million, up 112.4% year on year; Nebius reported $582.3 million from continuing operations, stated as 454% growth with its AI cloud segment up 514%.

The distinctive thing in both accounts is a large operating margin and a large loss at the same time. CoreWeave's adjusted EBITDA was $1,510 million — a 59% margin — alongside an operating loss of $(49) million and a net loss of $(626) million. Interest expense that quarter was $(640) million: the interest bill alone was more than thirteen times the operating loss, and slightly larger than the net loss itself. Nebius shows the same shape smaller, with interest expense up from $4.8 million to $119.1 million as three rounds of convertible notes began to cost. The operations generate cash at a high margin; the capital structure required to own the assets consumes it.

The building is easier to grasp against revenue. In the first half of 2026 CoreWeave spent $14,117 million of capital expenditure while recognising $4,653 million of revenue — about three times as much invested as earned. Nebius spent $8,130.3 million against $981.3 million, more than eight times. That is not growth funded from operations. It is capital-markets access converted into installed capacity as quickly as possible: Nebius raised $4,337.5 million of convertible notes, $2,846.7 million from treasury share sales and $2,000.0 million from prefunded warrants in six months.

Backlog is the number the sector is argued about, and it needs reading rather than quoting. CoreWeave's approximately $104 billion is defined in its own release as remaining performance obligations *plus* other amounts management estimates will be recognised under committed contracts. The first half is an accounting measure with rules; the second is an estimate. Treating the total as a contract value overstates what has been established.

Capacity is increasingly stated in watts rather than servers. CoreWeave reported active power up about 500 MW to 1.5 GW against approximately 3.7 GW contracted — meaning roughly 2.2 GW secured but not yet energised. That is a claim on grid connections, transformers and turbines whose lead times run to years, which is why it connects directly to the electricity collections here. It is also a schedule rather than an asset.

One caution generalises beyond this page. The metrics neoclouds are discussed in — run-rate revenue, contracted gigawatts, named customers, concentration, guidance — are largely absent from the documents that carry legal weight. Nebius's Form 6-K exhibit contains none of them. CoreWeave's release says outright that guidance would come on the call instead. The most-quoted numbers are spoken; the filed ones are the ones nobody repeats.

And the largest analytical hole is stated plainly: nothing read here establishes what useful life these accelerators are assumed to have. Whether one is depreciated over six years or three changes the reported economics of every company in this category.