新云厂商 CoreWeave 与 Nebius 的现金储备增速超过营收

CLOUD

The biggest clouds in the United States and China had existing direct businesses selling compute, storage, and networking capacity when the GenAI boom started, and many of them also have ancillary businesses selling advertising or other large scale media or software services to consumers. So it has been relatively easy for them to pivot.

The neoclouds, by contrast, were created more recently, sometimes out of dubious cryptocurrency farms and explicitly to undercut the high prices that the big clouds can command for GPU-accelerated systems that support GenAI training and inference. And ironically, some of the biggest model builders – including the hyperscalers and clouds as well as model builders like Anthropic and OpenAI – are becoming their customers.

It is pretty funny when a big cloud buys capacity on a neocloud to support an AI model builder. But with the demand for compute being so much larger than supply, it all comes down to who has the GPU or XPU allocations and what price they think they can get for access to them.

Few of the neoclouds are public companies, so we don’t get much insight into their inner workings. But CoreWeave, one of the former bitcoiners from New Jersey, and Nebius, a spinout of Russian Internet conglomerate Yandex now based in the Netherlands, are public and are bellwethers of a sort for this segment of the datacenter market. Here is a review of its business from 2022 through 2025, when CoreWeave was doing its pivot to GenAI and when it was utterly dependent on Microsoft for its initial sales:

It is not clear how dependent CoreWeave is currently on Microsoft for its revenues in the second quarter of 2026, but the company has said that it wants no one company to have more than 35 percent of its revenues in the near term and has inked long-term deals with OpenAI ($22.4 billion), Meta Platforms ($32 billion), Google, and Anthropic, which covers the largest model builders in the world. Hedge fund Jane Street, which is perhaps the most aggressive of the Wall Streeters to embrace GenAI, has made a $1 billion strategic investment in CoreWeave, which presumes that Jane Street will be renting time on CoreWeave iron as well.

Drilling down into the quarterly numbers, CoreWeave has amassed a $104.2 billion revenue backlog through the end of the June quarter, and since that time up through this week, it has added another $25 billion to the backlog.

In Q2, CoreWeave’s revenues grew by a factor of 2.1X to $2.58 billion and it reduced its operating loss to $49 million, but the company’s net loss ballooned to $626 million. The company spent $9.4 billion on capital expenses and exited the quarter with $6.41 billion in cash and equivalents to help fund the next round of buildouts.

CoreWeave finished up Q2 with 51 datacenters around the world, and with 1,500 MW of active power installed, we reckon that this is around 1.1 million GPUs across its fleet. CoreWeave has 3,700 MW of contracted power to support its buildouts for the big model builders and Microsoft. Thanks to its tight relationship with Nvidia, which has been fostering the neoclouds to put competitive pressure on the big clouds who are all making their own CPUs and accelerators, CoreWeave is the first company to bring up a “Vera-Rubin” NVL72 system, which it did during Q2 2026.

Nebius is a lot smaller than CoreWeave, but growing fast and building its own cash hoard to build more datacenter capacity.

Nebius has one-tenth the revenue but a little more than one half the capital expenses of CoreWeave, which means Nebius is in its hypergrowth phase to try to catch the 2027 opportunity for GenAI systems.

The company’s deferred revenue is, at just shy of $6 billion, a tiny fraction of what CoreWeave has lined up. Nebius top brass have said they have over $40 billion in customer capacity commitments. This has not been characterized, but it is widely believed that Microsoft and Meta Platforms are the anchor customers at Nebius.

The interesting thing about Nebius is how its contracted power is growing for this year. Back in August 2025, the company said it would have more than 1 GW, and then in November 2025 that was lifted to more than 2.5 GW. In February this year, it was stepped up to more than 3 GW, and then in May it was projected to be more than 4 GW. As the June quarter was ended, the capacity contracted by the end of 2026 was set at 5 GW. This quarter, Nebius started inking short-term contract deals lasting three to six months and sees a “price opportunity in the $40 million to $50 million per MW range,” according to its report to Wall Street. The core of the business is still for contracts that last one to three years.

SPONSORED FEATURE: HPE and NVIDIA help organizations apply accelerated computing, software, enterprise infrastructure, networking, control plane, and services to the workloads and business outcomes that matter most.

October 1, 2026

September 30, 2026

AMD Buys World Labs For $8.2 Billion To Get AI Model Cred