CoreWeave is built around contracted gigawatts and a short list of take-or-pay customers. Nebius is pairing compute with a software and inference layer while funding growth through prepayments and asset-backed debt. Vultr is expanding geographic reach and dual-silicon options. Those are three different growth metrics and three different risk profiles, and for a CIO building an AI infrastructure plan, that is the practical reason to use more than one of them.

CoreWeave Is Built for the Largest Training Runs a Handful of Customers Can Fund

CoreWeave reported Q2 2026 revenue of $2.6 billion, up 112 percent year over year, and a contracted backlog of $104 billion, up 246 percent (CNBC, 2026). The backlog figure excludes more than $25 billion in new commitments signed in the first weeks of the third quarter (Yahoo Finance, 2026).

Meta and Anthropic are the primary anchors of the large take-or-pay contracts (TechTimes, 2026). Contracted power stood at 4.2 gigawatts by the time of the earnings call, and management raised 2026 capital expenditure guidance to $35 billion to $39 billion (Investing.com, 2026).

The financing is debt-heavy. Net loss was $626 million. Interest expense reached $640 million, more than double the prior-year figure. Adjusted EBITDA margin was 59 percent. Adjusted operating margin, which includes depreciation, was 5 percent (Yahoo Finance, 2026).

CoreWeave is the right choice for frontier-scale training runs, and the customer concentration behind the backlog is tighter than the headline number alone suggests.

Nebius Pairs Compute With a Layer CoreWeave Hasn't Built

Nebius grew second-quarter 2026 revenue 454 percent year over year to $582 million. Adjusted EBITDA turned positive at $236 million, a 41 percent margin, after a loss the year before (Yahoo Finance, 2026). The company raised its year-end contracted power target to 5 gigawatts and expects more than $9 billion in customer prepayments across 2026 (BigGo Finance, 2026).

Cash stood at $8 billion. Nebius closed a $775 million asset-backed facility tied to contracted cash flows rather than unsecured corporate debt, and capital expenditure ran $5.7 billion in the quarter against full-year guidance of $20 billion to $25 billion (BigGo Finance, 2026).

Token Factory, its inference and orchestration platform, along with a European data residency position, gives enterprise teams a managed software layer instead of pure rented capacity.

Vultr Is Growing by Adding Regions, Not Just Chasing Gigawatts

Vultr opened its 33rd global data center region in Milan in May 2026, its ninth in Europe. The network reaches 90 percent of the world's population within 2 to 40 milliseconds (BusinessWire, 2026). Neither CoreWeave nor Nebius matches that geographic spread.

Vultr operated self-funded for a decade before its first outside capital, a $333 million equity round led by LuminArx Capital Management and AMD Ventures in December 2024, followed by a $329 million credit facility in June 2025 (HPCwire, 2025). In December 2025 the company committed more than $1 billion to a single 50-megawatt Ohio cluster built around 24,000 AMD Instinct MI355X chips (Medium, 2026). Separate reporting indicates Vultr is pursuing roughly $1 billion in additional funding (Fierce Network, 2026).

Vultr runs both AMD and Nvidia hardware at scale and positions the dual-supplier model as a sovereignty and cost-control option for enterprises avoiding single-vendor lock-in (SiliconANGLE, 2026).

Vultr's chief marketing officer, Kevin Cochrane, has described the approach as matching infrastructure to a workload's cost and sovereignty requirements in whichever region it runs, rather than concentrating capacity in a handful of massive sites (SiliconANGLE, 2026).

Sovereignty Is Where Vultr's Footprint Pulls Ahead

Vultr's regional footprint doubles as its sovereignty pitch. Data residency and jurisdictional control depend on where compute sits, and 33 regions across six continents gives enterprises more options for keeping data in a specific country than either CoreWeave or Nebius currently offers (BusinessWire, 2026).

The dual-supplier model reinforces that pitch. Running AMD and Nvidia silicon at scale gives regulated industries and governments a hedge against the kind of hardware concentration some treat as its own sovereignty risk (SiliconANGLE, 2026).

CoreWeave's model works against a sovereignty pitch. Its scale runs through a short list of take-or-pay customers and concentrated contracted power, built for frontier training rather than distributed, jurisdiction-specific deployment. Nebius offers a partial answer through European data residency and a managed inference layer in Token Factory, but its primary growth metric remains contracted power and software margin, not geographic dispersion.

For workloads where jurisdictional control and silicon diversification matter more than raw training scale, Vultr is the clearest fit of the three.

Each Growth Metric Carries Its Own Concentration Risk

CoreWeave's revenue depends on a short list of take-or-pay customers renewing at scale. Nebius is building capacity ahead of fully utilized demand, backed by prepayments and a debt facility tied to future contracts. Vultr is smaller on raw compute, and its next expansion stage depends on closing additional outside capital on the expected terms.

Spreading AI workloads across the three converts each company's specific risk into a portfolio problem instead of a single point of failure.

Unknowns and Uncertainties

Whether CoreWeave's concentration around Meta and Anthropic holds through renewal cycles beyond 2026 has not been tested at this scale. Whether Nebius's mix of shorter, higher-priced capacity deals changes its longer-term margin profile is not yet visible in a full year of results. Whether Vultr closes the additional funding reported by The Information is unconfirmed.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by