The ‘neocloud’ category has splintered into five companies with radically different playbooks, and buyers who treat them as interchangeable are leaving money on the table. CoreWeave is the heavyweight by revenue, posting $2.575 billion in Q2 2026 and a backlog north of $104 billion, but it also carries a $626 million net loss and a premium price list. Nebius, meanwhile, grew revenue 454% year over year to $582.3 million and is the only provider publishing on-demand B300 pricing — a signal it’s positioning itself as the forward-looking option for buyers planning around NVIDIA’s next architecture.
Lambda and Crusoe remain private, but both are barreling toward public markets. Lambda’s $1.5 billion Series E and Microsoft deal covering GB300 NVL72 systems give it serious supply, and it undercuts everyone on B200 at $6.69 per GPU-hour. Crusoe is the contrarian: it’s the only one with AMD MI300X and MI355X on its rate card, and its 4.9 GW of contracted infrastructure across five campuses suggests it’s building for hyperscaler-scale customers rather than individual developers.
Then there’s Groq, which is arguably not a neocloud at all anymore. After licensing its inference technology to NVIDIA in a deal CNBC pegged around $20 billion, the company raised $650 million in June 2026 and another $350 million in August, joined the NVIDIA Cloud Partner program, and now plans to add NVIDIA GPU capacity alongside its LPU cloud. That’s a pivot from chip designer to infrastructure provider — and a bet that its LPU differentiation can coexist with commodity GPU supply.
The pricing data, read directly from provider pages on August 21, 2026, tells a story of fragmentation. CoreWeave commands Platinum status in SemiAnalysis ClusterMAX 2.0 and charges accordingly. Lambda wins on raw B200 price but offers no spot tier. Crusoe is the only AMD option. Nebius is the only one showing B300 numbers. For buyers, the real question isn’t who’s cheapest — it’s whose contract structure and power pipeline match your workload timeline. CoreWeave’s 4.2 GW+ contracted across 51 data centers and Nebius’s 5 GW target both suggest capacity is being locked up years ahead, which means the spot market may get tighter before it gets better.
