2026 neocloud GPU providers ranked by pricing and contracted power

The article ranks five neocloud GPU providers—CoreWeave, Nebius, Lambda, Crusoe, and Groq—based on published pricing, contracted power, and other key metrics. CoreWeave, a public company listed on Nasdaq, reported Q2 2026 revenue of $2.575 billion, up 112% year over year, with a revenue backlog of $104 billion. Its contracted power reached 4.2 GW+ across 51 data centers, and it completed the industry’s first bring-up and validation of NVIDIA’s Vera Rubin NVL72. Nebius, also a public company, reported Q2 2026 revenue of $582.3 million, up 454% year over year, with an AI cloud segment EBITDA margin of 49.7%. The company raised its 2026 contracted power target to 5 GW and plans to deploy over 1 GW annually starting in 2027. Lambda, a private company, closed a $1.5 billion Series E in November 2025 and a $1 billion credit facility in May 2026. It has a multi-year agreement with Microsoft covering tens of thousands of NVIDIA GPUs. Crusoe, another private company, raised $1.375 billion in October 2025 and is building two large data centers for Oracle, OpenAI, and Microsoft, with contracted power reaching 4.9 GW. Groq, which launched GroqCloud in 2024, licensed its technology to NVIDIA in December 2025 and now focuses on inference-cloud infrastructure, operating 13 data centers and planning to scale to 200+ MW in 2027. Pricing varies significantly, with Lambda offering the cheapest B200 instance at $6.69, while CoreWeave charges a premium with H100 rates of $6.16 per GPU-hour. Nebius offers the lowest B300 rate at $7.85, and Crusoe is the only AMD option with H200 priced at $4.29. Groq’s LPU is sold per token, not per GPU-hour. The article also highlights the importance of contracted power, with Nebius disclosing that its Q2 2026 deals averaged over $20 million per MW, and its first capacity auction. CoreWeave and Nebius are the only public providers with demonstrated NVL72 rack-scale experience, while Lambda and Crusoe offer more affordable options for mid-size training and budget-sensitive experiments. GroqCloud remains the unique offering for low-latency inference at high token volume, with NVIDIA’s LPX platform incorporating Groq’s licensed technology.