AI News AI资讯 5h ago Updated 2h ago 更新于 2小时前 49

Best GPU Neoclouds 2026: CoreWeave, Nebius, Lambda, Crusoe, and Groq Ranked by Published Pricing and Contracted Power 2026年最佳GPU新云服务商:CoreWeave、Nebius、Lambda、Crusoe和Groq按公开定价与签约算力排名

CoreWeave stands as the only Platinum-rated GPU cloud provider in SemiAnalysis ClusterMAX 2.0, commanding 10-15% premium pricing over competitors on H100, H200, and B200 GPUs, with Q2 2026 revenue of $2.575B (+112% YoY) and a $104B revenue backlog Nebius leads in pricing transparency as the only provider publishing B300 on-demand pricing, with AI cloud ARR reaching $3.0B and Q2 revenue surging 454% YoY to $582M Lambda offers the lowest published B200 on-demand rate at $6.69/GPU-hr but lacks a sp CoreWeave是SemiAnalysis ClusterMAX 2.0中唯一的Platinum评级GPU云服务商,在H100/H200/B200上收取+10-15%溢价定价 Nebius是唯一发布B300按需定价的提供商,AI云ARR达$3.0B,签约功率目标5GW Lambda发布最低B200按需价格($6.69/GPU-hr),但与Microsoft签订多年期数十亿美元GPU供应协议 Crusoe是唯一在价目表上提供AMD MI300X/MI355X的提供商,签约功率4.9GW,管道容量40GW+ Groq完成LPU技术授权给NVIDIA(约$20B),估值$3.5B,计划添加NVID

70
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • CoreWeave stands as the only Platinum-rated GPU cloud provider in SemiAnalysis ClusterMAX 2.0, commanding 10-15% premium pricing over competitors on H100, H200, and B200 GPUs, with Q2 2026 revenue of $2.575B (+112% YoY) and a $104B revenue backlog
  • Nebius leads in pricing transparency as the only provider publishing B300 on-demand pricing, with AI cloud ARR reaching $3.0B and Q2 revenue surging 454% YoY to $582M
  • Lambda offers the lowest published B200 on-demand rate at $6.69/GPU-hr but lacks a spot pricing tier, while Crusoe is the sole provider featuring AMD MI300X/MI355X on its rate card
  • Groq has pivoted to become an NVIDIA Cloud Partner after licensing its LPU technology to NVIDIA in December 2025, planning to add NVIDIA GPU capacity alongside its existing LPU inference cloud across 13 data centers scaling from 54 MW to 200+ MW by 2027
  • The neocloud sector shows extreme capital intensity with combined contracted power exceeding 15 GW across the five providers, massive prepayment deals (70% prepaid at Nebius), and a clear divergence between public companies (CoreWeave, Nebius) and private players heading toward IPOs

Why It Matters

The GPU neocloud market has matured into a highly competitive infrastructure layer critical for AI training and inference, with pricing transparency becoming a key differentiator as buyers demand visibility into per-GPU-hour costs across next-generation silicon. For AI practitioners and researchers, understanding these provider distinctions—particularly around hardware availability (NVIDIA vs. AMD), pricing models (on-demand vs. spot vs. per-token), and contractual structures—directly impacts deployment strategy, cost optimization, and access to cutting-edge GPUs like the B300 and Vera Rubin NVL72.

Technical Details

  • Hardware Roadmaps: CoreWeave achieved the industry's first Vera Rubin NVL72 bring-up and validation in Q2 2026; Nebius serves as a launch partner for Vera CPU and Quantum-X800 CPO; Lambda has multibillion-dollar agreements covering GB300 NVL72 systems; Crusoe uniquely offers AMD MI300X/MI355X; Groq plans NVIDIA GPU capacity alongside its LPU inference cloud
  • Pricing Structure: H100 on-demand ranges from $3.85 (Nebius) to $6.16/node (CoreWeave); B200 on-demand starts at $6.69/GPU-hr (Lambda); Nebius is the sole B300 on-demand publisher; CoreWeave offers up to 60% off reserved pricing with free egress; Lambda and Crusoe use sales-quoted pricing with free egress; Groq uses per-token pricing for inference API
  • Power & Capacity: CoreWeave operates 1.5 GW active / 4.2 GW+ contracted across 51 data centers; Nebius targets 5 GW contracted by YE2026 with 1 GW/year deployment starting 2027; Lambda aims for 1 GW connected by YE2026 and 5 GW contracted; Crusoe has 4.9 GW contracted with 40 GW+ pipeline across 15 US data centers; Groq scaling from 54 MW to 200+ MW across 13 DCs
  • Financial Scale: CoreWeave guiding $12.4-13.2B revenue and $35-39B capex for 2026 with $31B recourse debt; Nebius raised $4.3B converts in H1 plus $5.0B converts priced Aug 19; Lambda closed $1.5B+ Series E and $1B credit facility; Crusoe raised $1.375B Series E at >$10B valuation with $7.1B JV construction loan; Groq raised $1B in 2026 (Jun + Aug) at $3.5B valuation with NVIDIA participation
  • Enterprise Contracts: CoreWeave holds anchor deals with Meta, Anthropic, and OpenAI; Nebius has Microsoft ($17.4B to $19.4B) and Meta ($3B + up to $27B); Lambda has multibillion-dollar Microsoft agreement; Crusoe secured Oracle/OpenAI Abilene 1.2 GW deal; Groq licensed LPU tech to NVIDIA ($20B reported) with 6M+ developers on GroqCloud

Industry Insight

The neocloud market is consolidating around two distinct models: full-spectrum GPU providers (CoreWeave, Nebius, Lambda) competing on hardware breadth and pricing transparency versus specialized inference players (Groq) pivoting toward hybrid architectures after licensing deals, suggesting the industry is bifurcating between training-focused and inference-optimized infrastructure plays.

Prepayment and anchor-deal economics are becoming critical survival metrics—Nebius reported 70% of Q2 deals were prepaid at $40M+ per MW on short-term contracts versus $12M on base 2026 pricing, indicating buyers are locking in capacity amid supply constraints while providers use prepayments to finance massive capex programs.

The AMD presence through Crusoe represents a strategic diversification play as NVIDIA GPU supply remains constrained, but the lack of AMD pricing transparency (sales-quoted only) versus NVIDIA's published rates suggests customers still prioritize NVIDIA for production workloads despite cost premiums of 10-15%.

TL;DR

  • CoreWeave是SemiAnalysis ClusterMAX 2.0中唯一的Platinum评级GPU云服务商,在H100/H200/B200上收取+10-15%溢价定价
  • Nebius是唯一发布B300按需定价的提供商,AI云ARR达$3.0B,签约功率目标5GW
  • Lambda发布最低B200按需价格($6.69/GPU-hr),但与Microsoft签订多年期数十亿美元GPU供应协议
  • Crusoe是唯一在价目表上提供AMD MI300X/MI355X的提供商,签约功率4.9GW,管道容量40GW+
  • Groq完成LPU技术授权给NVIDIA(约$20B),估值$3.5B,计划添加NVIDIA GPU容量 alongside LPU推理云

为什么值得看

本文是2026年GPU Neocloud市场的权威横向对比,覆盖定价、功率容量、硬件路线图、合同结构和独立质量评级,为AI基础设施采购决策提供关键参考。对云服务商选择、供应链策略和技术路线规划具有重要参考价值。

技术解析

  • CoreWeave:Q2 2026收入$2.575B(+112% YoY),收入积压约$104B(排除Q3新增$25B+),已激活功率1.5GW,签约功率4.2GW+(51个数据中心),完成行业首个Vera Rubin NVL72上线验证,2026年CapEx指引$35-39B
  • Nebius:Q2收入$582M(+454% YoY),AI云ARR $3.0B,调整后EBITDA利润率49.7%,Q2 CapEx约$5.7B,现金储备$8.0B,签约功率目标5GW,短期货值$40M+/MW vs 基础$12M/MW
  • Lambda:B200按需$6.69/GPU-hr(最低),无spot层级,与Microsoft签订多年期数十亿美元协议(含GB300 NVL72),2026年H2计划IPO
  • Crusoe:唯一AMD MI300X/MI355X定价,签约功率4.9GW(Abilene 1.2GW已 energized),管道40GW+,Series E $1.375B(估值>$10B),JV建设贷款$7.1B
  • Groq:13个数据中心,功率54MW→200+MW(2027),NCP认证(8/12),LPU技术授权NVIDIA,6M+开发者,$350M Series A(估值$3.5B,NVIDIA参投)

行业启示

  • 市场分层加剧:CoreWeave凭借Platinum评级和溢价定价占据高端,Lambda/Nebius以低价策略争夺市场份额,AMD MI300X/MI355X提供NVIDIA替代方案,供应商多元化趋势明显
  • 资本密集化:头部玩家CapEx指引达$35-39B(CoreWeave)、$5.7B/季度(Nebius),签约功率目标普遍GW级,GPU云成为资本密集型基础设施赛道
  • 技术路线融合:Groq从纯LPU转向NVIDIA GPU+LPU混合架构,反映推理优化与通用GPU的融合趋势,NCP认证和LPU授权标志专用芯片与通用云平台的协同演进

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPU GPU Inference 推理 Training 训练 Deployment 部署 Chip 芯片