2026 年最佳 GPU Neocloud 排名：CoreWeave、Nebius、Lambda、Crusoe 与 Groq 按公开定价和签约电力对比

“新云”（neocloud）标签现在涵盖五家业务模式非常不同的公司。CoreWeave 和 Nebius 是上市公司，会发布季度业绩并向美国证券交易委员会（SEC）提交报告（CoreWeave 提交 10-Q 表格，Nebius 作为外资私人发行人在 6-K 表格上提交）。Lambda 和 Crusoe 是私有公司，正朝 IPO 方向发展。Groq 自 2024 年起在其 LPU 架构上运行 GroqCloud；在 2025 年 12 月将该技术授权给 NVIDIA 之后，独立的 Groq 完全重新聚焦于推理云基础设施。本文从买家真正关心的方面对它们进行比较：公开的定价、部署和合同功率、硬件路线图、合同结构以及独立的质量评级。

CoreWeave（纳斯达克代码：CRWV）报告 2026 年第二季度收入为 25.75 亿美元，同比增长 112%，收入积压约为 1040 亿美元。该积压量不包括第三季度初新增的 250 亿美元以上的新承诺。截至 6 月 30 日，活跃功率达到 1.5 吉瓦，合同功率约为 3.7 吉瓦；CoreWeave 的数据中心页面显示，51 个数据中心合同功率现在超过 4.2 吉瓦。公司在本季度完成了业界首次 NVIDIA Vera Rubin NVL72 的启动和验证。净亏损为 6.26 亿美元，主要因 6.4 亿美元的净利息支出所致。（CoreWeave 2026 年第二季度发布：https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx，CoreWeave AI 数据中心页面：https://coreweave.com/ai-data-centers）

2026 年全年指引为收入 124–132 亿美元，资本支出 350–390 亿美元。管理层预计到年末活跃功率将超过 1.85 吉瓦。（CNBC：https://www.cnbc.com/2026/08/11/coreweave-crwv-q2-earnings-report-2026.html）

Nebius Group（纳斯达克代码：NBIS）报告 2026 年第二季度集团收入为 5.823 亿美元，同比增长 454%。AI 云收入为 5.749 亿美元，同比增长 514%，年化运行率收入为 30 亿美元。AI 云部门的调整后 EBITDA 利润率为 49.7%。第二季度资本支出约为 57 亿美元，公司季度末现金为 80 亿美元。Nebius 将 2026 年底的合同功率目标提高到 5 吉瓦，并计划从 2027 年开始每年部署超过 1 吉瓦。

Lambda是私人的。2025年11月，公司完成了由TWG Global领导的超过15亿美元的E轮融资，并于2026年5月完成了一项价值10亿美元的高级担保信贷设施（由摩根大通主导安排商）完成。Michel Combes，前SoftBank International和Sprint首席执行官，于2026年5月成为首席执行官;联合创始人Stephen Balaban转任首席技术官。Lambda与Microsoft签订了一项数十亿美元、为期多年的协议，涵盖数万块NVIDIA GPU，包括GB300 NVL72系统。《信息》报道IPO目标为2026年上半年，但截至目前尚无公开S-1。

《漂流记》是私人的。2025年10月，Valor Equity Partners和Mubadala Capital领衔，公司筹集了13.75亿美元的E轮股票，估值超过100亿美元。彭博社报道，2026年7月，《克鲁索》正与公司洽谈筹集约30亿美元，估值接近300亿美元。Crusoe为Oracle和OpenAI（第一个星际之门站点）在德克萨斯州阿比林建造了1.2 GW的校园。2026年3月，微软宣布为Microsoft设立一个独立的900兆瓦阿比林园区，使整个站点装机容量达到约2.1 GW。2026年6月9日，Crusoe表示其合同AI基础设施在五个美国园区和Crusoe Cloud中达到4.9 GW，开发管线超过40 GW。

Groq 于 2024 年 3 月围绕其 LPU 推断芯片推出了 GroqCloud。2025年12月，NVIDIA签署了Groq推理技术的非独占许可，据CNBC报道，交易金额约为200亿美元。创始人Jonathan Ross、Sunny Madra总裁及其他员工加入了NVIDIA。Groq保持独立，GroqCloud继续运营。此后，Groq于2026年6月筹集了6.5亿美元，2026年8月以35亿美元估值完成3.5亿美元的A轮融资，计划由英伟达参与。Groq目前运营13个数据中心，预计2027年将从54兆瓦扩展到200+ MW。2026年8月12日，公司加入了NVIDIA云合作伙伴计划，并表示计划“未来”实现NVIDIA加速计算的上线;GPU 容量尚未发布为公开服务。亚当·温特是首席执行官;Disruptive 的 Alex Davis 担任执行主席。

以下价格均直接取自各提供商于2026年8月21日的定价页面。CoreWeave 列出了每个节点的价格；每 GPU 的价格是节点价格除以 GPU 数量。CoreWeave 的页面还显示了相同的每 GPU 价格，作为其“推理单 GPU”价格。Groq 未发布 GPU 小时定价；GroqCloud 按令牌计费，而 NVIDIA GPU 集群访问是计划中的产品，由销售部门处理。CoreWeave 的现货价格因地区而异；下表所示为北美价格。

CoreWeave 收取溢价，并发布了最完整的显卡清单。其 H100 每 GPU 小时价格为 $6.16，比 Nebius 和 Lambda 高约 60%。SemiAnalysis 报告称，在托管的 Slurm 或 Kubernetes 集群上，每 GPU 小时价格，CoreWeave 比 Nebius、Crusoe、Lambda 和 Fluidstack 高出大约 10–15%，其定价更接近大型超大规模云服务商。CoreWeave 也是四家中唯一公布 GB200 NVL72 按需价格的公司。

Nebius 在 Blackwell 上低于 CoreWeave，并且是唯一提供 B300 按需的供应商。其 B200 的价格为 $7.15，低于 CoreWeave 的 $8.60，而其 $7.85 的 B300 价格在此组中没有公布的竞争对手。可抢占价格大约比按需价格低 45%。

Lambda 持有最便宜的 B200 发布实例，价格为 $6.69。其 1-Click Clusters（16 至 2,000+ GPUs，2周至1年期限）价格更高：16、64 和 256+ GPU 的 B200 每小时价格分别为 $9.86、$9.36 和 $8.87。Lambda 没有现货层。第三方追踪显示，随着需求超过供应，Lambda 在 2025–2026 年间将其 H100 SXM 价格从 $2.99 提升至 $3.99。

Crusoe 是唯一的 AMD 选项，也是最便宜的 H200，价格为 $4.29。其 H100 价格为 $3.90，接近 Nebius 和 Lambda。B200、GB200 和 MI355X 需与销售进行沟通。

标价低估了市场。Nebius 披露，其 2026 年第二季度交易的年合同价值平均超过每兆瓦 2,000 万美元，2026 年基础价格接近每兆瓦 1,200 万美元，而第三季度短期容量交易的成交价超过每兆瓦 4,000 万美元。公司还进行了首次容量拍卖。

Vera Rubin：CoreWeave 表示已于 2026 年第二季度完成 Vera Rubin NVL72 的行业首个启动及系统级验证。Nebius 表示已收到首批 Vera Rubin NVL72 系统，并在生产可用前验证计算、网络和编排功能。Lambda 是 GTC 2026 期间 NVIDIA Vera CPU 和 Quantum-X800 InfiniBand 封装光学的发布合作伙伴。Crusoe 宣布支持 Vera CPU 和 NVIDIA Nemotron 3 型号的发布。四家公司均未公布 Rubin 的定价。

Blackwell Ultra（B300 / GB300）：Nebius 是唯一提供公开 B300 按需价格的供应商。CoreWeave 将 GB300 NVL72 和 HGX B300 列为联系销售，B300 即用节点现货价格为每 8-GPU 节点 35.84 美元。Lambda 的微软合同包括 GB300 NVL72 系统。据报道，Nebius 的微软合同涵盖超过 10 万颗 GB300 芯片。

AMD：五家公司中，仅 Crusoe 在其按需卡上提供 AMD Instinct MI300X，并通过销售提供 MI355X。CoreWeave、Nebius 和 Lambda 在其公开页面上仍仅提供 NVIDIA 产品。Groq 的 LPU 是该组中另一款非 NVIDIA 加速器，但按代币出售，而非按 GPU 小时计费。

Groq LPU 和 NVIDIA LPX：NVIDIA 在 GTC 2026 上宣布其 LPX 平台，整合了许可的 Groq 推理技术。Groq 表示，其资金将用于改造现有数据中心，部署 LPX 系统和 NVIDIA GPU 集群。Groq 自有的 LPU 仍然是 GroqCloud 代币 API 的核心引擎。

来源：Nebius 关于 Meta 协议的 6-K：https://www.sec.gov/Archives/edgar/data/1513845/000110465926027886/tm268879d1_6k.htm，Nebius 关于有担保融资的 6-K：https://www.sec.gov/Archives/edgar/data/0001513845/000110465926084452/tm2620683d1_ex99-1.htm，DCD 关于 Nebius–Meta：https://www.datacenterdynamics.com/en/news/nebius-signs-3bn-deal-with-meta-says-current-available-capacity-is-sold-out-as-it-targets-25gw-by-end-of-2026/，DCD 关于 Lambda–NVIDIA：https://www.datacenterdynamics.com/en/news/microsoft-signs-multi-billion-dollar-cloud-capacity-deal-with-lambda/

Nebius–Meta 结构值得仔细阅读。2026 年 3 月的协议包括从 2027 年初开始的 120 亿美元 5 年期专用集群订单。进一步的订单使 Meta 可以访问指定集群的未售出容量，并要求 Meta 在 Nebius 无法将其出售给其他人时购买该容量。这实际上为受覆盖的集群设定了利用率底线。（SEC 6-K：https://www.sec.gov/Archives/edgar/data/1513845/000110465926027886/tm268879d1_6k.htm）

前沿规模训练（10,000+ GPU，多年）：CoreWeave 和 Nebius 是此类公开提供商中具有 SEC 报告的财务规模和经过验证的 NVL72 机架级操作经验的两个公司。CoreWeave 拥有可靠的业绩记录和白金评级。Nebius 提供更优的承诺定价，并且现在正在销售 2027 年的容量。两家公司手中都拥有 Vera Rubin 单元。

中型训练和微调（16–512 GPU，数周至数月）：Lambda 的 1-Click Clusters 和 Nebius 的按需 B200/B300 是最易访问的。Crusoe 的 H200 以 $4.29 是适用于大内存工作负载的最便宜 Hopper 类选项。

预算敏感型实验：Nebius 可抢占 H100 每 GPU $2.15 和 CoreWeave 现货 H100 每 GPU $2.46 是这里公布的最低价格。Lambda 没有现货级别。

AMD GPU 评估：在这一组中，Crusoe 是唯一的选择。

高令牌量低延迟推理：GroqCloud 的每令牌 LPU 端点仍然是独特的产品。Nebius Token Factory 和 Crusoe Managed Inference 是基于 GPU 的替代方案，并公布了按令牌计费的价格。

需要公开资产负债表的买家：仅 CoreWeave 和 Nebius 发布季度财报并向 SEC 提交。两家公司年度报表均经过审计；本文使用的第二季度数据为未经审计的季度结果。

Asif Razzaq 是 Marktechpost Media Inc. 的 CEO。作为一位富有远见的企业家和工程师，Asif 致力于利用人工智能的潜力为社会造福。他最近的努力是推出人工智能媒体平台 Marktechpost，该平台以其对机器学习和深度学习新闻的深入报道而脱颖而出，这些报道既技术可靠，又易于广大受众理解。该平台月访问量超过 200 万次，显示了其在受众中的受欢迎程度。

The ‘neocloud’ label now covers five companies with very different business models. CoreWeave and Nebius are public companies that publish quarterly results and file reports with the SEC (CoreWeave on Form 10-Q, Nebius as a foreign private issuer on Form 6-K). Lambda and Crusoe are private and heading toward IPOs. Groq has run GroqCloud on its LPU architecture since 2024; after licensing that technology to NVIDIA in December 2025, the independent Groq refocused entirely on inference-cloud infrastructure. This article compares them on what actually matters to a buyer: published pricing, deployed and contracted power, hardware roadmap, contract structure, and independent quality ratings.

CoreWeave (Nasdaq: CRWV) reported Q2 2026 revenue of $2.575 billion, up 112% year over year, with a revenue backlog of roughly $104 billion. That backlog excludes more than $25 billion of new commitments added in early Q3. Active power reached 1.5 GW and contracted power reached about 3.7 GW as of June 30; CoreWeave’s data center page now states 4.2 GW+ contracted across 51 data centers. The company completed the industry’s first bring-up and validation of NVIDIA Vera Rubin NVL72 during the quarter. Net loss was $626 million, driven by $640 million in net interest expense. (CoreWeave Q2 2026 release：https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx, CoreWeave AI data centers page：https://coreweave.com/ai-data-centers)

Full-year 2026 guidance is $12.4–13.2 billion in revenue and $35–39 billion in capex. Management expects more than 1.85 GW of active power by year end. (CNBC：https://www.cnbc.com/2026/08/11/coreweave-crwv-q2-earnings-report-2026.html)

Nebius Group (Nasdaq: NBIS) reported Q2 2026 group revenue of $582.3 million, up 454% year over year. AI cloud revenue was $574.9 million, up 514%, with annualized run-rate revenue of $3.0 billion. The AI cloud segment posted a 49.7% adjusted EBITDA margin. Q2 capex was about $5.7 billion, and the company ended the quarter with $8.0 billion in cash. Nebius raised its year-end 2026 contracted power target to 5 GW and plans to deploy more than 1 GW per year starting in 2027.

Lambda is private. It closed a Series E of more than $1.5 billion led by TWG Global in November 2025 and a $1 billion senior secured credit facility (J.P. Morgan lead arranger) in May 2026. Michel Combes, former CEO of SoftBank International and Sprint, became CEO in May 2026; co-founder Stephen Balaban moved to CTO. Lambda has a multibillion-dollar, multi-year agreement with Microsoft covering tens of thousands of NVIDIA GPUs, including GB300 NVL72 systems. The Information has reported an IPO target of H2 2026, but no public S-1 exists as of this writing.

Crusoe is private. It raised a $1.375 billion Series E at a valuation above $10 billion in October 2025, led by Valor Equity Partners and Mubadala Capital. In July 2026, Bloomberg reported Crusoe was in talks to raise roughly $3 billion at a valuation near $30 billion. Crusoe builds the 1.2 GW Abilene, Texas campus for Oracle and OpenAI (the first Stargate site). In March 2026 it announced a separate 900 MW Abilene campus for Microsoft, bringing the full site to about 2.1 GW. On June 9, 2026, Crusoe said its contracted AI infrastructure reached 4.9 GW across five US campuses and Crusoe Cloud, with a development pipeline above 40 GW.

Groq launched GroqCloud in March 2024 around its LPU inference chip. It changed shape in December 2025, when NVIDIA signed a non-exclusive license for Groq’s inference technology in a transaction reported by CNBC at about $20 billion. Founder Jonathan Ross, president Sunny Madra, and other staff joined NVIDIA. Groq stayed independent, and GroqCloud kept running. Since then, Groq raised $650 million in June 2026 and a $350 million Series A in August 2026 at a $3.5 billion valuation, with planned NVIDIA participation. Groq now operates 13 data centers and expects to scale from 54 MW to 200+ MW in 2027. On August 12, 2026 it joined the NVIDIA Cloud Partner program and said it plans to bring NVIDIA accelerated computing online “in the future”; that GPU capacity is not yet a published service. Adam Winter is CEO; Alex Davis of Disruptive is Executive Chairman.

Prices below were read directly from each provider’s pricing page on August 21, 2026. CoreWeave lists per-node prices; per-GPU values are the node price divided by GPU count. CoreWeave’s page also shows the identical per-GPU figure as its “inference single GPU” price. Groq does not publish GPU-hour pricing; GroqCloud is priced per token, and NVIDIA GPU cluster access is a planned offering handled by sales. CoreWeave spot prices vary by region; the figures below are the North America rates.

CoreWeave charges a premium and publishes the most complete card. Its H100 rate of $6.16/GPU-hr is about 60% above Nebius and Lambda. SemiAnalysis reports CoreWeave commands roughly 10–15% more per GPU-hour than Nebius, Crusoe, Lambda, and Fluidstack on managed Slurm or Kubernetes clusters, and that its pricing sits closer to the big hyperscalers. CoreWeave is also the only one of the four publishing a GB200 NVL72 on-demand price.

Nebius undercuts CoreWeave on Blackwell and is the only one with B300 on-demand. Its B200 rate of $7.15 is below CoreWeave’s $8.60, and its $7.85 B300 rate has no published competitor in this group. Preemptible rates run roughly 45% below on-demand.

Lambda holds the cheapest published B200 instance at $6.69. Its 1-Click Clusters (16 to 2,000+ GPUs, 2-week to 1-year terms) are priced higher: $9.86, $9.36, and $8.87 per B200-hour at 16, 64, and 256+ GPUs respectively. Lambda has no spot tier. Third-party trackers note Lambda raised its H100 SXM rate from $2.99 to $3.99 over 2025–2026 as demand outran supply.

Crusoe is the only AMD option and the cheapest H200 at $4.29. Its H100 rate of $3.90 is close to Nebius and Lambda. B200, GB200, and MI355X require a sales conversation.

List prices understate the market. Nebius disclosed that its Q2 2026 deals averaged more than $20 million in annual contract value per MW, that the 2026 base is priced near $12 million per MW, and that Q3 short-term capacity deals are clearing above $40 million per MW. The company also ran its first capacity auction.

Vera Rubin : CoreWeave says it completed the industry’s first bring-up and system-level validation of Vera Rubin NVL72 in Q2 2026. Nebius says it has received its first Vera Rubin NVL72 systems and is validating compute, networking, and orchestration ahead of production availability. Lambda was a GTC 2026 launch partner for NVIDIA’s Vera CPU and Quantum-X800 InfiniBand co-packaged optics. Crusoe announced launch support for the Vera CPU and NVIDIA Nemotron 3 models. None of the four publishes Rubin pricing yet.

Blackwell Ultra (B300 / GB300) : Nebius is the only provider with a public B300 on-demand price. CoreWeave lists GB300 NVL72 and HGX B300 as contact-sales, with a B300 spot rate of $35.84 per 8-GPU node. Lambda’s Microsoft contract includes GB300 NVL72 systems. Nebius’s Microsoft contract was reported to cover more than 100,000 GB300 chips.

AMD : Crusoe is the only one of the five with AMD Instinct MI300X on its on-demand card and MI355X available via sales. CoreWeave, Nebius, and Lambda remain NVIDIA-only on their public pages. Groq’s LPU is the other non-NVIDIA accelerator in this group, but it is sold per token, not per GPU-hour.

Groq LPU and NVIDIA LPX : NVIDIA announced its LPX platform, incorporating licensed Groq inference technology, at GTC 2026. Groq says its capital will fund fit-out of existing data centers with LPX systems and NVIDIA GPU clusters. Groq’s own LPU remains the engine behind GroqCloud’s token API.

Sources: Nebius 6-K on Meta agreement：https://www.sec.gov/Archives/edgar/data/1513845/000110465926027886/tm268879d1_6k.htm, Nebius 6-K on secured financing：https://www.sec.gov/Archives/edgar/data/0001513845/000110465926084452/tm2620683d1_ex99-1.htm, DCD on Nebius–Meta：https://www.datacenterdynamics.com/en/news/nebius-signs-3bn-deal-with-meta-says-current-available-capacity-is-sold-out-as-it-targets-25gw-by-end-of-2026/, DCD on Lambda–NVIDIA：https://www.datacenterdynamics.com/en/news/microsoft-signs-multi-billion-dollar-cloud-capacity-deal-with-lambda/

The Nebius–Meta structure is worth reading closely. The March 2026 agreement includes $12 billion of dedicated 5-year cluster orders starting early 2027. A further order gives Meta access to unsold capacity on specified clusters, and obliges Meta to buy that capacity if Nebius cannot sell it to others. That is effectively a floor on utilization for the covered clusters. (SEC 6-K：https://www.sec.gov/Archives/edgar/data/1513845/000110465926027886/tm268879d1_6k.htm)

Frontier-scale training (10,000+ GPUs, multi-year) : CoreWeave and Nebius are the two public providers here with SEC-reported financial scale and demonstrated NVL72 rack-scale operating experience. CoreWeave has the reliability track record and Platinum rating. Nebius has better committed pricing and is selling 2027 capacity now. Both have Vera Rubin units in hand.

Mid-size training and fine-tuning (16–512 GPUs, weeks to months) : Lambda’s 1-Click Clusters and Nebius’s on-demand B200/B300 are the most accessible. Crusoe’s H200 at $4.29 is the cheapest Hopper-class option for memory-heavy jobs.

Budget-sensitive experiments : Nebius preemptible H100 at $2.15 and CoreWeave spot H100 at $2.46 per GPU are the lowest published rates here. Lambda has no spot tier.

AMD GPU evaluation : Crusoe is the only choice in this group.

Low-latency inference at high token volume : GroqCloud’s per-token LPU endpoints remain the distinctive offering. Nebius Token Factory and Crusoe Managed Inference are the GPU-based alternatives with published per-token rates.

Buyers needing a public balance sheet : Only CoreWeave and Nebius publish quarterly results and file with the SEC. Both companies’ annual statements are audited; the Q2 figures used here are unaudited quarterly results.

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.