公司指南

Lambda (Lambda Labs):GPU雲端與硬體 AI 服務

Lambda,簡稱Lambda Labs,是一家美國公司,按小時租用NVIDIA GPU用於訓練和運行AI模型,為Microsoft等公司建立大型GPU集群,並銷售深度學習工作站和伺服器。

聆聽此指南AI 語音敘述 · ~8 min

本頁閱讀時間 9 分鐘
  1. 概述
  2. 重點摘要
  3. 深入探討
  4. 為 Lambda Cloud 上為期三天的微調運行製定預算
  5. 現實世界的實施
  6. Lambda 的未來 (Lambda Labs):人工智慧的 GPU 雲端和硬體
  7. 付諸實踐
  8. 資料來源與延伸閱讀
  9. 不斷探索
  10. 常見問題

概述

這很重要,因為像它這樣的專用 GPU 雲端(通常稱為 Neocloud)現在提供了前沿人工智慧背後的大部分運算能力。

重點摘要

  1. Lambda 按小時租用 NVIDIA GPU,為大客戶建立大型 GPU 集群,並提供名為 Lambda Stack 的預先配置軟體包。
  2. 該公司前身是一款 2015 年火爆的自拍藝術應用程序,該應用程式的 AWS 法案促使其創始人構建了自己的 GPU 伺服器。
  3. 2026 年 10 月,Lambda 列出的 H100 SXM GPU 價格為每 GPU 小時 3.99 美元至 4.29 美元,B200 GPU 價格為每 GPU 小時 6.69 美元至 6.99 美元。
  4. Microsoft於2025年11月與Lambda簽署了價值數十億美元的GPU部署協議,Anthropic於2026年8月做出了350億美元的承諾。
  5. Sprint 前執行長 Michel Combes 於 2026 年 5 月成為 Lambda 首席執行官,聯合創始人 Stephen Balaban 擔任首席技術長。

深入探討

Lambda is a company that sells access to graphics processing units, or GPUs, the chips that do most of the arithmetic in modern AI. It does this in three main ways: renting individual GPU machines by the hour through Lambda Cloud, renting or building large clusters of thousands of GPUs for big customers, and, historically, selling physical workstations and servers preloaded with AI software. For most readers the relevant product is the cloud: a developer can sign in, choose a machine with one to eight NVIDIA GPUs, connect over SSH, and start training or running a model within minutes, paying only for the time used.

The company's story starts with face recognition. Next Platform's 2020 profile reports that Stephen Balaban was the first hire at Perceptio, a neural-network face recognition startup that Apple acquired in August 2013, and that he founded Lambda a few months later with his twin brother Michael Balaban; TechCrunch and Verdict give 2012 as the founding year. In 2015 the founders built a selfie-to-art-style app on a single AWS GPU server. It went viral, reaching about 30,000 sign-ups a day at its peak, and the AWS bill climbed to roughly $40,000 a month. Building their own GPU machines for about $60,000, hosted at first in a friend's house, cut that bill by about 95%, and that experience became the business. At the end of 2016 Lambda began selling AI servers, followed in 2017 by its Quad workstations and Blade servers, and in 2018 it formally announced Lambda GPU Cloud and released Lambda Stack.

Renting a GPU on Lambda works much like any cloud virtual machine, with fewer layers. According to Lambda's documentation, a user picks an instance type and one of 13 regions, adds an SSH key, and optionally attaches a persistent filesystem in the same region. The default image is Ubuntu 22.04 with Lambda Stack, which bundles NVIDIA drivers, CUDA, cuDNN, NCCL, PyTorch, TensorFlow, JAX, Triton and JupyterLab, so a training script can usually run without installing a framework first. Only SSH port 22 is open by default. Filesystems are billed per gigabyte per month and keep billing after the machine is deleted, a common source of surprise charges.

Prices on Lambda's public pricing page in October 2026 illustrate the range. An NVIDIA B200 with 180 GB of memory costs $6.69 per GPU-hour on an 8-GPU instance and $6.99 on a single-GPU instance. An H100 SXM with 80 GB costs $3.99 to $4.29 per GPU-hour depending on instance size, an H100 PCIe $3.29, an A100 40 GB $1.99, and older cards such as the V100 as little as $0.79. For multi-node work, 1-Click Clusters reserved for two weeks to a year are listed at $5.54 to $6.16 per H100 GPU-hour and $8.87 to $9.86 per B200 GPU-hour, falling with cluster size. Longer and larger commitments are priced on request.

From about 2025 the company's scale changed sharply. Verdict reported a $480 million Series D in February 2025 and a Series E of more than $1.5 billion in November 2025, led by TWG Global, aimed at building what the company called gigawatt-scale AI factories. In November 2025 Microsoft signed a multibillion-dollar agreement for Lambda to deploy tens of thousands of NVIDIA GPUs, including GB300 NVL72 rack systems; CEO Stephen Balaban told TechCrunch that Lambda had worked with Microsoft for more than eight years. In May 2026, Bloomberg reported that former Sprint chief executive Michel Combes became CEO, with Stephen Balaban moving to chief technology officer and Michael Balaban to chief product officer. In October 2026 TechCrunch, citing The Wall Street Journal, reported that Lambda was raising up to $4 billion at a $14.5 billion pre-money valuation ahead of a planned 2027 public offering, and that its backlog of contracted orders had grown from $15 billion in June to $50 billion in September.

Compared with the large general-purpose clouds such as Amazon Web Services, Microsoft Azure and Google Cloud, Lambda offers a narrower menu focused on GPUs, simple hourly pricing and a preconfigured machine learning image, rather than a broad catalog of managed services. Compared with other GPU-focused providers such as CoreWeave and Nebius, which TechCrunch groups with Lambda as NVIDIA-backed neoclouds, the differences come down to price, available GPU models, cluster networking, regions and contract terms, all of which change often. Buying hardware outright remains an option for steady, round-the-clock workloads; renting suits bursty experiments and projects that need the newest chips without a capital purchase.

There are practical limits. Popular GPU types are not always available in every region, and on-demand capacity for the newest chips can run out at busy times. Each instance lives in one region, and filesystems must be in the same region. The GH200 instance uses an ARM processor, so software built only for x86 will not run on it. Lambda's documentation also prohibits cryptocurrency mining. At the company level, TechCrunch noted in October 2026 that a large share of Lambda's backlog depends on a single customer, Anthropic, which concentrates risk.

In practice, using Lambda requires an account with a payment method, an SSH key, and a plan for data: where the dataset lives, how it gets to the region, and how checkpoints are saved to a persistent filesystem before the instance is terminated. The main cost is GPU-hours, so the most effective savings come from testing code on a cheap single-GPU instance first, then scaling up only when the job is known to work.

04工作範例

為 Lambda Cloud 上為期三天的微調運行製定預算

  1. 查看定價頁面並注意 2026 年 10 月列出的 8x H100 SXM 費率,每 GPU 小時 3.99 美元。

  2. 乘以:8 個 GPU x 3.99 美元 x 72 小時 = 2,298.24 美元計算費用(不含任何銷售稅)。

  3. 新增儲存:在同一區域中為資料集和檢查點建立持久檔案系統,記住它按每月每 GB 計費,直到刪除。

  4. 在大規模運行之前,在單一較便宜的 GPU 執行個體(例如每小時 1.29 美元的 A10)上測試訓練腳本一小時。

  5. 使用預設 Lambda Stack 映像啟動 8-GPU 實例,附加檔案系統,並在訓練期間向其儲存檢查點。

  6. 訓練完成後,複製結果,終止實例,並刪除檔案系統(如果不再需要)。

它顯示了什麼

使用 8 個 H100 GPU 運行 72 小時的運算預算約為 2,300 美元,外加少量儲存費用。運行後對照計費頁面核對;較大的間隙通常意味著實例或檔案系統仍在運行。

現實世界的實施

一個研究團隊從 Lambda Cloud 主機啟動了 8 個 NVIDIA H100 SXM 實例,2026 年 10 月的售價為每 GPU 小時 3.99 美元,對模型進行了幾天的微調,然後終止它以停止計費。

根據 TechCrunch 報導,Microsoft 於 2025 年 11 月與 Lambda 簽署了一項價值數十億美元的協議,部署數萬個 NVIDIA GPU,包括 GB300 NVL72 系統。

一家新創公司預留了由 64 個 H100 GPU 組成的一鍵式集群數週來訓練模型,並為每 GPU 小時支付 5.85 美元,預留期限為兩週到一年。

某大學實驗室在自己的 Ubuntu 伺服器上安裝 Lambda Stack(Lambda 的免費 BSD 許可安裝腳本),以便一步取得 NVIDIA 驅動程式、CUDA、cuDNN、PyTorch 和 JAX 的匹配版本。

Anthropic 於 2026 年 8 月下旬向 Lambda 做出了 350 億美元的承諾,TechCrunch 報導,這在很大程度上推動了 Lambda 訂單積壓的增長。

Lambda 的未來 (Lambda Labs):人工智慧的 GPU 雲端和硬體

Lambda's direction is tied to the wider market for AI computing. In October 2026 TechCrunch, citing The Wall Street Journal, reported a planned 2027 initial public offering, pushed back from 2026 amid market uncertainty, and a $50 billion order backlog driven largely by a $35 billion Anthropic commitment. Whether those contracts translate into steady revenue, and how dependent the company remains on a few very large customers, are open questions that a public filing would make clearer. Operationally, the challenge is building enough data center capacity and power for new generations of NVIDIA hardware. Lambda's November 2025 Series E was explicitly framed around gigawatt-scale AI factories, and its product pages in 2026 already list NVIDIA's Vera Rubin platform alongside B200 and H200 systems. Competition with CoreWeave, Nebius and the large hyperscale clouds affects both prices and access to scarce chips, and list prices have changed repeatedly as new GPU generations arrive. For individual developers and researchers, the practical question is whether self-serve, hourly GPU access stays a priority as the company shifts toward multibillion-dollar contracts. Readers should check the current pricing page and documentation before planning a project, since instance types, regions and terms change often.

付諸實踐

為什麼這很重要

  • 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
  • 成本與預算. 商業條款和部署選項會影響長期成本和風險。
  • 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。

留意

  • 發佈公告可能會超過實際生產工作流程的穩定性。
  • API 定價或政策轉變可能會在一夜之間打破假設。
  • 單一供應商依賴性增加了鎖定和遷移成本。

後續步驟

  1. 使用您自己的任務和資料集評估提供者。
  2. 在整合之前查看隱私、安全和法律條款。
  3. 維護跨模型或供應商的後備計劃。
  4. 監控發行說明,以便路線圖的變更不會讓團隊感到意外。

資料來源與延伸閱讀

  1. Lambda拉姆达定价
  2. Lambda documentation按需实例
  3. LambdaLambda Stack: Ubuntu 的深度學習軟體
  4. The Next Platform偶然的人工智能系统制造商
  5. VerdictLambda 融資超過 15 億美元用於擴展 AI 超級運算基礎設施
  6. TechCrunchLambda 與 Microsoft 簽署價值數十億美元的人工智慧基礎設施協議
  7. BloombergAI 雲端供應商 Lambda 任命前 Sprint 執行長為新領導者
  8. TechCrunch人工智慧運算新創公司 Lambda 將在計劃 IPO 之前籌集 40 億美元

不斷探索

免費電子報

獲取每日 AI 簡報

每個工作日上午提供三則經過驗證的 AI 新聞,以淺顯易懂的英文撰寫。永久免費,無廣告。

每個工作日一封電子郵件。一鍵取消訂閱。我們絕不出售或分享您的地址。

測試一下自己

參加 Lambda (Lambda Labs):用於 AI 的 GPU 雲端和硬體測驗

每個答案皆有即時回饋,通過課程後可獲得具備可驗證 ID 的可分享證書。

開始測驗

支持免費的 AI 教育。AI Understanding 是一個 501(c)(3) 非營利組織——絕無廣告,絕無付費牆。 捐款

常見問題

什麼是 Lambda (Lambda Labs):AI 的 GPU 雲端和硬體?

Lambda,簡稱Lambda Labs,是一家美國公司,按小時租用NVIDIA GPU用於訓練和運行AI模型,為Microsoft等公司建立大型GPU集群,並銷售深度學習工作站和伺服器。這很重要,因為像它這樣的專用 GPU 雲端(通常稱為 Neocloud)現在提供了前沿人工智慧背後的大部分運算能力。

Lambda 目前的主要業務是什麼?

Lambda 是一家 GPU 雲端供應商,也銷售深度學習工作站和伺服器,主要圍繞 NVIDIA 硬體建置。

什麼是「Lambda 堆疊」?

Lambda Stack 預先安裝了 CUDA、cuDNN、PyTorch 和 TensorFlow,因此機器可以開箱即用地進行 ML。

Lambda 主要提供誰的 GPU?

Lambda 與 NVIDIA 密切相關,提供 A100、H100、H200 和 Blackwell 系列 GPU。

Lambda 的一鍵式叢集使用什麼網路技術來連接 GPU 來執行大型訓練作業?

一鍵式叢集使用 NVIDIA Quantum InfiniBand 來實現多節點訓練所需的高頻寬、低延遲互連。

Lambda 通常被描述為哪一類雲端公司?

Lambda 是一個專門從事人工智慧 GPU 運算的“neocloud”,與 CoreWeave 和大型超大規模企業競爭。