AI & Infrastructure CapexAI 与基础设施资本开支
DigitalOcean Holdings, Inc.
DOCN 4.5Price and target zones价格与目标区间
Above fair value高于合理价值
What the company does公司做什么
DigitalOcean gives cloud computing to developers, startups, and small businesses. It rents virtual servers, storage, databases, and networking on demand. Customers pay by usage, with simple and low prices. The company also adds AI and machine-learning tools, such as GPU servers and a GenAI platform.DigitalOcean 为开发者、初创公司和小型企业提供云计算服务。它按需出租虚拟服务器、存储、数据库和网络。客户按用量付费，价格简单且低廉。公司还增加了人工智能和机器学习工具，例如 GPU 服务器和 GenAI 平台。
Earns from营收来源 DigitalOcean earns money when customers use its cloud platform. Customers pay monthly fees for servers, storage, databases, and add-on services. More platform usage and more high-spend customers raise the revenue. FY2025 revenue was $901.4M, up 15% from the year before.DigitalOcean 在客户使用其云平台时赚取收入。客户为服务器、存储、数据库和附加服务支付月费。平台用量增加和高消费客户增多会提升收入。2025 财年营收为 9.014 亿美元，同比增长 15%。
Read the full description查看完整介绍
DigitalOcean Holdings, Inc., through its subsidiaries, operates an agentic inference cloud platform in North America, Europe, Asia, and internationally. The company provides AI and Digital Native Enterprises build, run, and scale intelligent applications for growing technology companies. It also offers infrastructure-as-a-service (IaaS) solutions comprising compute, storage, and networking products, including cloud firewalls, managed load balancers, NAT gateways, and virtual private cloud software, as well as IP address management and domain name system management. In addition, the company provides platform-as-a-service (PaaS) and software-as-a-service (SaaS) solutions, such as managed databases; managed Kubernetes and container registry; application platform to build, deploy, and scale applications; Functions, a serverless compute solution; and Uptime for real-time uptime and latency alerts, as well as managed hosting and DigitalOcean Marketplace, a platform where developers can find pre-configured applications and solutions. Further, it offers artificial intelligence (AI)/machine learning (ML) applications comprising GPU droplets; bare metal GPUS, which provides access to a GPU server without any virtualization layer and gives developers with customizable server for their use case; and Jupyter Notebooks that provides cloud workspace and managed interactive development environment for exploring data and training, and building machine learning models. Its customers use its platform in various industry verticals, such as online gaming, fintech, and cybersecurity, as well as for a range of use cases, including building and hosting websites, web and mobile applications development, AI integration, and building AI products and applications. DigitalOcean Holdings, Inc. was incorporated in 2012 and is headquartered in Broomfield, Colorado.
Regional breakdown地区营收构成 share of revenue by region按地区划分的营收占比
North America is DigitalOcean's largest market. It holds customers in the United States and Canada. The company runs data centers in New York, San Francisco, and Toronto for this region.北美是 DigitalOcean 最大的市场。它包含美国和加拿大的客户。公司为该地区在纽约、旧金山和多伦多运营数据中心。
- Core cloud compute核心云计算 — Droplet virtual servers and managed services.Droplet 虚拟服务器和托管服务。
Europe is the second-largest market. It covers customers across the European Union and the United Kingdom. Data centers in Amsterdam, Frankfurt, and London serve this region.欧洲是第二大市场。它涵盖欧盟和英国的客户。位于阿姆斯特丹、法兰克福和伦敦的数据中心服务于该地区。
- Managed databases托管数据库 — Managed database and Kubernetes services.托管数据库和 Kubernetes 服务。
Asia is a strong growth market. It includes India, Singapore, and nearby countries. Data centers in Bangalore and Singapore serve local developers.亚洲是强劲的增长市场。它包括印度、新加坡及邻近国家。位于班加罗尔和新加坡的数据中心服务于当地开发者。
- AI/ML toolsAI/ML 工具 — GPU Droplets and GenAI platform.GPU Droplet 和 GenAI 平台。
The rest of the world covers all other regions. This includes Australia, South America, Africa, and the Middle East. A data center in Sydney serves the Asia-Pacific edge.世界其他地区涵盖所有其他区域。这包括澳大利亚、南美、非洲和中东。位于悉尼的数据中心服务于亚太边缘地区。
- Object storage对象存储 — Spaces object storage and networking.Spaces 对象存储和网络服务。
Company facts公司资料
- Founded成立时间
- 2012
- Headquarters总部
- Broomfield, Colorado
- Employees员工人数
- 1,462
- Industry行业
- Software - Infrastructure
- Country国家
- United States
Investment thesis投资逻辑 from the watchlist来自关注列表
✂️ ANALYZED 8/12 — first file ever on a position held without one. TRIM. The dead $70+ is replaced (see below). The business is executing better than almost anything else held: Q2 revenue $281M +29% and ACCELERATING (from ~14% a year ago), ARR $1,125M +29% with a record $93M incremental ARR (+191%), AI customer ARR $234M +212% = 21% of total, NDR 102% (three-year high), adj EBITDA 40% margin, RPO $894M up >10x, first nine-figure annual commitment signed, guidance raised twice (FY26 now $1.170-1.180B, +30-31%), FY27 ≥50% growth reaffirmed. Rule of 40 = 69.
What's working
- And the AI revenue is the good kind — 85% is inference + attached core cloud, NOT bare-metal GPU rental; inference +~800% YoY, 6,000+ customers in 90 days, tokens +30x in 60 days
- Liability management was competent: ~$472M of the 0% 2030 converts repurchased in July at ~96% share-for-share offset; ~$153M principal left; the ~$312M of 2026 notes maturing 12/01 is covered by $767M cash
- Correct that, and the leverage objection disappears
Red flags
- But the capital intensity did not disappear — it moved off the capex line. Q2 reported capex was only $41.6M (15% of revenue) against a FY25 run-rate of $268.5M. Meanwhile gross PP&E went $1,290.6M → $2,249.1M in twelve months (+$958.5M) and finance leases went $274.4M → $608.9M (+122%, +60% in Q2 alone). The company added ~$1B of assets on ~$400M of operating cash flow and reported a positive FCF margin throughout — because its adjusted-FCF definition explicitly excludes equipment acquired under financing arrangements and finance leases. CFO: "we closely match our cash outflow with our revenue by financing equipment." This is [[pattern-ai-build-inflates-earnings-while-destroying-fcf]] — and AKAM, which puts the same spend on the capex line where a screener sees it, has the uglier and more honest disclosure.
- 155MW of committed capacity against $1.18B of revenue, and no GPU utilization figure has ever been disclosed.
- Concentration is doubling: the $1M+ cohort went 9% → 23% of ARR, top 25 customers = 20% — and management retired NDR as a headline metric in the same quarter it printed a three-year high. That is not evidence anything is wrong; it is the removal of the instrument that would show you
- Gross margin 59.9% → ~55% — the AI growth arrives at structurally lower margin
- ROIC 9.8% vs an 11-12% WACC (beta 1.59) with invested capital +46% YoY
- No cost advantage — Hetzner is 3-5x cheaper on equivalent compute; the moat is a decade-deep tutorial/SEO funnel, and agentic provisioning is exactly what devalues it — the AI wave driving the revenue erodes the acquisition channel. Moat 4.0
- Dilution, not retirement: 91.9M → 116.9M shares in eight months; buybacks effectively over ($600M/$488M in 2022-23 → $60M/$82M)
- Insiders: zero open-market buys, selling clustered at the top — CFO 25,000 @ $152.50 and 10,000 @ $170.07, director exercise-and-same-day-sell @ $147.62, plus Access Industries out for ~$557M at $148-162
- Two of the five legs of the +295% move were mechanical — Russell 1000 inclusion 6/29 and a 16.35%-of-float short base
Watch
- The old "87% D/A leverage" line is STALE and struck — a ~$900M Q1 equity raise flipped the balance sheet: equity −$28.7M → +$930.7M, net debt $1,041M → $154M (6/30)
- Vendor multiples are void in three places:
PE(ttm) 60.94uses an EPS inflated by a ~$120M tax-valuation-allowance release (FY25 pretax ~$139M vs net income $259.3M — [[pattern-net-margin-above-operating-margin-is-a-tripwire]] + [[pitfall-tax-valuation-allowance-round-trip-breaks-eps]]); tax-normalised trailing P/E is ~129x, not 60.9x; andPE(fwd) 72.40sitting above trailing is [[pitfall-forward-pe-above-trailing-pe-flags-an-inflated-base]] firing. Graham ($20.95, or $14.40 normalised) is not a target — it just says there is no floor within sight of this price. ⚖️ Real multiples: 13.4x EV/FY26 revenue 34.5x EV/adj EBITDA ~112x EV/adj FCF (0.9% FCF yield) ~97x forward non-GAAP EPS. 🔍 Reverse-DCF: $15.8B EV needs $1.2-1.6B of steady-state FCF ≈ $6-8B of revenue, 5-7x today's, i.e. ~26% compounding for seven straight years against Hetzner, Cloudflare and three hyperscalers. FV $70-165 (centre $105) — deliberately wide, because a 30-50% grower with a 4/10 moat genuinely has that distribution. Spot $134 sits between base and bull, ~28% above centre. Entry $75-90. ✂️ Trim $150 — = 12x EV/NTM revenue (NTM ≈ $1.45B). Dollar form is deliberate and the reason is mechanical: the site derives a multiple-trim from Yahoo's forward P/E, whose implied EPS of $1.85 is ~34% above the company's own $1.35-1.40 non-GAAP guide, and the trailing EPS it would otherwise key off carries the tax release. Hold the stub, do not add, let $150 work.* - Cluster note: DOCN passes THIS sleeve's test (a datacenter-buildout pause hurts it — 155MW committed), but it is on the opposite side of the trade from AVGO/KLAC/AMAT — it BUYS the capacity they SELL. Its failure mode is AI demand and cost of capital, not the capex cycle. Its true peer is AKAM.
Would upgrade on
- Upgrade: an adjusted-FCF definition that INCLUDES lease-financed equipment first disclosure of GPU utilization, if healthy NDR restored as a metric any insider open-market purchase.
✂️ **8/12 已分析 — 一个长期持有却从未有过分析文件的仓位，这是第一份。 TRIM（减持）。 原来那条已经失效的 $70+ 被替换（见下）。 **
这门生意的执行力比投资组合里几乎任何一家都好：Q2 收入 $281M，+29% 且在加速（一年前只有约 14%），ARR $1,125M，+29%，单季新增 ARR 创纪录 $93M（+191%），AI 客户 ARR $234M，+212%，占总量 21%，净收入留存率 102%（三年新高），调整后 EBITDA 利润率 40%，剩余履约义务（RPO）$894M，同比增长超过 10 倍，签下第一笔九位数的年度承诺合同，指引年内上调两次（FY26 现为 $1.170-1.180B，+30-31%），并重申 FY27 ≥50% 的收入增长。 **40 法则得分 = 69。 **
✅ 而且这部分 AI 收入是好的那一种 —— 85% 来自推理服务和随之带动的核心云业务，而不是裸金属 GPU 出租；推理收入同比增长约 800%，90 天内新增 6,000 多个客户，60 天内 token 用量增长 30 倍。 ✅ 负债管理做得称职：7 月以约 96% 的股份对冲比例回购了约 $472M 的 2030 年零息可转债；只剩约 $153M 本金；12 月 1 日到期的约 $312M 2026 年票据由 $767M 现金完全覆盖。
🔴 旧的「87% 资产负债率」那句话已经过时，予以删除 —— Q1 一笔约 $900M 的股权融资翻转了资产负债表：股东权益从 −$28.7M 变成 +$930.7M，净负债从 $1,041M 降到 $154M（截至 6/30）。 ✅ 修正这一点之后，杠杆这个反对理由就消失了。
🚩 但资本密集度并没有消失 —— 它只是从资本开支这一行挪走了。 Q2 报告的资本开支只有 $41.6M（占收入 15%），而 FY25 的年化水平是 $268.5M。 与此同时，固定资产原值在十二个月里从 $1,290.6M 增加到 $2,249.1M（+$958.5M），融资租赁负债从 $274.4M 涨到 $608.9M（+122%，其中仅 Q2 就 +60%）。 这家公司用约 $400M 的经营现金流增加了约 $10 亿的资产，而且全程报告为正的自由现金流利润率 —— 因为它自定义的调整后自由现金流明确排除了通过融资安排和融资租赁取得的设备。 CFO 原话：「我们通过为设备融资，让现金流出与收入紧密匹配。 」**这就是 [[pattern-ai-build-inflates-earnings-while-destroying-fcf]] —— 而 AKAM 把同样的支出放在资本开支这一行、让任何筛选器都能看见，它的披露更难看，也更诚实。 **
🚩 对上 $1.18B 的收入，已承诺的容量高达 155MW，而且公司从未披露过任何 GPU 利用率数字。 🚩 客户集中度正在翻倍：$1M+ 客户群占 ARR 的比重从 9% 升到 23%，前 25 大客户占 20% —— 而管理层恰恰在净收入留存率创下三年新高的这个季度，宣布不再把它列为重点指标。 这本身不能证明出了什么问题；它只是拿走了那个能让你看出问题的仪表。 🚩 毛利率从 59.9% 降到约 55% —— AI 带来的增长在结构上是更低毛利的。 🚩 ROIC 9.8%，低于 11-12% 的加权平均资本成本（贝塔 1.59），而投入资本同比增长 46%。 🚩 没有成本优势 —— 同等配置下 Hetzner 便宜 3-5 倍；护城河是一个积累了十年的教程/SEO 流量入口，而智能体自动部署恰恰就是让它贬值的东西 —— 推动收入的那波 AI 浪潮，正在侵蚀它的获客渠道。 护城河评分 4.0。 🚩 是稀释，不是回购：八个月内股本从 91.9M 增加到 116.9M；回购实际上已经停止（2022-23 年的 $600M/$488M → $60M/$82M）。 🚩 内部人：零公开市场买入，卖出集中在高点 —— CFO 以 $152.50 卖出 25,000 股、以 $170.07 卖出 10,000 股，一位董事以 $147.62 当日行权即卖出，另有 Access Industries 在 $148-162 区间套现约 $557M。 🚩 这轮 +295% 涨幅的五条腿里，有两条是机械性的 —— 6/29 纳入罗素 1000 指数，以及一个占流通股 16.35% 的空头底仓。
❌ 供应商给出的估值倍数在三处失效：PE(ttm) 60.94 用的是被约 $120M 递延所得税估值备抵转回抬高的每股收益（FY25 税前约 $139M，而净利润 $259.3M —— [[pattern-net-margin-above-operating-margin-is-a-tripwire]] 加 [[pitfall-tax-valuation-allowance-round-trip-breaks-eps]]）；按正常税率还原后的历史市盈率约为 129x，不是 60.9x；而 PE(fwd) 72.40 高于历史市盈率，正是 [[pitfall-forward-pe-above-trailing-pe-flags-an-inflated-base]] 在触发。 格雷厄姆内在价值（$20.95，还原后 $14.40）不是目标价 —— 它只说明**在这个价位附近根本不存在任何支撑底。 **
⚖️ 真实的倍数：13.4x 企业价值/FY26 收入 34.5x 企业价值/调整后 EBITDA 约 112x 企业价值/调整后自由现金流（0.9% 的自由现金流收益率）· 约 97x 前瞻非 GAAP 市盈率。 🔍 反向 DCF：$15.8B 的企业价值需要 $1.2-1.6B 的稳态自由现金流，约等于 $6-8B 的收入，是今天的 5-7 倍，也就是要在与 Hetzner、Cloudflare 和三家超大规模云厂商的竞争中连续七年以约 26% 复合增长。
合理价值 $70-165（中枢 $105）—— 区间刻意留宽，因为一家增速 30-50%、护城河只有 4/10 的公司，其结果分布本来就是这么散。 现价 $134 落在基准情形和乐观情形之间，比中枢高约 28%。 买入区间 $75-90。
✂️ 减持 $150 —— = 12x 企业价值/未来十二个月收入（未来十二个月约 $1.45B）。 采用美元形式是刻意的，原因是机制性的：网站是用雅虎的前瞻市盈率去反推倍数对应的美元价位，而那个隐含的每股收益 $1.85 比公司自己 $1.35-1.40 的非 GAAP 指引高出约 34%，而它本来会用的历史每股收益又带着那笔税务转回。 两个盈利分母都是坏的；收入是唯一干净的那个。 按 FY27 的指引收入，这个价位会自动上移到约 $180 —— 请在每次 /analyze 时重新审定这个 12x 的判断。 税负门槛约为 (57÷107)×15 ≈ 8 个百分点的预期下跌才值得提前了结；合理价值中枢比现价低 22%，门槛已经跨过 —— 但在 $107 的规模上，这是一个关于卖出纪律的决定，不是关于钱的决定。 *继续持有这个零头，不要加仓，让 $150 这条规则去工作。 **
⚠️ 集群备注：DOCN 通过了本板块的检验标准（数据中心建设放缓会伤到它 —— 已承诺 155MW），但它和 AVGO/KLAC/AMAT 处在这笔交易的相反一侧 —— 它是在买入那些公司卖出的产能。 它的失效模式是 AI 需求和资金成本，不是资本开支周期。 **它真正的同类是 AKAM。 **
Historical results历史业绩 past fiscal years — these values do not change过去财年——这些数值不会变动
Profit margins利润率 the share of revenue that stays as profit营收中留存为利润的比例
of every $1 of revenue每 $1 营收中的占比
- Return on equity净资产收益率
- 62%
- Revenue growth营收增长
- +29%
Competitors竞争对手
Why customers stay客户为什么留下
DigitalOcean serves small teams that the big clouds do not focus on. It gives simple prices and easy tools. This keeps developers loyal and lowers support costs. The revenue also spreads across many regions, so no single market controls the business.DigitalOcean 服务于大型云厂商不重点关注的小型团队。它提供简单的价格和易用的工具。这使开发者保持忠诚并降低支持成本。收入还分布在多个地区，因此没有单一市场主导业务。
Related companies相关公司 🔴 AI capex — BUYERS of the capacity (split out 8/12)
Separated from the sellers because they sit on the opposite side of the same trade. DOCN, AKAM and ORCL buy the capacity AVGO/KLAC/AMAT sell.
Do not read a low reported capex number as low capital intensity — check finance leases and the adjusted-FCF definition first. DOCN's own metric excludes lease-financed equipment; AKAM's does not.