业务分析 · ORCL / NYSE · 自下而上单位经济
Oracle 算力租赁：556.6 亿美元 CapEx 到底买了什么，能把 P/E 压到几倍？
先把机房、供电、GPU、网络、租金和融资一层层算清，再讨论 FY2030 的 6–7 倍是否可信。
Business analysis · ORCL / NYSE · Bottom-up unit economics
Oracle AI infrastructure: what did $55.7B of CapEx buy, and how far can it compress P/E?
We bridge data halls, power, GPUs, networking, contract pricing and financing before judging whether a 6–7x FY2030 P/E is credible.
1. 556.6 亿美元不是都拿去盖楼
Oracle FY2026 报告资本开支 556.63 亿美元；扣除 45.92 亿美元具有重大融资成分的客户预付款、33.45 亿美元制造商付款时点融资后，公司定义的净现金支出为 477.26 亿美元。官方披露 SEC filed earnings exhibit
181 个数据中心地点几乎全部租赁。FY2026 新取得的经营及融资租赁使用权资产约 231.92 亿美元，另外还有 2600 亿美元尚未起租的数据中心租赁承诺。因此，出租方承担了大量楼壳建设，Oracle 的现金 CapEx 更偏向 GPU、服务器、网络、存储，以及租赁机房内部的配电和冷却。官方披露 FY2026 10-K
1. The $55.7B was not all spent on buildings
Oracle reported $55.663B of FY2026 capital expenditure. After $4.592B of customer prepayments with a significant financing component and $3.345B of manufacturing-payment timing finance, company-defined net cash outlay was $47.726B.Source-reported SEC filed earnings exhibit
Substantially all of Oracle's 181 data-center locations are leased. New operating- and finance-lease right-of-use assets were about $23.192B in FY2026, while uncommenced data-center lease commitments reached $260B. Lessors therefore finance much of the shell; Oracle cash CapEx skews toward GPUs, servers, networking, storage and tenant fit-out.Source-reported FY2026 10-K
会计下界
$29.3Bcomputer / network / machinery 账面原值净增加；不含混在 CIP 中的设备。
分析基准
$41.7B现金 CapEx 的 75%，对应大量楼壳由租赁方融资。
极端上界
$52.8B把全部 CIP 增量都视为服务器/网络；实际 CIP 也含 leasehold。
2. 两种算法都指向约 50 万张 GPU 等价
算法 A：拿设备投入除以完整算力栈成本
不能用 3–4 万美元的裸 B200 报价直接除 CapEx，因为还缺服务器 CPU、NVSwitch、InfiniBand、内存、存储、光模块、安装和备件。以 Stargate 全栈投入与 Vantage 机房开发成本之差作锚，完整 compute stack 约为每张 6.5–9.0 万美元，基准约 8.3 万美元。分析假设 Vantage Lighthouse campus
算法 B：拿已交付功率乘每 MW GPU 密度
NVIDIA GB200 NVL72 约 120kW/72 卡，GB300 进一步提高机架功率；把网络、存储、冗余和 PUE 算进去，大模型集群实际约 350–500 张/MW。OpenAI 的“5GW 超过 200 万颗芯片”直接对应约 400 张/MW。Oracle 则称 FY2026 已交付超过 1.2GW。官方 + 推导 NVIDIA GB200 guide OpenAI chip-density disclosure Oracle CEO commentary
结论：约 50 万张 GPU 等价，合理区间 45–60 万。它不是 Oracle 披露的显卡数量，也不代表全部是同一代 NVIDIA GPU；它是把不同代际 GPU、网络和存储折算成当前 Blackwell 级别算力栈的容量指标。
2. Two independent methods converge near 500,000 GPU equivalents
Method A: compute-equipment investment divided by installed stack cost
Dividing CapEx by a $30k–$40k bare B200 quote is wrong because it omits CPUs, NVSwitch, InfiniBand, memory, storage, optics, deployment and spares. Triangulating Stargate's full-stack investment against Vantage's campus cost suggests $65k–$90k per installed compute-stack equivalent, with an $83k base.Analyst assumption Vantage Lighthouse campus
Method B: delivered power multiplied by practical GPU density
An NVIDIA GB200 NVL72 rack is about 120kW for 72 GPUs, and GB300 raises rack power further. Adding networking, storage, redundancy and PUE yields roughly 350–500 GPU equivalents per MW. OpenAI's “over two million chips across 5GW” directly implies about 400 per MW; Oracle says it delivered more than 1.2GW in FY2026.Reported + derived NVIDIA GB200 guide OpenAI chip-density disclosure Oracle CEO commentary
Conclusion: roughly 500,000 GPU equivalents, with a 450,000–600,000 range. Oracle does not disclose the card count; this is a current-Blackwell-equivalent capacity measure across mixed GPU generations, networking and storage.
3. 大单成交价约 3.8 美元/GPU小时，不是目录价 10–18 美元
OCI 公布的按需目录价大约为 H100/H200 每卡时 10 美元、B200 14 美元、GB200 16 美元、GB300 18 美元；但超大客户会获得长约折扣。AWS H100 Capacity Block 与 Google 长约/弹性价格大约在 4.3–4.9 美元区间。公开价格 OCI price list AWS Capacity Blocks Google Cloud pricing
OpenAI 披露与 Oracle 的合作“超过 3000 亿美元、五年、最高 4.5GW”。把它机械均摊，得到每 GW 每年约 133 亿美元；再用约 40 万张/GW，反推约 3.81 美元/GPU小时。公司口径 + 推导 OpenAI 4.5GW disclosure
3. Mega-contract realization is about $3.8/GPU-hour, not the $10–$18 list price
OCI's published on-demand list prices are roughly $10/GPU-hour for H100/H200, $14 for B200, $16 for GB200 and $18 for GB300. Mega-customers receive long-duration discounts. AWS H100 Capacity Blocks and Google committed/flex pricing sit around $4.3–$4.9.Public pricing OCI price list AWS Capacity Blocks Google Cloud pricing
OpenAI describes its Oracle partnership as more than $300B over five years and up to 4.5GW. A mechanical allocation yields about $13.3B of annual revenue per GW; at roughly 400,000 GPU equivalents per GW, that implies about $3.81/GPU-hour.Issuer claim + derived OpenAI 4.5GW disclosure
| 情景 | GPU/MW | 实现价格 | 利用率 | 每 GW 年收入 | 毛利率 |
|---|---|---|---|---|---|
| 保守 | 350 | $3.30/h | 90% | $9.1B | 30% |
| 基准 | 400 | $3.805/h | 97.5% | $13.0B | 35% |
| 乐观 | 500 | $4.50/h | 98% | $19.3B | 40% |
基准单卡年电费约 1840 美元，仅占约 3.25 万美元收入的 5.7%；真正的大头是服务器折旧、机房租赁、网络/存储和融资，而不是电费。电价采用美国工业电价参考。U.S. EIA
Base electricity cost is about $1,840 per GPU-year, only 5.7% of roughly $32,500 of revenue. Server depreciation, data-center leases, network/storage and financing matter much more than power. The electricity input uses the U.S. industrial-price reference. U.S. EIA
4. 算力收入真正能把 P/E 压到哪里
Oracle 给出的长期 AI infrastructure 非 GAAP 毛利率是 30%–40%，代表性合同为 35%。该毛利率已经把折旧、机房租金、电力和直接运营成本放在毛利之上；我们再扣除 6% 增量研发/销售/管理费用、每 GW 14 亿美元归属融资成本和 17% 税率。管理层目标 Oracle AI analyst-day deck
4. How far can compute revenue actually compress P/E?
Oracle targets a long-run 30%–40% non-GAAP AI-infrastructure gross margin, with 35% in its representative contract. That gross margin already sits after depreciation, data-center lease cost, power and direct operations. We then deduct 6% incremental R&D/S&M/G&A, $1.4B of attributable financing cost per GW and a 17% tax rate.Management target Oracle AI analyst-day deck
| 规模情景 | AI 容量 | AI 年收入 | AI 增量 EPS | 总调整后 EPS | 静态 P/E @ $143.81 |
|---|---|---|---|---|---|
| FY26 已交付块满产上限 | 1.2GW | $15.6B | +$0.68 | $7.51 | 19.2x |
| OpenAI 4.5GW 全部投产 | 4.5GW | $58.5B | +$2.77 | $10.57 | 13.6x |
| 本站 FY30 独立基准 | 7.5GW | $97.5B | +$4.61 | $12.41 | 11.6x |
| 接近管理层目标的资本高效情景 | 11.5GW | $172.0B | +$12.04 | $20.54 | 7.0x |
最关键的反推：若采用本站基准单位经济，达到管理层 FY2030 的 21 美元调整后 EPS，大约需要 21.5GW；若价格、毛利和客户供资都接近乐观情景，则约需 11.9GW。后者相当于 450–500 万张 GPU 等价，是公开的 Oracle–OpenAI 4.5GW 规模的约 2.6 倍。管理层目标因此隐含了其他大客户、客户供资硬件，以及传统数据库/应用利润继续增长。FY2028–FY2030 的 21 美元仍是长期目标，不是已兑现利润。Oracle financial targets
11.5GW 预设产生约 1720 亿美元 AI 算力收入，已经略高于管理层 FY2030 的 1660 亿美元 OCI 总收入目标；它的用途是显示“把 EPS 做到约 21 美元所需的资本效率”，而不是与管理层分部口径完全一致的收入预测。
The key backsolve: at our base unit economics, reaching management's $21 adjusted FY2030 EPS needs roughly 21.5GW. If pricing, margin and customer funding all approach the high case, the requirement falls to about 11.9GW—roughly 4.5–5.0 million GPU equivalents and about 2.6 times the disclosed 4.5GW Oracle–OpenAI program. The target therefore embeds other mega-customers, customer-funded hardware and continued database/application profit growth. The $21 remains a long-term target, not realized earnings. Oracle financial targets
The 11.5GW preset produces about $172B of AI-compute revenue, already slightly above management's $166B FY2030 total OCI revenue target. It is a capital-efficiency stress test for reaching roughly $21 of EPS, not a segment forecast that reconciles exactly to management's presentation.
FY26 1.2GW 已有一部分收入包含在当前 6.83 美元正常化 EPS 中，因此把 +0.68 美元全部再加一次是偏乐观上限，不是正式预测。4.5GW、7.5GW 和 11.5GW 情景使用不同的核心业务 EPS 与融资负担，具体见下方可调模型。
Some revenue from the FY2026 1.2GW is already in the current $6.83 normalized EPS, so adding the full $0.68 again is an optimistic ceiling, not a formal forecast. The 4.5GW, 7.5GW and 11.5GW cases use different core-EPS and funding assumptions in the calculator below.
5. 可调单位经济模型
先用资本开支计算可购买与可上电 GPU，再用总 GW 规模计算收入、净利润、EPS 和 P/E。所有紫色标签均为本站假设，可自行调整。
A. CapEx → GPU
部署数取“买得到”和“电力装得下”的较小值。
B. GW → 收入 → EPS → P/E
点击预设，或移动滑块检查自己的假设。
6. 这项研究最容易错在哪里
- GPU 数量没有披露。“约 50 万张”是设备投入、功率密度和已交付 GW 的三角验证，不是 Oracle 官方数字。
- GW 口径可能不同。园区总电力、关键 IT 负载和已安装 GPU 的可用功率不是一回事；模型用 OpenAI 芯片密度作实务校准。
- 3000 亿美元不是已确认收入。OpenAI 称合作超过 3000 亿美元且最高 4.5GW，Oracle 没有公开客户级收入确认表、取消条款和逐站开机进度。
- 客户供资决定 EPS。750 亿美元“预付 + 客户自带硬件部分”没有拆分；真正的 FY2026 重大融资成分预付款只有 45.92 亿美元。客户自带 GPU 会显著提高 Oracle 的资本回报。
- P/E 不是现金流。FY2026 自由现金流为负 236.86 亿美元；折旧、租赁和利息的爬坡可能远落后或领先于收入确认。
证伪阈值：若每 GW 实现收入长期低于 100 亿美元、AI 毛利率低于 30%、或归属融资成本高于约 20 亿美元/GW，则独立基准 EPS 会明显低于 12 美元，当前股价对应的 FY2030 P/E 更接近 14–18 倍。反之，只有当总交付规模接近 12GW、实现价不低于约 4 美元/GPU小时、毛利率接近 40%，且大量 GPU 由客户供资时，7 倍才变得可解释。
6. Where this analysis can be wrong
- GPU count is undisclosed. Roughly 500,000 is a triangulation of equipment spend, power density and delivered GW—not an Oracle-reported figure.
- GW definitions vary. Campus utility capacity, critical IT load and GPU-ready power are not interchangeable; the model uses OpenAI's chip density as a practical calibration.
- $300B is not recognized revenue. OpenAI describes more than $300B and up to 4.5GW, but Oracle has not published customer-level recognition, cancellation terms or site-by-site commissioning.
- Customer funding determines EPS. The $75B “prepaid plus customer-supplied hardware portion” is not split; actual FY2026 prepayments with significant financing components were only $4.592B. BYOH materially improves Oracle's return on capital.
- P/E is not cash flow. FY2026 free cash flow was negative $23.686B; depreciation, lease and interest ramps can lead or lag revenue recognition.
Falsification threshold: if realized revenue stays below $10B/GW-year, AI gross margin falls below 30%, or attributable financing exceeds roughly $2B/GW, independent-base EPS falls materially below $12 and FY2030 P/E at today's price is closer to 14–18x. A 7x outcome becomes explainable only near 12GW, at least roughly $4/GPU-hour, close to 40% gross margin and substantial customer-funded hardware.
主要来源
- Oracle investor-relations stock informationORCL security identity and market-price page; price snapshot frozen 2026-08-20
- Oracle FY2026 10-Kcapex, PP&E, leases, debt, depreciation, RPO
- Oracle FY2026 Q4 resultsOCI revenue, FY27 guidance, financing, customer-funded hardware
- Oracle FY2026 Q4 filed exhibitreported capex and net cash outlay bridge
- Oracle 2025 AI infrastructure analyst-day deck30-40% AI gross margin and representative contract economics
- Oracle 2025 financial analyst-day deckFY2028-FY2030 long-term revenue and EPS targets
- Oracle CEO Q4 commentarydelivered GW, contracted capacity and GPU utilization
- OpenAI: Oracle Stargate 4.5GW partnershipover $300B over five years and up to 4.5GW
- OpenAI: Stargate chip-density disclosureover two million chips across 5GW
- NVIDIA DGX GB200 NVL72 hardware guide72 GPUs and approximately 120kW per rack
- NVIDIA GB300 NVL72 reference architectureGB300 rack power and cluster components
- Vantage Lighthouse campus$15B and 902MW data-center development anchor
- OCI global price listpublic GPU list prices
- AWS EC2 Capacity Blocks pricingcommitted H100 market price anchor
- Google Cloud accelerator-optimized pricingH100/H200/B200 committed and flex prices
- U.S. EIA electricity pricesU.S. industrial electricity-price reference