NVIDIA Vera Rubin NVL72 產業分析
NVIDIA Vera Rubin NVL72 Industry Analysis

以每 MW 有效 Token 產量 實際 GPU 利用率 模型遷移成本與五年總持有成本為核心指標
A decision framework centered on effective tokens per MW, effective GPU utilization, model migration cost, and five year total cost of ownership.

管理層決策報告 版本日期 2026 年 10 月 12 日
Executive decision report  Version date 12 October 2026

圖 1 Vera Rubin NVL72 整機櫃 3D 爆炸圖與供應鏈示意  Figure 1 Vera Rubin NVL72 rack scale 3D exploded view and supply chain illustration

核心結論：Rubin 的峰值算力與單櫃價格不能單獨支持投資決策。採購應以相同模型 相同品質門檻與完整設施用電，驗證每 MW 可交付的合格 AI 工作量與五年單位經濟性。
Core conclusion: Rubin peak performance and rack price alone cannot justify an investment. Procurement should validate qualified AI work per MW and five year unit economics using the same models, quality thresholds, and full facility power measurements.

執行摘要
Executive Summary

Vera Rubin NVL72 是 NVIDIA 第三代機櫃級 AI 架構，整合 72 顆 Rubin GPU 36 顆 Vera CPU 20.7 TB HBM4 NVLink 6 800G 網路 液冷與 AI 軟體。其投資價值在於電力受限資料中心能否提升有效 Token 與 Agent 任務完成量。
Vera Rubin NVL72 is NVIDIA third generation rack scale AI architecture, integrating 72 Rubin GPUs, 36 Vera CPUs, 20.7 TB of HBM4, NVLink 6, 800G networking, liquid cooling, and AI software. Its investment value depends on whether a power constrained data center can increase effective tokens and completed agent tasks.

建議決策門檻
Recommended Decision Gates

核心指標 Core KPI	最低建議門檻 Recommended minimum	採購判斷 Procurement test
每 MW 有效 Token 產量 Effective tokens per MW	較現有平台提升至少 2 倍 At least 2x versus current platform	以客戶模型與完整機房電表實測 Measure customer models with facility meters
有效 GPU 利用率 Effective GPU utilization	商用穩態 60% 以上 At least 60% in steady production	排除等待 重試與不合格輸出 Exclude waiting, retries, and rejected output
模型遷移成本 Model migration cost	低於五年 TCO 的 15% Below 15% of five year TCO	包含雙軌運行與品質回歸 Include dual run and quality regression
五年 TCO Five year TCO	每百萬有效 Token 降低 30% 30% lower cost per million effective tokens	採基準 樂觀 壓力三情境 Use base, upside, and stress cases
投資回收期 Payback period	24 至 30 個月內 Within 24 to 30 months	依算力銷售收入或內部節省計算 Use compute revenue or internal savings

 

管理層結論
Management Conclusions

• Rubin 適合電力緊張 且 Agentic AI 長上下文 MoE 推論量充足的 AI Factory。
  Rubin is best suited to power constrained AI factories with substantial long context, MoE, and agentic workloads.
• 有效 GPU 利用率低於 40% 時，高峰值效能通常無法抵銷閒置資本。
  Below 40% effective GPU utilization, peak performance usually cannot offset idle capital.
• 所有效能與成本聲明必須由 30 至 90 天 POC 與三相電表重新驗證。
  All performance and cost claims should be revalidated through a 30 to 90 day POC and three phase facility metering.

一 產品定位與技術規格
1 Product Positioning and Technical Specifications

NVL72 面向前沿模型訓練 後訓練 Test time scaling Agentic AI 長上下文推論 物理 AI 模擬與主權 AI。72 顆 GPU 形成單一 NVLink 域，再以 800 Gb/s InfiniBand 或 Ethernet 擴展至多機櫃。
NVL72 targets frontier model training, post training, test time scaling, agentic AI, long context inference, physical AI simulation, and sovereign AI. Its 72 GPUs operate as one NVLink domain and scale across racks through 800 Gb/s InfiniBand or Ethernet.

項目 Item	Vera Rubin NVL72 Vera Rubin NVL72	決策意義 Decision relevance
GPU 與 CPU GPU and CPU	72 Rubin GPUs 與 36 Vera CPUs 72 Rubin GPUs and 36 Vera CPUs	完整 CPU GPU 協同運算域 Integrated CPU GPU compute domain
GPU 記憶體 GPU memory	20.7 TB HBM4 20.7 TB HBM4	支援大型 MoE 與長 KV Cache Supports large MoE models and long KV cache
HBM 頻寬 HBM bandwidth	約 1,400 至 1,580 TB/s About 1,400 to 1,580 TB/s	官方頁面仍有版本差異 Official pages still show version differences
推論算力 Inference compute	3,600 PFLOPS NVFP4 3,600 PFLOPS NVFP4	峰值不能取代模型級實測 Peak value does not replace model benchmarks
訓練算力 Training compute	2,520 PFLOPS NVFP4 2,520 PFLOPS NVFP4	受精度策略與通訊效率影響 Depends on precision and communication efficiency
NVLink 6 NVLink 6	216 TB/s 機櫃總頻寬 216 TB/s aggregate rack bandwidth	降低跨 GPU 專家與張量通訊瓶頸 Reduces expert and tensor communication bottlenecks
機櫃與冷卻 Rack and cooling	48 RU 100% 液冷 入口水溫最高 45 C 48 RU, fully liquid cooled, inlet water up to 45 C	既有風冷機房通常需要改造 Most air cooled facilities require upgrades

 

整機櫃模組
Rack Scale Building Blocks

1. 18 組 1RU 運算托盤，每組 2 顆 Vera CPU 與 4 顆 Rubin GPU。
   18 one RU compute trays, each with two Vera CPUs and four Rubin GPUs.
2. 9 組 NVLink 6 交換器托盤，形成 72 GPU 單一 NVLink 域。
   Nine NVLink 6 switch trays create one 72 GPU NVLink domain.
3. 144 個以上 ConnectX 9 800G 介面與 18 個以上 BlueField 4 DPU。
   More than 144 ConnectX 9 800G interfaces and at least 18 BlueField 4 DPUs.
4. 4 組 3RU Power Shelf 供應 54 V DC 液冷 Busbar。
   Four 3RU power shelves supply a 54 V DC liquid cooled busbar.

二 四項核心投資指標
2 Four Core Investment Metrics

四項指標必須在同一工作負載 電價 PUE 品質標準與時間窗內計算，否則不同平台的比較沒有可比性。
All four metrics must use the same workloads, electricity price, PUE, quality standard, and measurement window; otherwise platform comparisons are not valid.

指標 Metric	回答的問題 Decision question	主要資料來源 Primary evidence	權重 Weight
每 MW 有效 Token 產量 Effective tokens per MW	同樣電力可交付多少合格 AI 工作量 How much qualified AI work can the same power deliver	電表 推論服務 品質評測 SLA Facility meters, serving logs, quality tests, SLA	35% 35%
實際 GPU 利用率 Effective GPU utilization	已購 GPU 有多少時間產生有效工作 How much purchased GPU time creates useful work	Telemetry 排程器 Profiler Telemetry, scheduler, profiler	25% 25%
模型遷移成本 Model migration cost	平台切換需要多少資源與風險 What resources and risks platform switching requires	人月 雙軌費用 驗證與停機 Labor months, dual run, validation, downtime	15% 15%
五年總持有成本 Five year TCO	每單位有效產出的完整五年成本 Full five year cost per unit of effective output	CAPEX OPEX 融資 殘值 CAPEX, OPEX, financing, residual value	25% 25%

 

統一決策公式
Unified Decision Formula

最終比較應使用每百萬有效 Token 或每項合格 Agent 任務的五年全成本，而不是單櫃採購價格。
The final comparison should use the five year cost per million effective tokens or per qualified agent task, not the purchase price of a rack.

三 每 MW 有效 Token 產量
3 Effective Token Output per MW

分母必須採資料中心實際電表讀值，納入 GPU CPU HBM 網路 儲存 CDU 水泵 UPS 與配電損耗，不能只採 GPU TDP。
The denominator must use actual facility meter readings and include GPU, CPU, HBM, network, storage, CDU, pumps, UPS, and power distribution losses. GPU TDP alone is insufficient.

符號 Symbol	定義 Definition	量測方式 Measurement
T raw T raw	原始輸入與輸出 Token 或任務產量 Raw input and output tokens or task throughput	服務端計量與模型日誌 Serving metrics and model logs
Q Q	品質係數 Quality factor	正確率 任務完成率 人工或自動評測 Accuracy, task completion, human or automated evaluation
S S	SLA 係數 SLA factor	TTFT P95 P99 延遲與逾時率 TTFT, P95, P99 latency, and timeout rate
A A	客戶接受率 Customer acceptance factor	排除重試 錯誤與無效輸出 Exclude retries, errors, and invalid output
P facility P facility	機房端實際用電 MW Measured facility power in MW	PDU UPS 與冷卻電表 PDU, UPS, and cooling meters

 

示例與 Agentic AI 修正
Example and Agentic AI Adjustment

假設每小時產出 100 億 Token，品質係數 95%，SLA 90%，接受率 98%，則有效 Token 為 83.79 億。若機房負載 0.216 MW，每小時每 MW 有效產量約 388 億 Token。此數字僅示範計算方法。
If raw output is 10 billion tokens per hour, quality is 95%, SLA attainment is 90%, and acceptance is 98%, effective output is 8.379 billion tokens. At 0.216 MW facility load, the result is about 38.8 billion effective tokens per MW hour. This is an illustrative calculation only.

• Prefill 與 Decode 分開量測，保留輸入輸出長度分布。
  Measure prefill and decode separately and retain input and output length distributions.
• 同時報告每 MW 完成任務數 工具呼叫成功率與每項任務成本。
  Also report completed tasks per MW, tool call success, and cost per completed task.
• 以相同品質標準比較 Rubin Blackwell AMD Helios 與雲端平台。
  Compare Rubin, Blackwell, AMD Helios, and cloud platforms under the same quality standard.

四 實際 GPU 利用率
4 Effective GPU Utilization

nvidia smi 的 GPU 繁忙率不等於商業有效利用率。等待記憶體與網路 重試 不合格 Batch 或品質不合格輸出都可能讓表面利用率偏高。
GPU busy reported by nvidia smi is not the same as commercially effective utilization. Memory and network stalls, retries, rejected batches, and low quality output can all inflate apparent utilization.

構成 Component	示例值 Example	說明 Meaning
硬體可用率 Hardware availability	98% 98%	排除故障與維護降級 Excludes faults and maintenance derating
排程使用率 Scheduled utilization	85% 85%	GPU 已分配給有效工作 GPU is assigned to valid work
實際運算效率 Compute efficiency	70% 70%	相對可達算力 排除等待與瓶頸 Relative achieved compute after stalls and bottlenecks
有效工作接受率 Accepted work rate	95% 95%	排除錯誤 重試及品質不合格 Excludes errors, retries, and rejected quality
有效 GPU 利用率 Effective GPU utilization	55.4% 55.4%	98% × 85% × 70% × 95% 98% x 85% x 70% x 95%

 

建議營運門檻
Recommended Operating Thresholds

營運階段 Operating stage	有效 GPU 利用率 Effective utilization	判斷 Interpretation
導入與測試期 Implementation and test	35% 至 50% 35% to 50%	允許模型優化與排程調整 Allows optimization and scheduling changes
正常商用期 Steady production	55% 至 70% 55% to 70%	具有基本資本效率 Basic capital efficiency
成熟大規模環境 Mature scaled environment	70% 至 80% 70% to 80%	需要成熟平台與穩定需求 Requires mature platform and stable demand
長期高於 85% Sustained above 85%	需審慎 Investigate	可能缺乏維護與尖峰容量 May lack maintenance and peak headroom

 

• 同步監控 Tensor Core HBM NVLink Scale out 網路與 KV Cache。
  Monitor Tensor Core, HBM, NVLink, scale out networking, and KV cache together.
• 將排程等待 重試 Checkpoint 與故障降級納入有效利用率。
  Include scheduling waits, retries, checkpoints, and fault derating in effective utilization.

五 模型遷移成本
5 Model Migration Cost

模型從 H100 H200 B200 或 GB300 遷移至 Rubin，通常涉及 CUDA 框架 Kernel 量化 分片 推論服務與營運流程的共同調整。
Moving models from H100, H200, B200, or GB300 to Rubin usually requires coordinated changes to the CUDA stack, kernels, quantization, sharding, serving, and operating processes.

成本項目 Cost item	範圍 Scope	估算基礎 Estimation basis
軟體移植 Software porting	CUDA 驅動 容器 PyTorch Triton TensorRT LLM CUDA, drivers, containers, PyTorch, Triton, TensorRT LLM	工程人月與測試環境 Engineering months and test environment
Kernel 優化 Kernel optimization	NVFP4 FP8 Attention MoE 稀疏運算 NVFP4, FP8, attention, MoE, sparse compute	效能工程人月 Performance engineering months
模型轉換 Model conversion	權重格式 量化 Checkpoint 與分片 Weights, quantization, checkpoints, sharding	模型數量與參數規模 Model count and parameter scale
推論重構 Serving redesign	Prefill Decode 分離 KV Cache 與路由 Prefill decode separation, KV cache, routing	服務架構複雜度 Serving architecture complexity
驗證 Validation	準確率 安全 偏差 回歸 合規 Accuracy, safety, bias, regression, compliance	測試集與審查週期 Test suites and review cycles
雙軌與停機 Dual run and downtime	Blackwell 與 Rubin 並行 流量切換 回退 Parallel Blackwell and Rubin, traffic cutover, rollback	4 至 12 週算力與營運損失 Four to twelve weeks of compute and operational impact

 

控制門檻
Control Gates

• 遷移成本控制在五年 TCO 的 10% 至 15%。
  Keep migration cost within 10% to 15% of five year TCO.
• 核心服務雙軌運行 4 至 12 週，保留可回退版本。
  Run critical services in parallel for four to twelve weeks and retain rollback versions.
• 量化後重新驗證品質 安全 合規與 SLA。
  Revalidate quality, safety, compliance, and SLA after quantization.
• 若回收期超過 18 個月，應重新檢視平台切換範圍。
  Reassess migration scope if migration payback exceeds 18 months.

六 五年總持有成本
6 Five Year Total Cost of Ownership

五年 TCO 應覆蓋設備 機房改造 網路與儲存 電力與冷卻 軟體 維保 人力 融資 停機 遷移與殘值。
Five year TCO should cover equipment, facility upgrades, network and storage, power and cooling, software, support, labor, financing, downtime, migration, and residual value.

基準情境試算
Base Case Illustration

成本項目 Cost item	五年金額 US$M Five year amount US$M	口徑 Basis
NVL72 設備 NVL72 equipment	7.803 7.803	依外部成本模型 非正式報價 External cost model, not an official quote
機房改造分攤 Facility retrofit allocation	1.500 1.500	配電 UPS 液冷 管路與消防 Power distribution, UPS, liquid cooling, piping, fire protection
網路與儲存 Network and storage	1.200 1.200	Scale out 交換器 光模組與共享儲存 Scale out switches, optics, shared storage
五年電力與冷卻 Five year power and cooling	1.135 1.135	平均 IT 180 kW PUE 1.20 電價 0.12 美元 Average IT 180 kW, PUE 1.20, power at US$0.12 per kWh
軟體與企業支援 Software and support	0.900 0.900	授權 管理平台與五年支援 Licenses, management platform, five year support
維保備品 Maintenance and spares	0.600 0.600	PSU 泵浦 快接頭與零件 PSU, pumps, quick disconnects, parts
維運人力 Operations labor	0.750 0.750	平台 網路 設施與資安分攤 Platform, network, facility, and security allocation
融資 保險 稅費 Finance insurance taxes	0.700 0.700	依資本結構估算 Estimated by capital structure
模型遷移 Model migration	0.800 0.800	工程 雙軌及驗證 Engineering, dual run, validation
停機與風險準備 Downtime and contingency	0.400 0.400	非預期中斷與更換 Unplanned interruption and replacement
第五年殘值 Year five residual value	-1.000 -1.000	出售或降階使用 Resale or lower tier reuse
五年 TCO 合計 Total five year TCO	13.788 13.788	示範性基準情境 Illustrative base case

 

電力敏感度
Power Cost Sensitivity

情境 Scenario	五年 TCO US$M Five year TCO US$M	主要假設 Main assumptions
樂觀 Upside	12.3 12.3	高利用率 低電價 快速遷移 良好殘值 High utilization, low electricity price, fast migration, strong residual value
基準 Base	13.8 13.8	180 kW PUE 1.20 標準維保與遷移 180 kW, PUE 1.20, standard support and migration
壓力 Stress	16.5 16.5	記憶體與電價上升 延遲部署 低殘值 Higher memory and power prices, delayed deployment, low residual value

 

七 單櫃成本結構與價格判讀
7 Rack Cost Structure and Price Interpretation

單櫃估計 US$7,803,148 接近外部市場研究的 US$7M 至 9M 區間，但不屬於 NVIDIA 官方售價。不同報價是否含 HBM 網路 儲存 CDU 軟體與維保，可能造成數百萬美元差異。
The estimated rack cost of US$7,803,148 is close to external market research ranges of US$7 million to US$9 million, but it is not an official NVIDIA price. Whether a quote includes HBM, networking, storage, CDU, software, and support can change the total by several million dollars.

類別 Category	估算金額 US$ Estimated amount US$	比例 Share	Rubin 時代判讀 Rubin era interpretation
GPU 核心算力 GPU compute	3,960,000 3,960,000	50.7% 50.7%	仍為最大成本項 Still the largest cost item
記憶體與儲存 Memory and storage	2,001,600 2,001,600	25.6% 25.6%	HBM4 LPDDR5X 與 NVMe 價格敏感 Sensitive to HBM4, LPDDR5X, and NVMe pricing
高速通訊 High speed interconnect	720,000 720,000	9.2% 9.2%	NVLink 6 800G 與 CPO 重要性上升 NVLink 6, 800G, and CPO become more important
CPU CPU	180,000 180,000	2.3% 2.3%	36 顆 Vera CPU 36 Vera CPUs
PCB 與 ABF PCB and ABF	137,070 137,070	1.8% 1.8%	高層板與低損耗材料 High layer count and low loss materials
電源 Power	76,000 76,000	1.0% 1.0%	可能低估高密度配電改造 May understate high density power conversion
液冷 Liquid cooling	72,080 72,080	0.9% 0.9%	可能未含外部 CDU 與機房管路 May exclude external CDU and facility piping
機構與組裝 Mechanicals and assembly	656,398 656,398	8.4% 8.4%	機櫃 滑軌 組裝與測試 Rack, rails, assembly, and testing

 

八 供應鏈與台灣產業機會
8 Supply Chain and Taiwan Industry Opportunities

NVIDIA 表示 Vera Rubin 生產生態涵蓋 30 國 350 多座工廠，其中台灣約 150 家合作夥伴。正式列名 技術能力 送樣驗證與量產訂單必須分開判讀。
NVIDIA states that the Vera Rubin production ecosystem spans more than 350 factories in 30 countries, including about 150 partners in Taiwan. Official naming, technical capability, sampling, qualification, and production orders must be distinguished.

環節 Segment	代表性公司 Representative companies	機會與限制 Opportunity and constraint
晶片與平台 Chips and platform	NVIDIA NVIDIA	GPU CPU NVSwitch NIC DPU 與軟體 GPU, CPU, NVSwitch, NIC, DPU, and software
先進製程封裝 Advanced process and packaging	TSMC ASE SPIL TSMC, ASE, SPIL	先進節點與高密度封裝 供應配置待確認 Advanced nodes and high density packaging; allocation needs confirmation
HBM4 HBM4	SK hynix Samsung Micron SK hynix, Samsung, Micron	價格 良率與供應分配為核心風險 Price, yield, and allocation are core risks
系統與 ODM Systems and ODM	Dell HPE Lenovo Supermicro Foxconn QCT Wistron Wiwynn Inventec Dell, HPE, Lenovo, Supermicro, Foxconn, QCT, Wistron, Wiwynn, Inventec	多家已列入 NVIDIA 生態 Multiple vendors are named in the NVIDIA ecosystem
PCB ABF CCL PCB ABF CCL	欣興 南電 景碩 台光電 台燿 Unimicron, Nan Ya PCB, Kinsus, Elite Material, Taiwan Union Technology	高層低損耗材料需求提高 Demand rises for high layer count and low loss materials
液冷 Liquid cooling	奇鋐 雙鴻 台達電 AVC, Auras, Delta Electronics	冷板 Manifold CDU 快接與系統整合 Cold plates, manifolds, CDU, quick disconnects, system integration
電源 Power	台達電 光寶 群電 Delta Electronics, Lite On, Chicony Power	高壓 AC 54 V DC 與機櫃電源 High voltage AC, 54 V DC, and rack power
機構滑軌 Mechanicals and rails	川湖及機構廠 King Slide and mechanical suppliers	高承重 快速維護與模組化 High load capacity, serviceability, modularity

 

價值轉移方向
Value Migration

• HBM4 與先進封裝占用更多系統價值。
  HBM4 and advanced packaging capture more system value.
• 800G 1.6T CPO 與低損耗材料直接影響 GPU 利用率。
  800G, 1.6T, CPO, and low loss materials directly affect GPU utilization.
• 液冷與電源由輔助設備升級為 AI Factory 產能約束。
  Liquid cooling and power move from support systems to AI factory capacity constraints.

九 競爭格局與主要風險
9 Competitive Landscape and Key Risks

平台 Platform	核心優勢 Core advantage	主要限制 Key constraint	適用情境 Best fit
NVIDIA Vera Rubin NVIDIA Vera Rubin	CUDA NVLink 完整 AI Factory 與供應鏈 CUDA, NVLink, full AI factory stack and ecosystem	價格高 平台綁定 供應集中 High price, platform lock in, supply concentration	前沿模型 Agentic AI 多機櫃 Frontier models, agentic AI, multi rack scale
AMD Helios AMD Helios	72 GPU 開放架構 UALink ROCm 與 EPYC 整合 72 GPU open architecture, UALink, ROCm, EPYC integration	軟體成熟度與規模驗證 Software maturity and scale validation	成本競爭與開放標準 Cost competition and open standards
Google TPU Google TPU	垂直整合與大型 Pod 規模 Vertical integration and large pod scale	主要依賴 Google Cloud Primarily tied to Google Cloud	Google 生態大型訓練 Large training within Google ecosystem
AWS Trainium AWS Trainium	AWS 服務整合與雲端成本控制 AWS integration and cloud cost control	跨雲與地端可攜性有限 Limited portability across clouds and on premises	AWS 原生客戶 AWS native customers
客製 ASIC Custom ASIC	特定模型成本與能效最佳化 Optimized cost and efficiency for specific models	NRE 高 週期長 彈性低 High NRE, long schedule, low flexibility	超大規模穩定工作負載 Hyperscale stable workloads

 

主要風險與控制
Key Risks and Controls

風險 Risk	影響 Impact	控制措施 Control
HBM4 與封裝供應 HBM4 and packaging supply	延遲出貨與成本上升 Shipment delay and higher cost	鎖定配額 替代配置 價格條款 Secure allocation, alternatives, price clauses
機房電力不足 Insufficient facility power	無法上架或降頻運行 Unable to deploy or forced derating	先完成電力與 PUE 容量驗證 Validate power and PUE capacity first
液冷可靠性 Liquid cooling reliability	洩漏 高溫與停機 Leaks, overheating, downtime	水質 流量 壓差與洩漏偵測 Water quality, flow, pressure, leak detection
網路瓶頸 Network bottleneck	GPU 等待與有效利用率下降 GPU stalls and lower effective utilization	端到端 800G 壓力測試 End to end 800G stress tests
軟體成熟度 Software maturity	效能倍率無法重現 Performance claims cannot be reproduced	自有模型 POC 與版本鎖定 Own model POC and version control
出口管制 Export controls	交付與終端用途受限 Delivery and end use restrictions	BIS End User 與法遵審查 BIS, end user, and compliance review

 

十 30 至 90 天 POC 與採購驗收計畫
10 30 to 90 Day POC and Procurement Acceptance Plan

階段 Phase	期間 Timing	工作內容 Work	主要交付 Deliverable
基線 Baseline	第 1 至 2 週 Weeks 1 to 2	固定模型 資料集 SLA 電價與 PUE Freeze models, datasets, SLA, power price, PUE	現有平台基準 Current platform baseline
部署 Deployment	第 2 至 4 週 Weeks 2 to 4	Rubin 軟體 網路 儲存 液冷與監控 Rubin software, network, storage, cooling, monitoring	可重複部署映像 Repeatable deployment image
效能 Performance	第 4 至 6 週 Weeks 4 to 6	Prefill Decode MoE 長上下文與 Agent Prefill, decode, MoE, long context, agents	Token MW 與延遲曲線 Token per MW and latency curves
韌性 Resilience	第 6 至 8 週 Weeks 6 to 8	故障 降級 回復 Checkpoint 與液冷事件 Faults, derating, recovery, checkpoints, cooling events	RAS 與恢復報告 RAS and recovery report
財務 Finance	第 8 至 10 週 Weeks 8 to 10	TCO 遷移成本與營收敏感度 TCO, migration cost, revenue sensitivity	三情境財務模型 Three scenario financial model
決策 Decision	第 10 至 12 週 Weeks 10 to 12	採購 租賃 延後或混合方案 Buy, lease, defer, or hybrid	投資委員會決策包 Investment committee package

 

驗收 KPI
Acceptance KPIs

KPI KPI	驗收要求 Acceptance requirement	證據 Evidence
每 MW 有效產量 Effective output per MW	達到 POC 約定提升倍數 Achieve agreed POC improvement	設施電表與服務日誌 Facility meters and serving logs
有效 GPU 利用率 Effective GPU utilization	穩態 60% 以上或合約值 At least 60% steady state or contracted value	Telemetry 與排程紀錄 Telemetry and scheduler records
品質 Quality	不低於既有平台容忍值 No worse than agreed baseline tolerance	固定評測集與安全測試 Fixed evaluation set and safety tests
延遲 Latency	P95 P99 達標 Meet P95 and P99 targets	壓力測試與真實流量 Load tests and production traffic
可用率 Availability	符合 SLA 與維修時間 Meet SLA and repair time	故障注入與事件紀錄 Fault injection and incident records
遷移成本 Migration cost	不超過核定預算 Within approved budget	工時 雙軌與雲端帳單 Labor, dual run, and cloud invoices
五年 TCO Five year TCO	基準與壓力情境可接受 Base and stress cases acceptable	經財務審核模型 Finance reviewed model

 

十一 最終建議
11 Final Recommendations

Vera Rubin NVL72 適合已具備高密度液冷 800G 以上網路 穩定模型需求與成熟平台團隊的 AI Factory。若供電與液冷改造尚未完成，或 GPU 需求利用率不足，應先採雲端租用 NVL4 或分批部署降低資本風險。
Vera Rubin NVL72 is suitable for AI factories that already have high density liquid cooling, 800G or faster networking, stable model demand, and a mature platform team. If power and cooling upgrades are incomplete or utilization demand is insufficient, cloud rental, NVL4, or phased deployment can reduce capital risk.

建議決策順序
Recommended Decision Sequence

1. 以現有平台建立每 MW 有效 Token 與任務成本基線。
   Establish effective tokens per MW and task cost baselines on the current platform.
2. 使用相同模型 品質門檻與資料集執行 Rubin POC。
   Run a Rubin POC using the same models, quality thresholds, and datasets.
3. 把有效 GPU 利用率拆成可用率 排程率 計算效率與接受率。
   Decompose effective utilization into availability, scheduling, compute efficiency, and acceptance.
4. 完成模型遷移工作包與雙軌預算。
   Complete the model migration work package and dual run budget.
5. 建立五年 TCO 三情境並計算每百萬有效 Token 成本。
   Build three five year TCO scenarios and calculate cost per million effective tokens.
6. 只有在壓力情境回收期低於 30 個月時才進入大規模採購。
   Proceed to scale procurement only if stress case payback is below 30 months.

核心判斷
Core Decision

若 Rubin 在相同電力限制下將有效 AI 工作量提高 2 倍以上，有效 GPU 利用率穩定達 60% 至 70%，且五年每百萬有效 Token 成本降低 30% 以上，即使單櫃接近 US$8M 仍可能具有經濟合理性。反之，高峰值效能只會轉化為昂貴的閒置資產。
If Rubin delivers more than 2x effective AI work under the same power limit, sustains 60% to 70% effective GPU utilization, and reduces five year cost per million effective tokens by more than 30%, a rack price near US$8 million may still be economically justified. Otherwise, peak performance becomes expensive idle capacity.

參考資料
References

1. NVIDIA Vera Rubin NVL72 product page 2026. https://www.nvidia.com/en-in/data-center/vera-rubin-nvl72/
2. NVIDIA DGX Vera Rubin NVL72 Datasheet 2026. https://resources.nvidia.com/en-us-dgx-systems/dgx-vera-rubin-nvl72-datasheet
3. NVIDIA DGX Vera Rubin NVL72 User Guide Hardware 2026. https://docs.nvidia.com/dgx/dgxvr-user-guide/hardware.html
4. NVIDIA Newsroom Vera Rubin Ramps Into Full Production 31 May 2026. https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory
5. NVIDIA Blog Vera Rubin NVL72 Efficiency for AI Agents updated 15 September 2026. https://blogs.nvidia.com/blog/vera-rubin-nvl72-efficiency-ai-agents/
6. AMD Helios Rackscale Solution 2026. https://www.amd.com/en/products/rackscale-solutions/helios.html
7. Google Cloud TPU documentation and Next 2026 infrastructure announcements. https://cloud.google.com/tpu
8. AI 機櫃成本結構 使用者提供之外部成本模型 2026. AI rack cost structure based on a user supplied external cost model 2026.

重要聲明
Important Notice

本報告用於產業研究與投資採購評估。外部價格 供應鏈分配與效能倍數可能隨記憶體價格 軟體版本 OEM 配置及測試條件變動。正式決策應以具約束力的報價 技術附件 POC 結果及法律與財務審查為準。
This report is intended for industry research and investment or procurement evaluation. External pricing, supply allocation, and performance multiples may change with memory prices, software versions, OEM configurations, and test conditions. Final decisions should rely on binding quotations, technical schedules, POC results, and legal and financial review.