The $2 trillion in cloud capex is expected to drive nearly 20 GW of new compute capacity by 2027
AI summary card
The $2 trillion in cloud capex is expected to drive nearly 20 GW of new compute capacity by 2027
Morgan Stanley believes AI data center buildouts at AMZN, GOOGL, META, and MSFT are entering the capacity realization phase, with 2027 new capacity of about 19.5 GW, led by GOOGL, while NVDA still contributes roughly 60% of the incremental capacity.
- The report expects AMZN, GOOGL, META, and MSFT to add about 14 GW / 20 GW of capacity in 2026 / 2027, with 2027 capacity roughly 3x 2025 levels.
- GOOGL is expected to add about 7 GW in 2027, ahead of AMZN and MSFT by about 5 GW each and META by about 3.4 GW; if external compute deals are included, META's effective incremental capacity is about 4 GW.
- The current-generation NVDA GPU buildout of a 1 GW data center can cost up to about 2x as much as current-generation in-house ASICs, but NVDA leads in compute performance per watt by about 2-8x versus in-house ASICs.
- About 50% or more of AMZN, MSFT, and META 2026 capex is for forward buildouts that will come online in 2027 and beyond; GOOGL is only about 10%, implying higher execution and component inflation risk.
- Racks/servers remain the largest investment item per GW, networking accounts for about 20%, power enclosures are in the high-single-digit to low-double-digit range, and HBM and DRAM may also push costs higher.
Report interpretation
Overview
The report builds a bottom-up per-GW cost and capacity rollout model around roughly $2 trillion of hyperscale cloud data center capex from 2024 to 2027, and evaluates the AI compute capacity, cost differences, forward buildout pace, and data center component inflation driven by GPUs and in-house ASICs. Its core conclusion is that as GenAI enters the inference era, compute capacity remains a key bottleneck for industry growth and competitiveness, and a large share of the next two years of capex will begin to convert into usable capacity.
Core views
The report believes that AMZN, GOOGL, META, and MSFT will add about 14 GW and 20 GW of capacity in 2026 and 2027, respectively. GOOGL is expected to add the most capacity in 2027, at about 7 GW, mainly to support Gemini training, GCP growth, and GenAI features in core products such as Search and YouTube; AMZN and MSFT are each expected to add about 5 GW, while META is expected to add about 3.4 GW, though META's effective incremental capacity is about 4 GW if external hyperscale compute purchases are included. At the chip level, NVDA remains the leader in incremental capacity, accounting for about 60% of total new capacity in 2026/2027, but the share of GOOGL TPU and AMZN Trainium is rising.
Analysis framework
The report uses a bottom-up capacity model, combining each company's estimated chip counts, rack counts, per-GW cost, chip mix, and data center construction costs to estimate the GW capacity associated with current and future hyperscale cloud capex. It also compares NVDA Blackwell and Vera Rubin with in-house ASICs such as TPU and Trainium in terms of capex, compute performance per watt, HBM capacity, networking, and software optimization.
Methodology notes
cost per GW
By breaking out chip, rack, server, networking, power enclosure, and memory costs, the report estimates the capex required to build 1 GW of AI data center capacity and uses that to infer future online capacity for each cloud vendor.
cost/token/watt
The report not only compares chip and rack capex, but also emphasizes how compute performance per watt, HBM capacity, network efficiency, and software optimization affect inference economics.
future capacity capex share
By distinguishing between capacity going online in the current period and power enclosures, land, equipment, memory, and chip purchases pre-built for 2027 and beyond, the report assesses the second-order growth in future capex and execution risk.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- AMZN / AWSHyperscale cloud vendor and user of the Trainium in-house ASIC
- Strengths
- 2026/2027 capacity catch-up is clear, about 50% of 2026 capex is for future capacity, and AWS demand plus the Trainium roadmap provide long-term support.
- Weaknesses
- Early GenAI capacity buildout was slower, and Trainium adds only about 2 GW per year, roughly half the scale of TPU.
- Comparison
- About 4.7 GW of new capacity in 2027, below GOOGL but close to MSFT.
- Risks
- Data center delays, component inflation, and Trainium performance and ecosystem progress falling short of expectations.
- GOOGL / Google Cloud / TPUExpected to be the hyperscale cloud vendor with the highest new capacity in 2027
- Strengths
- TPU adoption is rising, 2027 new capacity is about 6.8 GW, and it supports Gemini, GCP, and GenAI features in core Search and YouTube products.
- Weaknesses
- Only about 10% of 2026 capex is for future capacity, leaving relatively little forward-build buffer.
- Comparison
- 2027 new capacity leads AMZN, MSFT, and META; the share of in-house ASIC capacity is higher than GPU capacity.
- Risks
- Execution risk, component shortages, and inflation risk are relatively high, and the allocation of TPU output between third-party data centers and self-use still needs to be watched.
- MSFT / AzureA major buyer of AI infrastructure and NVDA GPUs
- Strengths
- About 55% of 2026 capex is for capacity in 2027 and beyond, supporting long-term growth in Azure, Azure AI, and M365.
- Weaknesses
- Compared with AWS and GCP, the report believes MSFT is almost entirely dependent on NVDA, leaving less flexibility in chip supply and cost.
- Comparison
- About 4.7 GW of new capacity in 2027, similar to AMZN.
- Risks
- NVDA supply, GPU cost, cloud demand conversion, and the sustainability of capex.
- METAA large-scale AI compute demand player, but without a standalone hyperscale cloud business
- Strengths
- Effective incremental capacity in 2026/2027, including external deals, is about 4 GW, with first-party data and potential consumer AI agent productization opportunities.
- Weaknesses
- It lacks cloud revenue as a hedge, so returns depend more heavily on AI productization, advertising, and the realization of new revenue streams.
- Comparison
- Its own 2027 new capacity is about 3.4 GW, below GOOGL, AMZN, and MSFT.
- Risks
- Insufficient AI model productization, a heavy burden from external compute commitments, and difficulty proving ROIC.
- NVDACore supplier of AI GPUs and incremental capacity
- Strengths
- Accounts for about 60% of incremental capacity in 2026/2027, leads in compute performance per watt by about 2-8x versus current-generation in-house ASICs, and has a strong HBM ecosystem and allocation capability.
- Weaknesses
- Building a 1 GW data center with current-generation NVDA GPUs can cost up to about 2x as much as current-generation in-house ASICs, and Blackwell/Rubin rack costs are high.
- Comparison
- In-house ASICs are cheaper on a capex basis, but NVDA leads in performance per watt and ecosystem maturity.
- Risks
- TPU and Trainium catch-up, customers directly buying memory to reduce bundle value, and rising next-generation GPU costs may affect procurement pace.
- In-house ASICs / TPU / TrainiumAn alternative path for cloud vendors to lower AI compute capex
- Strengths
- Construction cost per GW is significantly lower than NVDA GPUs, and GOOGL and AMZN can improve system efficiency through full-stack control.
- Weaknesses
- Current performance per watt still lags NVDA, and gaps still need to be narrowed through networking, HBM capacity, and software optimization.
- Comparison
- Current-generation in-house ASIC rack costs are about $6 billion to $11 billion per GW, lower than about $20 billion per GW for NVDA GB300 and about $25 billion per GW for Vera Rubin.
- Risks
- Insufficient improvement in performance per watt, software ecosystem limitations, and HBM and advanced-node supply constraints.
Key data
- Cumulative hyperscale cloud capex from 2024 to 2027About $2 trillionThe report says hyperscale cloud capex is expected to exceed $1 trillion in 2027, and cumulative 2024-2027 capex is about $2 trillion.
- New capacity in 2025About 6.7 GWAbout 2.1 GW for GOOGL, 1.8 GW for AMZN, 1.4 GW for MSFT, and 1.4 GW for META.
- New capacity in 2026About 13.7 GWAbout 5.6 GW for GOOGL, 3.5 GW for AMZN, 2.6 GW for MSFT, and 1.9 GW for META.
- New capacity in 2027About 19.5 GWAbout 6.8 GW for GOOGL, 4.7 GW for AMZN, 4.7 GW for MSFT, and 3.4 GW for META.
- NVDA share of incremental capacityAbout 60%NVDA remains the main source of hyperscale cloud incremental capacity in 2026/2027, but the share of TPU and Trainium is increasing.
- Cost difference between NVDA GPU and in-house ASIC data center buildoutUp to about 2xBuilding a 1 GW data center with current-generation NVDA GPUs can cost up to about 2x as much as current-generation in-house ASICs such as TPU and Trainium.
- NVDA compute performance per watt advantageAbout 2-8xThe report believes NVDA significantly outperforms current-generation in-house ASICs on FLOPs/Watt.
- AMZN 2026 forward capexAbout $90 billion, about 50%Used for buildings, shells, facilities, power, and capacity that will come online in 2027 and beyond.
- MSFT 2026 forward capexAbout $104 billion, about 55%This shows substantial pre-build activity in AI infrastructure, supporting future growth in Azure, Azure AI, and M365.
- META 2026 forward capexAbout $85 billion, about 55%Used for chips, racks, and GW capacity that will come online in 2027 and beyond; META also has about $25 billion in annual external hyperscale cloud commitments.
- GOOGL 2026 forward capexAbout $20 billion, about 10%The lower forward-build share implies higher inflation, shortage, and data center execution risk.
- Networking cost as a share of per-GW costAbout 20%The report identifies networking as one of the largest investment items after racks.
Impact & implications
If the report's estimates prove correct, AI infrastructure will move from a capex investment phase into a capacity-release phase, and cloud revenue growth, AI productization, and ROIC will become the market's key validation points. For NVDA, although GPU unit costs are higher, its performance-per-watt advantage preserves its core position; for GOOGL and AMZN, in-house ASICs are expected to lower capex and strengthen vertical integration advantages; for META, because it lacks an independent hyperscale cloud business, its large compute spending depends more heavily on AI productization and incremental revenue proof to justify returns.
Risks
- Data center construction may be constrained by power, land, labor, contractors, electricians, and plumbers, delaying capacity online dates.
- DRAM and HBM pricing or supply tightness could increase rack costs, with a bigger impact during the NVDA Blackwell and Rubin cycles.
- GOOGL's low forward-build share could expose it to greater execution, inflation, and shortage risk.
- If in-house ASICs cannot narrow the performance-per-watt gap through networking, memory bandwidth, and software optimization, the lower capex advantage may be offset by inference efficiency.
- META's large compute investment needs to be justified through AI productization and incremental revenue, or the market may question the return on capex.
- Hyperscale cloud vendors may increasingly rely on fixed-rate long-term debt or innovative financing structures, so financing costs and balance sheet pressure need to be monitored.
What to watch
- The revenue range per MW/GW at hyperscale cloud vendors and what it implies for cloud revenue and ROIC.
- Progress at Trainium 4 and TPU 8 in cost/token/watt, performance per watt, HBM capacity, and software optimization.
- How much of GOOGL TPU output ultimately remains on GOOGL's balance sheet versus going into third-party data centers.
- Whether AMZN starts selling Trainium to third-party data centers; the report does not currently include that upside case.
- Whether power, land, and labor constraints affect the 2026-2027 timing of data center online dates.
- Whether Rubin plug-in memory modules lead cloud customers to buy more approved memory directly, thereby reducing NVDA bundle costs.
- Whether AMZN, MSFT, and META forward buildouts lead to a slowdown in future capex second-order growth.