Morgan Stanley expects $2 trillion in hyperscaler capex to materially unlock AI data center capacity by 2027
AI summary card
Morgan Stanley expects $2 trillion in hyperscaler capex to materially unlock AI data center capacity by 2027
Using a bottom-up cost-per-GW and chip shipment model, the report estimates the incremental AI data center capacity of AMZN, GOOGL, META, and MSFT. It concludes that 2027 incremental capacity will be about 20GW, with GOOGL adding the most. NVDA still accounts for about 60% of incremental capacity, but TPU and Trainium shares are rising.
- AMZN, GOOGL, META, and MSFT are expected to add about 14GW and 20GW of capacity in 2026 and 2027, respectively, with 2027 incremental capacity roughly 3x 2025 levels.
- GOOGL is expected to add about 7GW in 2027, the most among the four hyperscalers; AMZN and MSFT are each around 5GW, while META is about 3.4GW to 4GW of effective capacity.
- Building 1GW of data center capacity with current-generation NVDA Blackwell GPUs requires capex that can be as much as 2x current-generation Custom ASIC, but NVDA leads Custom ASIC by roughly 2x to 8x in compute performance per watt.
- About 50% or more of AMZN, MSFT, and META's 2026 capex is for forward builds that will come online in 2027 and beyond; GOOGL's share is only about 10%, implying higher inflation, shortage, and execution risk.
- Servers/racks remain the largest cost item per GW, at about 60%. Outside racks, networking is about 20%, power enclosures are in the high-single-digit to low-double-digit range, and HBM and DRAM may push costs higher.
Report interpretation
Overview
This report focuses on how compute capacity becomes the core constraint on AI development speed, cloud revenue growth, and the competitive advantage of big tech as GenAI enters the inference era. Morgan Stanley estimates that cumulative hyperscaler data center capex from 2024 to 2027 will approach $2 trillion, and 2027 annual capex could exceed $1 trillion. The report builds a bottom-up model of per-GW cost and chip/rack counts to estimate AMZN, GOOGL, META, and MSFT incremental AI data center capacity, forward-purchase scale, chip mix, and cost structure.
Core views
The key conclusions are: first, capacity is being released materially, with the four hyperscalers adding about 14GW and 20GW of capacity in 2026 and 2027, respectively, meaning 2027 incremental capacity is close to 3x 2025 levels. Second, GOOGL is expected to add about 7GW in 2027, ahead of AMZN and MSFT at roughly 5GW each, while META's own capex supports about 3.4GW, or about 4GW of effective incremental capacity when external cloud compute agreements are included. Third, NVDA remains the market leader and is expected to account for about 60% of incremental hyperscaler capacity in 2026-2027, though GOOGL TPU and AMZN Trainium shares are rising. Fourth, the cost to build 1GW with NVDA GPUs is materially higher than with Custom ASIC, but NVDA clearly leads in performance per watt, making cost per token per watt a key future battleground. Fifth, the large forward-build programs at AMZN, MSFT, and META help dampen future second-derivative capex growth and offset component inflation and data center delay risk, whereas GOOGL's forward-build ratio is lower.
Analysis framework
The report uses a bottom-up data center capex decomposition: it first estimates the per-GW build cost for different chip architectures, then combines chip units, rack counts, and go-live years by company to derive annual incremental capacity and capex needs. The analysis also separates GPU, TPU, Trainium, and other chip sources, and breaks cost into servers/racks, networking, power enclosures, memory, land, and facilities, in order to assess the quality of capex, the share of forward builds, and future execution risk across hyperscalers.
Methodology notes
Per-GW cost model
Estimates the capex required to build 1GW of AI data center capacity based on chip, rack, server, networking, power enclosure, memory, and facility costs, and uses that to compare the capital efficiency of NVDA GPU and Custom ASIC architectures.
Chip mix and capacity contribution
Estimates incremental GW capacity from NVDA GPU, GOOGL TPU, and AMZN Trainium chip units to determine each company's investment tilt between GPUs and in-house ASICs.
Forward purchasing and future capacity bridge
Splits current-year capex between capacity that comes online now and assets that come online in future years, used to assess second-derivative capex growth, component inflation hedging, and data center delivery risk.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- NVDACore GPU and rack supplier for AI data centers
- Strengths
- Expected to account for about 60% of incremental hyperscaler capacity in 2026/2027, and Blackwell plus the next Vera Rubin/Rubin Ultra generations lead Custom ASIC by about 2x to 8x in performance per watt.
- Weaknesses
- Building 1GW of data center capacity with current-generation NVDA GPUs can cost as much as about 2x Custom ASIC, and chip plus HBM costs may raise customer capex.
- Comparison
- Versus TPU and Trainium, NVDA is more expensive but stronger in performance per watt, HBM ecosystem, and supply allocation power.
- Risks
- Customer migration to in-house ASICs, rising HBM costs, higher next-generation GPU pricing, and margin pressure from hyperscalers directly sourcing memory.
- GOOGLThe hyperscaler with leading new AI capacity and a major TPU user
- Strengths
- Expected to add about 7GW in 2027, the highest among the four hyperscalers; TPU adds about 4GW in both 2026 and 2027, supporting GenAI product deployment for Gemini, GCP, and Search/YouTube.
- Weaknesses
- Only about 10% of 2026 capex is for capacity that comes online in 2027 and beyond, implying a low forward-build ratio.
- Comparison
- Compared with AMZN, MSFT, and META, GOOGL has the largest incremental capacity but the least future capacity buffer.
- Risks
- Component inflation, shortages, data center execution risk, and uncertainty over whether TPU supply remains on GOOGL's balance sheet or flows into third-party data centers.
- AMZNAWS AI infrastructure expansion and Trainium push
- Strengths
- Expected to add about 4.7-5GW in 2027; about $90bn, or about 50% of 2026 capex, is for capacity that comes online in 2027 and beyond, helping offset component inflation and data center delays.
- Weaknesses
- The report says AWS got off to a slower start on GenAI capacity buildout in 2023/2024, making 2026/2027 a catch-up phase.
- Comparison
- Trainium adds about 2GW per year, roughly half the scale of TPU, and represents about 50% of AMZN's annual incremental capacity.
- Risks
- Trainium external sales pace, AWS demand conversion, return on forward builds, and future changes in capex growth.
- MSFTDemand source for Azure and M365 AI infrastructure
- Strengths
- About $104bn, or about 55% of 2026 capex, is for capacity that comes online in 2027 and beyond, supporting long-term Azure, Azure AI, and M365 growth.
- Weaknesses
- Compared with AWS and GCP, MSFT is almost entirely dependent on NVDA and has less chip mix diversification.
- Comparison
- 2027 incremental capacity is about 4.7-5GW, similar to AMZN, but in-house ASIC contribution is low.
- Risks
- Dependence on NVDA supply and pricing, returns on prebuilt AI infrastructure, and conversion of incremental cloud and M365 revenue.
- METAAn internet platform with no hyperscale cloud business but very large AI compute investment
- Strengths
- Its own capex is expected to add about 3.4GW in 2027, and when external cloud compute transactions are included, effective incremental capacity in 2026 and 2027 is about 4GW; first-party data and potential consumer agent products may support model monetization.
- Weaknesses
- Without a hyperscale cloud business, it must prove ROIC through advertising, consumer products, and AI tool revenue.
- Comparison
- Unlike AMZN, GOOGL, and MSFT, META has no cloud infrastructure commercialization channel, making it more dependent on product innovation and incremental revenue.
- Risks
- Insufficient AI productization, operating expense pressure from external cloud commitments, and uncertainty over whether neocloud capacity needs to be resold or offloaded.
- TPU / Trainium / Custom ASICIn-house AI accelerators that replace or complement NVDA GPUs
- Strengths
- Current-generation Custom ASIC has materially lower per-GW build cost than NVDA GPUs, and GOOGL and AMZN can control more of the stack.
- Weaknesses
- Current-generation Custom ASIC lags NVDA by about 2x to 8x in compute performance per watt, and system-level optimization still needs work.
- Comparison
- TPU adoption is higher than Trainium; the report estimates TPU adds about 4GW in both 2026 and 2027, while Trainium adds about 2GW per year.
- Risks
- Insufficient progress in networking, HBM capacity, and software optimization, plus uncertainty over third-party sales and asset ownership.
Key data
- Cumulative hyperscaler capex from 2024 to 2027About $2 trillionThe report says aggregated hyperscaler data center spending from 2024 to 2027 will approach this scale.
- 2027 hyperscaler capexAbove $1 trillionThe chart shows AMZN, MSFT, GOOGL, and META combined at about $1,018bn in 2027.
- Incremental capacity from the four hyperscalersAbout 14GW in 2026 and about 20GW in 2027Combined incremental AI data center capacity for AMZN, GOOGL, META, and MSFT.
- 2027 incremental capacity rankingGOOGL about 6.8-7GW, AMZN about 4.7-5GW, MSFT about 4.7-5GW, META about 3.4GWMETA's effective capacity is about 4GW after including external hyperscale agreements.
- NVDA incremental capacity shareAbout 60%The report expects NVDA to still account for about 60% of aggregated incremental hyperscaler capacity in 2026/2027.
- NVDA GPU versus Custom ASIC per-GW build costCurrent-generation NVDA GPU can cost up to about 2x current-generation Custom ASICThe main difference comes from server/rack costs.
- NVDA performance-per-watt advantageAbout 2x to 8xThe report believes NVDA is significantly ahead of current-generation Custom ASIC in compute performance per watt.
- AMZN 2026 forward-build capexAbout $90bn, about 50%Used for buildings, shells, facilities, power, and capacity that come online in 2027 and beyond.
- MSFT 2026 forward-build capexAbout $104bn, about 55%Supports long-term growth in Azure, Azure AI, and M365.
- META 2026 forward-build capexAbout $85bn, about 55%Used for chips, racks, and GW capacity that come online in 2027 and beyond.
- GOOGL 2026 forward-build capexAbout $20bn, about 10%The report argues this implies higher inflation, shortages, and data center build execution risk.
- Main cost items outside racksNetworking about 20%, power enclosures about high-single-digit to low-double-digit percentages, HBM about mid-single-digit to mid-double-digit percentagesDRAM and HBM may create upside risk for rack prices.
Impact & implications
The investment implication of this report is that AI capex should not be viewed only through total spend, but also through per-GW cost, the timing of capacity online, chip mix, and whether spending is a forward build for future capacity. GOOGL's capacity expansion is the most aggressive, which is positive for Gemini, GCP, and GenAI deployment across Search/YouTube, but its lower forward-build share raises execution and inflation risk. The large forward-build programs at AMZN, MSFT, and META may imply near-term capex pressure, but they could also lead to a slower second derivative of capex growth and stronger supply assurance later. NVDA is more expensive, but its lead in performance per watt keeps it highly competitive; Custom ASIC will need improvements in networking, HBM capacity, and software optimization to narrow the system-level performance-per-watt gap.
Risks
- Data center construction may be delayed by practical constraints such as power, land, contractors, electricians, and plumbers.
- Rising HBM, DRAM, and advanced-node costs may increase server/rack costs and hurt per-GW capital efficiency.
- GOOGL's low forward-build ratio means that if it keeps expanding aggressively in 2027, it may face higher inflation, shortages, and execution risk.
- If Custom ASIC cannot narrow the performance-per-watt gap through networking, memory bandwidth, and software optimization, it may struggle to replace high-end NVDA GPUs.
- Because META lacks a cloud monetization channel, large-scale compute capex could pressure ROIC if AI productization and incremental revenue fall short.
- If hyperscalers increasingly use long-term debt or innovative financing structures, capital structure and financial risk could change.
What to watch
- Each hyperscaler's revenue per GW or revenue per MW range to assess cloud revenue growth and ROIC.
- The degree of performance-per-watt, HBM capacity, networking efficiency, and software stack improvement in Trainium 4 and TPU 8.
- Whether NVDA Rubin and Rubin Ultra restore NVDA's share of incremental capacity after launch.
- The split of GOOGL TPU shipments between internal use and third-party data centers.
- Whether AMZN begins selling Trainium to third-party data centers.
- The impact of HBM4, TSMC advanced nodes, and memory module procurement patterns on GPU rack costs and NVDA margins.
- Whether power, land, labor, and power enclosure delivery become bottlenecks for capacity online.
- Whether AMZN, MSFT, and META's forward builds lead to slower second-derivative capex growth after 2027.