$2 Trillion Flows into AI: The 2027 Compute Power Surge and the True Cost Picture

Overview

This report provides an in-depth analysis of the substantial capital expenditures by global hyperscale cloud providers on AI infrastructure between 2024 and 2027, along with the resulting changes in computing capacity. The report highlights that, as generative AI enters the inference phase, computing capacity has become a critical competitive barrier. Using a bottom-up modeling approach, Morgan Stanley estimates that the cumulative $2 trillion in investments will add approximately 20 gigawatts (GW) of new computing capacity by 2027. The report further breaks down the cost and energy-efficiency differentials across various chip architectures—GPU versus ASIC—and compares the investment strategies and forward-looking initiatives of leading cloud providers.

Core views

The scale of computing capacity expansion is staggering, with a peak expected in 2027. According to research reports, Amazon (AMZN), Google (GOOGL), Meta (META), and Microsoft (MSFT) are projected to add 14 GW and 20 GW of computing capacity in 2026 and 2027, respectively. By comparison, AWS has cumulatively added only about 5 GW over the past 18 years as of the end of 2024. The additional capacity planned for 2027 would be sufficient to meet the annual electricity consumption of more than 15 million U.S. households. Among them, Google is expected to contribute the largest incremental capacity in 2027—around 7 GW—followed by Amazon and Microsoft, each adding roughly 5 GW. Although Meta does not operate its own cloud services business, it is anticipated to achieve effective capacity additions of approximately 4 GW through internal development and external leasing. NVDA remains the market leader, but proprietary ASICs are gaining ground thanks to their cost advantages. In terms of chip types, NVIDIA (NVDA) will continue to dominate, accounting for roughly 60% of the new computing capacity added in 2026–2027. However, the share of custom-designed ASICs—such as Google’s TPU and Amazon’s Trainium—is steadily increasing. Google plans to ramp up ASIC deployments over 2026–2027, replacing GPU‑based capacity; about 50% of this will support Google Cloud. Meanwhile, Amazon is accelerating the rollout of its Trainium chips while continuing to procure NVIDIA GPUs to meet AWS demand. Microsoft, on the other hand, relies almost entirely on NVIDIA, primarily for external customers, whereas Meta currently uses NVIDIA for 75–80% of its internal workloads. Cost vs. Efficiency: NVDA Is More Expensive but More Efficient. Research based on underlying models indicates that building a 1 GW data center using the latest generation of NVIDIA GPUs—such as Blackwell—can entail capital expenditures up to twice those of comparable next‑generation proprietary ASICs like TPUs or Trainiums. The primary cost differential lies in server and rack expenses: NVIDIA racks cost roughly $20–25 billion per GW, compared to just $6–11 billion per GW for proprietary ASICs. Nevertheless, NVIDIA significantly outperforms proprietary ASICs in computational efficiency, delivering 2–8 times higher FLOPs per watt. This suggests that while upfront capital outlays may be higher, long-term operational efficiency and unit‑token costs could prove more competitive. Proprietary ASICs must close this gap through network optimization, expanded high‑bandwidth memory (HBM), and software tuning. Investment Strategies Diverge: Forward‑Looking Deployment vs. Immediate Delivery. Major cloud providers exhibit marked differences in the timing of their capital expenditures. Amazon, Microsoft, and Meta are making substantial forward‑looking investments—including land acquisition, power infrastructure, and equipment—projecting that more than 50% of their 2026 capital spending will support capacity slated to come online in 2027 or later. This approach helps smooth future capital expenditure growth and mitigates risks associated with component inflation and delivery delays. By contrast, Google allocates only about 10% of its 2026 budget to future‑oriented capacity, with the bulk directed toward near‑term deliveries, potentially exposing it to greater execution risks and upward pressure from component price volatility. Data Center Cost Structure: Networking and Power Infrastructure Are Key Drivers. Beyond chips and racks, networking equipment accounts for roughly 20% of total data center costs, making it the second-largest expense after servers. Powered shells represent another 6–15% of overall capital outlays. Additionally, fluctuations in the prices of high‑bandwidth memory (HBM) and standard DRAM further elevate rack‑level cost pressures. Notably, NVIDIA bundles memory sales within its Blackwell architecture, capturing additional revenue; under the upcoming Rubin architecture, cloud providers may gain opportunities to procure memory directly to reduce costs. However, the bundled model could still offer systemic efficiency advantages.

Analysis framework

Institutional analysts have adopted a bottom-up modeling approach to estimate computing capacity and capital expenditures. First, they estimate computing capacity (in GW) by tracking chip shipments. Drawing on projections of shipment volumes for various chip types—such as NVIDIA GPUs, Google TPUs, and Amazon Trainium—and factoring in each chip’s power consumption and performance metrics, the report calculates the annual deployed computing capacity of cloud providers, expressed in gigawatts (GW). This methodology sidesteps reliance on vague, company‑reported capital expenditure figures, instead reconstructing actual computing supply at the physical hardware level. Second, the report disaggregates the construction costs of a single GW of data center capacity. It develops a detailed cost model that breaks down the total capital expenditure for building 1 GW of data center infrastructure into seven components: 1) servers/racks (including GPUs/ASICs, CPUs, and HBM); 2) backup power systems; 3) cooling equipment; 4) electrical enclosures, electrical systems, and construction machinery; 5) networking equipment; 6) optical modules/cabling and other ancillary items; and 7) data center interconnectivity and security. By comparing the cost shares across these categories for different chip architectures, the analysis quantifies the sources of cost differentials between GPUs and ASICs. Finally, the report conducts cross‑validation and forward‑looking analysis. The theoretical capital expenditures derived from “number of chips × cost per GW” are benchmarked against companies’ reported or projected actual capital outlays, enabling an assessment of how much funding is allocated to “proactive capacity building”—namely, securing land, power infrastructure, and other resources in anticipation of future demand. This analysis sheds light on the distinct strategic paces adopted by firms in response to supply chain constraints and inflationary pressures.

Methodology notes

• Industry/ Sector Analysis FrameworkSupply-and-Demand Framework

  Computing power capacity serves as the core constraint on AI development.

  The research report identifies computing capacity (in GW) as a key determinant of the pace at which generative AI development and adoption advance. On the demand side, the explosive growth of AI applications has created a supply-demand imbalance, with computing power in short supply. On the supply side, by monitoring the volume of hardware deployments, the report forecasts the timing of capacity release, thereby assessing industry sentiment and the competitive landscape.

• Company Fundamentals and Financial FrameworkFree cash flow analysis

  Forward-Looking Breakdown of Capital Expenditures (Capex)

  The research report not only focuses on the total capital expenditure but also breaks it down into “current‑period deliveries” and “forward‑looking investments.” This analytical approach enables investors to discern how much of a company’s current cash outflows are earmarked for future growth, thereby facilitating a more precise assessment of future free‑cash‑flow pressures and growth potential.

• Valuation MethodologyPE/PEG valuation

  Price-to-Earnings Ratio Valuation Based on Forward EPS

  When rating specific companies such as AMZN, MSFT, and NVDA, the research report calculates target prices by multiplying the 2027 estimated earnings per share (EPS) by a designated price-to-earnings (P/E) multiple. For instance, assigning NVDA a 26x 2027 EPS multiple reflects a premium for its high growth and strong certainty of future prospects.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• NVIDIA Corp (NVDA.O)

  Key beneficiaries: providers of the vast majority of incremental computing power chips, with industry-leading energy efficiency.

  Strengths

  Its FLOPs-per-watt ratio is 2 to 8 times higher than that of competing products; its ecosystem is robust; and the Blackwell/Rubin architecture continues to evolve through iterative improvements.

  Weaknesses

  The construction cost per gigawatt is substantial, roughly twice that of self-developed ASICs, and the company faces competition from cloud service providers that are developing their own custom chips.

  Comparison

  Compared with TPU and Trainium, NVIDIA’s GPUs excel in performance and versatility but lag behind in cost efficiency for certain use cases.

  Risks

  Demand in the AI terminal market falls short of expectations; customers have significantly reduced their GPU procurement; and competitors have launched highly competitive custom hardware.

• Alphabet Inc. (GOOGL.O)

  Aggressive Expanders: Expected to add the largest share of computing power by 2027, with vigorous deployment of TPUs.

  Strengths

  The TPU ecosystem is well-established, with strong cost control; its search and YouTube businesses generate stable cash flow to support AI-related investments.

  Weaknesses

  The forward-looking capital expenditure ratio is low (only 10%), leaving the company exposed to elevated execution risks and component inflation pressures.

  Comparison

  Compared with AMZN and MSFT, GOOGL places greater emphasis on near-term delivery rather than long-term stockpiling, adopting a more aggressive yet riskier strategy.

  Risks

  Advertising revenue has slowed; low monetization rates for new AI products are putting pressure on profit margins; and data center construction is delayed.

• Amazon.com Inc (AMZN.O)

  Steady Catch-Up Player: Balancing GPUs and Trainium, with Substantial Upfront Investment

  Strengths

  AWS maintains a strong market position; it allocates a high proportion of its capital to forward-looking investments (over 50%), thereby hedging supply-chain risks; and its retail business generates robust cash flow.

  Weaknesses

  GenAI capacity building started relatively late and is still catching up; retail profit margins remain volatile.

  Comparison

  Compared with GOOGL, AMZN adopts a more conservative and prudent strategy, placing greater emphasis on long-term capacity reserves.

  Risks

  AWS revenue growth may slow or its profit margin may decline; retail merchandise margins could deteriorate; and the duration of investments may exceed expectations.

• Microsoft Corp (MSFT.O)

  Enterprise‑level leaders: Almost entirely reliant on NVDA, serving external clients.

  Strengths

  Azure and enterprise software (M365) exhibit strong synergies; the company maintains a high proportion of forward-looking investments (55%); and it boasts deep, long-standing customer relationships.

  Weaknesses

  Lacking in-house AI chips and entirely reliant on NVDA, the company’s cost control remains relatively weak.

  Comparison

  Unlike GOOGL and AMZN, MSFT’s computing power is primarily geared toward external enterprise clients rather than internal consumers.

  Risks

  Macroeconomic factors are weighing on IT spending; on-premises solutions are eroding cloud adoption; and AI adoption remains limited.

• Meta Platforms Inc (META.O)

  Internal driver: No cloud business; reliance is entirely on internal consumption and external leasing.

  Strengths

  It possesses a vast trove of first-party data; its advertising business has staged a robust recovery; and it maintains high operational efficiency.

  Weaknesses

  Without external cloud services, the return on computing-power investments hinges entirely on internal product innovation (such as Reels and AI agents); Reality Labs remains unprofitable.

  Comparison

  Unlike the other three, META’s computing power does not generate direct cloud revenue; rather, it serves as a tool to enhance advertising efficiency and user engagement.

  Risks

  Decreased user engagement; regulatory restrictions on ad targeting; expanding losses at Reality Labs; misaligned execution in data center construction.

Key data

• Cumulative Capital Expenditures from 2024 to 2027Approximately US$2 trillionThe total investment by hyperscale cloud providers in AI infrastructure
• New computing power capacity in 2027~20 GWCompared with approximately 6.7 GW in 2025, there is a significant increase.
• NVDA vs. ASIC Construction CostsNVDA is up about 2xNVDA Blackwell rack costs approximately $20–25 billion per GW, while ASICs cost roughly $6–11 billion per GW.
• NVDA vs. ASIC Energy Efficiency RatioNVDA by 2 to 8 times higherFloating-point operations per watt (FLOPs/Watt)
• Cost Proportion of Networking Equipment~20%The second-largest component of total data center costs, after servers and racks.
• Forward-looking capital expenditure ratios of Amazon, Microsoft, and Meta>50%The proportion of capital expenditures in 2026 allocated to capacity expansion for 2027 and beyond.
• Google’s forward-looking spending ratio~10%The proportion of capital expenditures in 2026 allocated to capacity expansion for 2027 and beyond.

Impact & implications

For cloud service providers, the rapid expansion of computing capacity implies potential acceleration in AI‑related revenue, but also entails substantial capital expenditures. Amazon, Microsoft, and Meta have demonstrated stronger supply-chain control and greater confidence in future demand by securing land and power resources well in advance, which could enable them to maintain more stable cost structures and delivery capabilities in the competitive landscape ahead. By contrast, Google, despite its sizable current investments, has relatively limited forward‑looking reserves, leaving it vulnerable to execution risks and inflationary pressures. For chip suppliers, NVIDIA, with its superior energy efficiency and robust ecosystem, will likely remain dominant in high‑end training and complex inference workloads, even at premium pricing. However, the cost advantages of custom chips such as Google’s TPU and Amazon’s Trainium in specific workloads are prompting cloud providers to increase the share of ASICs within hybrid architectures, posing a long‑term structural challenge to pure‑GPU vendors. For infrastructure providers, network equipment manufacturers—such as those supplying optical modules and switches—and power‑infrastructure firms will continue to benefit from this wave of investment, given their significant and indispensable share of overall costs.

Risks

• The AI terminal market demand failed to materialize as expected, prompting customers to significantly reduce their GPU purchases.
• Data center construction is subject to physical constraints such as power supply, labor shortages, and land availability, which can lead to delivery delays.
• Component inflation—such as rising prices for HBM and networking equipment—has driven up data center construction costs.
• In-house ASIC development has made slow progress in system-level power-efficiency optimization, failing to close the performance gap with NVIDIA.
• Macroeconomic downturns weigh on advertising revenues and corporate IT spending, thereby constraining cloud providers’ capital expenditure capacity.

What to watch

• Quarterly capital expenditure breakdowns and forward-looking capacity guidance disclosed by each cloud provider.
• NVDA’s Rubin architecture and the release cadence of its subsequent products, as well as customer adoption rates.
• The external sales performance of Google TPU and Amazon Trainium (if any), as well as the adoption rate among third-party data centers.
• The supply dynamics and price trends of High Bandwidth Memory (HBM).
• The actual monetization progress and user adoption rates of AI applications, such as Copilot and AI agents.