By Senior Industry Analyst Claire Lin (林巧珍)
AI giants are increasingly using capital investment, computing capacity, chip procurement, and long-term power contracts to build an interconnected AI ecosystem.
Taiwanese companies are deeply embedded in the global AI infrastructure supply chain. As infrastructure investment is increasingly shared through cross-investments, computing capacity commitments, and joint supplier projects, access to electricity is emerging as a critical constraint.
Disruptions in power availability or delays in data center deployment could create broader risks for Taiwanese suppliers. Companies therefore need to strengthen risk buffers across contracts, inventory, production capacity, and geographic deployment.
According to recent financial reports from major AI companies and forecasts for the second-quarter earnings season, capital expenditures by Microsoft, Alphabet, and Amazon are expected to grow significantly faster than their cloud revenues between 2024 and 2026. Annual capital expenditure growth is estimated at 45.2% to 116.1%, compared with cloud revenue growth of 18.5% to 95.7%.
This divergence highlights the increasing capital intensity of AI infrastructure development.
At the same time, growing market concerns over the returns on AI investment are making it increasingly difficult for individual companies to bear the full cost of infrastructure expansion. Tech companies are developing risk-sharing arrangements through cross-investments, long-term computing commitments, and equity-linked partnerships.
Diversifying the AI ecosystem
After investing more than US$10 billion (NT$318 billion) in OpenAI, Microsoft has expanded its AI ecosystem beyond a single partnership. It has invested in European AI company Mistral AI and used a non-asset acquisition structure to recruit the co-founders and core technical team of Inflection AI.
OpenAI has also sought to diversify its infrastructure. In addition to relying on Microsoft, it has partnered with Oracle to expand external computing capacity, reducing its dependence on a single provider for computing infrastructure.
Amazon and Alphabet have invested at least US$8 billion and US$2 billion, respectively, in Anthropic. Anthropic has agreed to use AWS and Google Cloud as major cloud providers. It has also committed to using Amazon's Trainium chips as one of the main hardware platforms for training and deploying future generations of its Claude models, while using Google's latest TPUs to train selected models.
Meta has committed to purchasing up to 6 GW of deployment capacity from AMD over the next five years. As part of the agreement, AMD issued performance-based warrants allowing Meta to acquire up to 160 million shares of AMD common stock at an exercise price of US$0.01 per share. Vesting is linked to the shipment of 6 GW of hardware as well as AMD's stock-price milestones.
AI companies are also seeking to secure electricity supply directly. Microsoft and Amazon have explored behind-the-meter models that would allow data centers to be located near nuclear power plants and connected through dedicated transmission infrastructure.
Amazon's initial plan encountered regulatory obstacles from the US Federal Energy Regulatory Commission. The company subsequently shifted toward a grid-connected model in 2025 while continuing to secure long-term power supply. It signed 20-year Power Purchase Agreements (PPAs) covering hundreds to thousands of megawatts with nuclear power providers including Talen Energy.
Microsoft similarly signed a 20-year PPA with Constellation to support the restart of Unit 1 at the Three Mile Island nuclear power plant. The agreement is intended to secure zero-carbon electricity for data centers located within the PJM Interconnection service area. Meta has also signed a 20-year low-carbon power purchase agreement with energy company Vistra.
Together, these agreements show that risk-sharing arrangements can distribute the financial burden of large-scale infrastructure investment across multiple companies. However, they also increase interdependence among AI companies.
AI tests grid capacity
The power requirements of next-generation AI systems are increasing rapidly. A single chip based on Nvidia's Blackwell architecture consumes approximately 1 kW, while rack power density has increased to between 60 kW and 120 kW. The next-generation Rubin platform is expected to push cluster-level power requirements to as much as 600 kW.
At the same time, major global computing hubs are facing increasingly limited grid capacity and shortages of low-carbon electricity. According to a June 2026 Capgemini report, 77% of surveyed power utility executives believe that growth in data center electricity demand will exceed their ability to expand power supply.
Taiwan faces similar constraints. Taipower has tightened its review process for new electricity applications, while the Ministry of Economic Affairs has announced that approvals for new data centers exceeding 5 MW north of Taoyuan will temporarily be suspended.
Securing generation capacity does not guarantee immediate access to electricity. Even after a data center or power generation facility is completed, operators may still need to wait for transmission lines, substations, and other grid infrastructure to be connected.
PJM Interconnection, the largest regional transmission organization in the United States, has indicated that average waiting times for new connection applications have increased from less than two years in the past to five years or, in some cases, as long as eight years.
If delays in power access become widespread across major data center markets, the effects could extend upstream into the technology supply chain.
This issue is particularly relevant to Taiwan because Taiwanese companies hold significant global market shares across semiconductor manufacturing, advanced packaging, and AI server manufacturing. If major overseas power hubs face electricity shortages or rationing that disrupt data center operations or delay new capacity deployment, the resulting impact would extend beyond order shifts among individual customers and could lead to a broader downward adjustment in orders across the supply chain.
Building resilience and reducing dependencies
As power constraints create greater uncertainty in AI infrastructure deployment, Taiwanese manufacturers need to shift part of their focus from capacity expansion toward operational and financial resilience.
Current practices among major Taiwanese ICT manufacturers suggest two main approaches: financial risk sharing and geographic diversification based on power availability.
Taiwanese manufacturers can use commercial contracts to reduce their exposure to delays or cancellations caused by energy-related disruptions. For customized components, manufacturers may adopt NCNR (non-cancellable, non-returnable) provisions or inventory buyback agreements.
For high-value next-generation projects, some manufacturers may also use consignment arrangements in which customers provide key materials, while requiring higher advance payments. Additional mechanisms can be introduced to allocate responsibility for excess or obsolete inventory. These arrangements can reduce manufacturers' exposure to inventory depreciation and financial losses if data center projects are postponed because of power constraints.
The contrast between the multi-gigawatt power requirements of major US AI companies and the 5 MW restriction on new data centers in northern Taiwan illustrates the physical limitations of expanding computing infrastructure in electricity-constrained locations.
When planning overseas expansion, Taiwanese companies could therefore incorporate electricity availability into site-selection decisions.
They may also work with customers to locate new computing projects in regions with relatively abundant low-carbon electricity and shorter grid-connection timelines. Potential examples include Texas, parts of Northern Europe, and selected locations in Southeast Asia with geothermal or hydropower resources.
Under this approach, Taiwanese manufacturers would not only respond to customer demand for equipment but also participate more actively in decisions concerning the geographic deployment of computing infrastructure.
Moving forward
As interdependence across the AI ecosystem increases, Taiwanese manufacturers should also strengthen scenario planning. Companies could conduct dynamic stress tests covering different levels of order reduction, ranging from a base-case scenario of approximately 20% to an extreme scenario of approximately 35%.
Such scenarios can be linked to predefined response mechanisms. For example, if power constraints create a high risk of project delays, manufacturers could temporarily freeze lead-time commitments to upstream suppliers to prevent excessive inventory accumulation. They could also establish material traceback or reallocation mechanisms with downstream customers to manage components originally allocated to delayed projects.
Instead of using excess capacity primarily to secure low-margin projects, Taiwanese manufacturers could use these periods to strengthen their technological capabilities.
R&D and engineering teams could be reassigned to next-generation platforms, including efforts to resolve manufacturing and assembly bottlenecks associated with new AI products. Production teams could also accelerate automation and digitalization of existing manufacturing lines. Such investments may place additional pressure on companies' short-term financial resources. However, they can also create technological and operational capabilities that support competitiveness when the next cycle of AI infrastructure investment begins.
* Market Intelligence & Consulting Institute (MIC) was established as a division of the Institute for Information Industry (III) in 1987 in Taiwan. MIC serves as an industry consultant, and a government think tank. With over 100 analysts, our focus on ICT research offers a deep understanding of tech trends and their geopolitical impacts. In partnership with the Taiwan government, we transform insights into policies that enhance the nation's industrial strength, boost technological competitiveness, and adeptly manage geopolitical challenges.
