The computing power of the five major US tech giants is still in short supply.
An unprecedentedly radical computing power construction plan is unfolding in the United States.
The computing power construction of US tech giants is advancing towards a scale of 10GW (Gigawatt is a unit of power. A 1GW computing power cluster can accommodate 200,000 NVIDIA GB200 chips, and the overall cost of building a 1GW computing power cluster exceeds $10 billion) and one million chips.
Capital expenditure (the financial indicator Capital Expenditure, abbreviated as CapEx) is a key indicator for observing computing power investment. The reason is that more than 80% of the capital expenditure of large technology companies is usually used for computing power procurement, such as building data centers, purchasing servers, chips, and storage network devices.
As of November 1st, the five major US technology companies - Microsoft, Amazon, Google, Meta, and Oracle - have all disclosed the progress of their capital expenditures in 2025 and 2026 during their earnings conference calls (Microsoft and Oracle calculate based on fiscal years. Microsoft's fiscal year is from July 1st of the previous year to June 30th of the current year, and Oracle's fiscal year is from June 1st of the previous year to May 31st of the current year).
Comprehensive information from the earnings conference calls shows that the capital expenditure of these five technology companies in 2026 will exceed $470 billion (approximately 3.3 trillion RMB). Conservatively calculated, the growth rate of capital expenditure in 2026 will be at least over 26% (conservatively calculated, the capital expenditures of Amazon and Google will only maintain the scale of 2025).
Data shows that in 2025, the total capital expenditure of the five major technology companies will exceed $370 billion (approximately 2.6 trillion RMB), with a total growth rate of about 64%. In 2024, the total capital expenditure of the five major technology companies was only $220.8 billion (approximately 1.6 trillion RMB), less than half of that in 2026 (for details, see "Will DeepSeek Reduce the Total Demand for Computing Power? The Four Major US Tech Giants' Computing Power Investment Will Not Decrease This Year").
The significant annual increase in the capital expenditure of large technology companies has also sparked discussions about whether AI (Artificial Intelligence) is over - bubbled. However, judging from the backlog of orders (i.e., remaining performance obligations, the financial indicator RPO, Commercial remaining performance obligation) in the financial reports of the five major US technology companies, its growth rate is exceeding that of capital expenditure.
Capital expenditure represents market supply, and the backlog of orders represents market demand. This means that the computing power of the five major US tech giants is in short supply.
With the continuous growth of capital expenditure and the backlog of orders, NVIDIA is the direct beneficiary. As of the third quarter of 2025, 88% of NVIDIA's revenue came from AI chips for data centers. NVIDIA disclosed during its Q4 2024 earnings conference call that half of this revenue came from large technology companies and cloud computing providers such as Microsoft, Amazon, Google, Meta, and Oracle.
Driven by these major customers, on October 29th, Eastern Time in the United States, NVIDIA's (NASDAQ: NVDA) market value exceeded $5 trillion, becoming the first company in history with a market value of over $5 trillion (for details, see "NVIDIA's Market Value Exceeds $5 Trillion, and Jensen Huang Says There Are $500 Billion in Orders to Be Delivered").
The entire computing power market, from upstream chip supply, to mid - stream cloud computing and software services, and then to downstream customer AI application needs, is forming a virtuous cycle.
Significant Growth in Computing Power Investment
How will the over $470 billion in capital expenditure of the five major US technology companies in 2026 be spent?
The latest earnings conference calls of Microsoft, Amazon, Google, Meta, and Oracle (for Microsoft and Oracle, it's the first quarter of fiscal year 2026; for Amazon, Google, and Meta, it's the third quarter of 2025) revealed the details. The conclusion is that these expenditures will include at least over 16GW of computing power and over 3 million AI chips.
Microsoft's management said that it expects the growth rate of capital expenditure in fiscal year 2026 (from July 1st, 2025, to June 30th, 2026) to be higher than that in fiscal year 2025 (from July 1st, 2024, to June 30th, 2025). Microsoft's capital expenditure in fiscal year 2025 was $64.6 billion, a year - on - year increase of 45.1%. Calculated at this growth rate, Microsoft's capital expenditure in fiscal year 2026 will reach as high as $93.7 billion.
Microsoft's management disclosed that the total scale of Microsoft's data centers will double in the next two years. In the third quarter of 2025, Microsoft announced a computing power cluster in Fairwater, Wisconsin, USA. This cluster will be launched in 2026 with a capacity of 2GW.
Amazon expects its capital expenditure in 2025 to be approximately $125 billion and to increase in 2026. In terms of computing power construction, Amazon added 3.8GW of computing power in the past 12 months. At least 1GW of computing power will be launched in the fourth quarter of 2025, and the total computing power will double in 2027.
Google's capital expenditure in 2025 will be between $91 billion and $93 billion, higher than the previous expectation of $85 billion. It is expected that the capital expenditure will increase significantly in 2026. In terms of computing power construction, Google announced in July this year that it will invest $6 billion to build a 1GW computing power center in Andhra Pradesh, southern India.
Meta expects its capital expenditure in 2025 to be between $70 billion and $72 billion, higher than the previous forecast of between $66 billion and $72 billion. Meta expects that the growth rate of capital expenditure in 2026 will be significantly higher than that in 2025. Among them, the computing power demand of the MSL (Meta Superintelligence Labs) team is growing the most strongly.
Currently, Meta is building a 1GW computing power cluster called "Prometheus" in Ohio, USA, and a 5GW computing power cluster called "Hyperion" in Louisiana, USA.
Oracle's capital expenditure in fiscal year 2026 (from June 1st, 2025, to May 31st, 2026) is approximately $35 billion. In 2025, Oracle launched at least two 1GW computing power clusters. This includes the "Stargate" project in Abilene, Texas, USA, with a target capacity of 2GW, and the "Stargate" project in the United Arab Emirates, with a target capacity of 1GW.
Based on the above information, the total power of the computing power clusters above 1GW that Microsoft, Amazon, Google, Meta, and Oracle have announced they are building themselves currently exceeds 16GW. This means that there will be at least over 3 million AI chips to be purchased and over $160 billion in new computing power investment in the next 6 to 24 months.
Goldman Sachs, an investment research institution, pointed out in a research report on June 28th this year that the global computing power supply in 2024 was 57GW. Based on this data, this additional 16GW of computing power is more than 28% of the global computing power in 2024.
A person in the data center industry explained that usually, a complete NVIDIA GB200 NVL72 cabinet (including 36 Grace CPUs and 72 Blackwell GPUs) has a total power consumption of approximately 120KW (kilowatts). After deducting heat dissipation, lighting, network, and power losses, a 1GW computing power cluster can accommodate at least 200,000 NVIDIA GB200 series chips, and a 16GW computing power cluster can accommodate at least 3 million NVIDIA GB200 series chips.
Another person in the data center industry told Caixin that the cost of building a 1GW computing power center in the United States is approximately between $10 billion and $20 billion. The cost includes three parts: first, server hardware - related, such as GPUs (Graphics Processing Units), CPUs (Central Processing Units), HBM (High - Bandwidth Memory), etc.; second, network - related, including switches, optical modules, fiber optic cables, etc.; third, the land, buildings, racks, power supply facilities, and liquid - cooling systems of the data center.
In addition to the above - mentioned additional at least 16GW of computing power, there is also a large amount of existing computing power that needs to be upgraded. Therefore, a part of the capital expenditure will be spent on this.
A strategic person from a Chinese cloud service provider told Caixin that the iteration cycle of AI chips is shorter. In the past, the depreciation cycle of CPU chips was generally 6 to 8 years, but for AI chips, it is usually only 3 to 6 years. Especially for training chips, the service life is only about 3 years. This means a shorter depreciation cycle and greater computing power investment.
The Backlog of Orders Is Also Growing
The rapid growth of the capital expenditure of the five major US technology companies is mainly due to the rapid growth of market demand. Currently, the speed of computing power construction cannot even keep up with the growth rate of the backlog of orders. That is to say, the computing power is actually in short supply.
Market demand involves a core indicator - the backlog of orders (i.e., remaining performance obligations, the financial indicator RPO, Commercial remaining performance obligation), which refers to the signed contracts waiting to be delivered.
The significant growth of the backlog of orders means that more and more customers are signing larger - scale and longer - term contracts for computing power usage with Microsoft, Amazon, Google, and Oracle. The business model of cloud computing and software companies is the subscription system, where customers pay in advance on a monthly/quarterly/annual basis. During the subscription period, customers consume the contract resources. The prepaid fees are not immediately recognized as revenue but are gradually recognized as income over the next 1 to 3 years as the resources are consumed.
As the backlog of orders continues to grow, tech giants will invest in corresponding computing power to fulfill their contracts. The financial report data shows that the growth rate of the backlog of orders of Microsoft, Amazon, Google, Meta, and Oracle is exceeding the growth rate of capital expenditure.
Caixin comprehensively counted the total remaining performance obligations of Microsoft, Amazon, Google, Meta, and Oracle from the second quarter of 2022 to the second quarter of 2025, which were $401.2 billion, $459.3 billion, $601.1 billion, and $1146 billion respectively. In these three years, the growth rate of the backlog of orders of the five major tech giants has been continuously increasing, with growth rates of 14.5%, 30.6%, and 90.7% respectively.
Even after deducting the $300 billion, five - year large - scale order that Oracle signed with OpenAI in the first quarter of 2025 (OpenAI's $300 billion order has an excessive impact on the overall growth rate, and there is a risk that it may not be fulfilled), the total remaining performance obligations of the five major tech giants in the second quarter of 2025 still reached $846 billion, a year - on - year increase of 41% (for details, see "OpenAI Signs a New $38 Billion Order with Amazon, and It's a Mystery How Nearly $600 Billion in Computing Power Orders Will Be Fulfilled").
In the latest earnings conference calls, every time the capital expenditure data is mentioned, investors will ask - whether the high - level capital expenditure can achieve a corresponding return on investment. Correspondingly, the management of Microsoft, Amazon, Google, Meta, and Oracle will mention the growth of the backlog of orders and state that the computing power is currently in short supply.
Microsoft's management said in the earnings conference call for the first quarter of fiscal year 2026 (i.e., the third quarter of 2025) that Microsoft has nearly $400 billion in backlog of orders, and the market demand is growing at a faster - than - expected rate. Moreover, the average duration of these contracts is two years. Most of the resources will be consumed in a short period of time.
Amazon's management disclosed in the earnings conference call for the third quarter of 2025 that as of the end of the third quarter, the backlog of orders of Amazon AWS had reached $200 billion, not including several new unannounced transactions reached in October. Amazon is currently significantly increasing its computing power capacity. Currently, the bottleneck of computing power capacity is mainly limited by energy, and in the future, it may be limited by chips.
Google's management disclosed in the earnings conference call for the third quarter of 2025 that the backlog of orders of Google Cloud reached $155 billion at the end of the third quarter. This is mainly due to the strong demand for enterprise - level AI. The number of billion - dollar orders signed by Google Cloud in the first nine months of 2025 exceeded the total of the previous two years. Although the company has been working hard to increase its capacity and speed up server deployment and data center construction, it is expected that there will still be a tight supply - demand situation in the fourth quarter of 2025 and in 2026.
Oracle's management was more aggressive in their judgment during the earnings conference call for the first quarter of fiscal year 2026 (from June 1st, 2025, to August 31st, 2025). They believe that Oracle's future backlog of orders is expected to grow to $5 trillion. Oracle's management said that the revenue of Oracle Cloud's IaaS (Infrastructure) business will increase by 77% to $18 billion in fiscal year 2025 and will reach $32 billion, $73 billion, $114 billion, and $144 billion in the next four years respectively.
Meta's management's view in the earnings conference call for the third quarter of 2025 was that Meta invested in a certain amount of computing power based on an optimistic assumption. Meta's founder and CEO, Mark Zuckerberg, mentioned that currently, Meta's applications (Facebook, Instagram, WhatsApp, Messenger, Threads) and advertising have always suffered from a shortage of computing power. Meta's computing power demand is constantly growing and can bring higher profits. The company is confident in developing more intelligent algorithms to optimize recommendations for applications and advertising, thereby achieving profitability.