Is U.S. AI Facing a Bigger Bubble or an Application Outbreak? An Analysis from Six Dimensions

As we enter the second half of 2025, the U.S. AI industry has reached an unprecedented climax. NVIDIA's market value has exceeded $4 trillion, and Microsoft's market value has also once broken through the $4 trillion mark. The annual capital expenditures of the four major U.S. AI giants are each approaching $100 billion. The valuations of U.S. AI unicorn enterprises have soared to hundreds of billions of dollars, and the annualized revenue of AI-native companies has exceeded $10 billion. The cost of poaching a top AI talent in Silicon Valley has even reached an astonishing $1 billion.

This round of AI revolution was triggered by OpenAI's successive releases of ChatGPT and GPT-4 from the end of 2022 to March 2023. People once assumed that GPT-5, GPT-6, and so on would be launched every one or two years, and that Artificial General Intelligence (AGI) would soon be realized. However, the launch of GPT-5 has been more difficult and delayed than expected. Large models are evolving towards multimodality, reasoning, and intelligent agents in the physical world. A true technological revolution is far more tortuous and "abnormal" than initially demonstrated, but its overall direction remains unchanged until a complete productivity solution is formed by combining with existing technologies on a newly built infrastructure.

The U.S. clearly stated in its newly released AI Action Plan that AI will bring an industrial revolution, a new information revolution, and even a "Renaissance"-like effect. Do these trends indicate that U.S. AI is becoming an even bigger bubble? Or is AI ushering in an outbreak of applications, driving more industries towards intelligent upgrading through innovation, improving labor productivity, bringing profound changes to the economic structure, and even achieving ultimate transformations in energy and computing? Under the banners of AGI and Artificial Super Intelligence (ASI), can we continue to expect Silicon Valley to expand the boundaries of AI technology, reduce costs, and promote the popularization of AI in the economy and society? To answer these questions, we need to observe from the following six aspects:

1. New Upsurge in AI Investment

In the second half of 2025, Silicon Valley giants have launched a new round of computing power arms race. Google, Microsoft, Amazon, and Meta have all invested in unprecedentedly large-scale capital expenditures, generally approaching the $100 billion mark.

Microsoft has concluded its 2025 fiscal year, and its guidance for capital expenditures in the next quarter has reached $30 billion, setting the tone for more than $120 billion in the full 2026 fiscal year. Meta has raised the lower limit of its capital expenditures for the 2025 fiscal year by $30 billion compared to the previous fiscal year. Since it has promised "similarly significant growth" in the 2026 fiscal year, this means that its expenditures will exceed $100 billion. Amazon had previously committed to $100 billion in capital expenditures this year.

The capital expenditures of these four tech giants (Big 4) are approaching $100 billion per quarter. These expenditures are driving U.S. economic growth. In 2025, capital expenditures related to AI computing power will account for approximately 2% of U.S. GDP, contributing 0.7 percentage points to GDP growth. The scale of AI investment is sufficient to stimulate U.S. economic growth.

However, behind this investment boom, there are also hidden risks of irrational expansion. As mentioned in the Oriental Finance Magazine report, the global investment in AI-related infrastructure such as chips and data centers has reached $1.5 trillion this year and is expected to exceed $2 trillion next year. In the first half of 2025, the average U.S. GDP growth was 1.6%, which was better than initially expected, but this growth was almost driven by AI investment. Data shows that AI-related capital expenditures contributed nearly half of U.S. GDP growth in the first half of the year, and even reached 92% under some statistical standards.

In Silicon Valley and Wall Street, the slogan "computing power is the future" has become an absolute consensus, and enterprises have fallen into a collective "fear of missing out" (FOMO), worrying that falling behind will lead to being eliminated by the times. However, the irrational expansion behind this investment boom is breeding multiple risks, causing the U.S. economy to accumulate increasing vulnerabilities under the superficial prosperity. For example, OpenAI's first-half revenue was only $4.3 billion, but its net loss was as high as $13.5 billion, and the net loss in the third quarter even exceeded $11.5 billion, falling into a strange cycle of "the more it develops, the more it loses money". A McKinsey survey revealed the industry's truth: nearly 80% of enterprises that have deployed AI have failed to achieve profit growth, and 95% of generative AI pilot projects have not brought direct financial returns.

2. Infrastructure Investment and Migration

The destination of these massive investments determines whether AI can move from technological breakthroughs to application revolutions, and the most important current focus is on AI infrastructure, where more and more new production factors - tokens - are being generated.

Microsoft and Google were once fierce competitors, but no one expected that before AI could disrupt Google's search business, it had already begun to disrupt Amazon's cloud services. In the second quarter, the revenue growth of Amazon AWS was only 17%, lower than that of its competitors Microsoft (39%) and Google (37%). If AWS does not make efforts, it will lose its position as the world's largest cloud giant to Microsoft by the end of next year. During the earnings call, Microsoft CEO Satya Nadella proudly announced that "Microsoft is leading in AI infrastructure construction", with a newly added scale of 2GW in the past year, surpassing any other cloud service provider.

In terms of the adoption speed of new-architecture chips, AWS is also slower than its main competitors. At Microsoft's Build Conference in May this year, Microsoft Vice President Scott Guthrie stated in his keynote speech that "some cloud providers, such as AWS, have not yet launched services based on GB200".

The weakness of Amazon Cloud Services is related to its unstructured vertical integration layout for AI. Compared with Google's video and search business, Meta's social network, and Microsoft's enterprise applications, Amazon's core business is e-commerce, and the breadth and depth of its current AI implementation are slightly lower. It also lacks self-developed cutting-edge large models.

As Nadella mentioned, Microsoft once took the lead in this AI revolution by virtue of its magical investment and cooperation model with OpenAI. Although OpenAI is no longer bound to Microsoft Azure since the beginning of the year, Microsoft Copilot still benefits from GPT; Google's Gemini has helped it get rid of the "innovator's dilemma". Due to the fact that the certainty of its own business and large models for cloud business is lower than that of other cloud giants, the intensity and growth rate of AWS's capital expenditures are the lowest among Silicon Valley cloud giants, even lower than Meta's. Moreover, Amazon is originally a heavy-asset tech giant, and a large part of its capital expenditures are used for logistics and warehousing.

Competition will intensify in the second half of the year, even though capital expenditures are eroding the giants' short-term profit margins and free cash flow. Both Microsoft and Amazon believe that AI-driven cloud services will continue to grow at a high speed. They also mentioned that the global IT stock is still migrating to the cloud, and this trend is mutually driven by the growth of AI workloads. Amazon stated that currently, 85% to 90% of global IT expenditures are still deployed on-premises, but this ratio will reverse in 10 or 15 years; and the migration may even be faster due to AI and other reasons. Microsoft's view is that "we are still far from the finish line".

This infrastructure investment and migration trend is not only changing the competitive landscape of cloud service providers but also profoundly affecting the entire IT industry chain. As pointed out in the research report of Microsoft Research (document 4), the migration of IT infrastructure to the cloud and the growth of AI workloads are mutually reinforcing. This not only drives the demand for cloud services but also promotes the upgrading of related hardware and software technologies, such as the development of high-performance chips, high-speed networks, and efficient data storage systems.

3. Stimulating the Manufacturing Industry

Investment in data centers has directly stimulated the demand for the construction and manufacturing industries and provided impetus for manufacturing investment itself and its upgrading.

The U.S. AI Action Plan hopes that AI will bring an industrial revolution. It clearly proposes to "empower U.S. workers" and "support the next generation of manufacturing" while expanding technology. At the same time, the U.S. has proposed to develop power grid and semiconductor manufacturing technologies that can match the speed of AI innovation and to cultivate a large number of skilled technical workers serving AI infrastructure.

Chips, drones, robots, and autonomous driving are the most strategically important manufacturing fields in the U.S. Currently, the "Silicon Valley Gang" has increasingly close ties with the Pentagon, setting off an upsurge in the intelligentization of defense equipment.

The expansion of this infrastructure not only brings a leap in computing power but also systematically reactivates the team of the new generation of blue-collar workers and engineers. Each data center project requires a large number of well-trained construction workers, electricians, electronics, and electrical engineers.

OpenAI's data center in Texas has a capacity of 1.2GW, xAI's data center in Tennessee also has a capacity of 1.2GW, Meta's data center in Louisiana has a capacity of 2GW, and Amazon's data center in Indiana has a capacity of up to 2.2GW. In the past year, all giants have reached milestone nuclear power agreements. In particular, the "recommissioning" of Microsoft's Three Mile Island Nuclear Power Plant is quite symbolic.

This set of skills can be replicated and migrated. Virginia is known by Americans as the "Global Data Center Capital" with an astonishing density of data centers. The regional manufacturing and engineering skills flywheel effect has taken shape and has spilled over to neighboring states such as Maryland. In the U.S. Rust Belt, former steel factories are being transformed into data centers.

This stimulation of the manufacturing industry is not limited to the construction of data centers. As mentioned in the Framework View article, the U.S. "Star Gate" plan involves a large number of manufacturing links, such as the production of high-end AI chips, the manufacturing of data center equipment, and the construction of supporting new energy facilities. These investments have not only driven the growth of the U.S. manufacturing industry but also promoted the upgrading of manufacturing technologies. For example, the production of high-end AI chips requires more advanced manufacturing processes, which has promoted the development of the U.S. semiconductor manufacturing industry.

4. Token Economy

In AI data centers, as electricity is continuously consumed, tokens are constantly being produced. These tokens will fill the digital space and profoundly change the software product and service industry.

The training of large models is still continuing to expand, and reasoning has just begun. The anxiety about computing power has not disappeared; it comes not only from the guarantee of total amount but also from the rapid reduction of computing power costs to achieve the commercial closed-loop of AI. Meta has just invested heavily to form an AI "dream team" with the ambitious goal of creating personal super intelligence. The guarantee of sufficient computing power was a promise when it was poaching talents. It is said that this team, which currently has only 50 people, can control 30,000 GPU cards.

Every business under the umbrella of these giants benefits from AI. The monthly active users of AI functions in all Microsoft products have exceeded 800 million. Amazon CEO Andy Jassy admitted that the demand for AI computing power exceeds supply, and AWS's order backlog has exceeded $195 billion, a year-on-year increase of 25%; Microsoft CTO Amy Hood disclosed that this figure has reached $368 billion, a year-on-year increase of 37%.

Reinforcement learning technology and reasoning models have ushered in the era of intelligent agents (Agentic). The application of Agentic will lead to an exponential growth in token consumption. The number of tokens consumed by a single agent is 4 times that of model dialogue, and when it comes to multi-agent systems, the consumption starts from 15 times! Currently, in terminal-side and embodied intelligence applications, the most bottlenecked area is the computing efficiency defined by tokens/joule.

The market's consumption of tokens is experiencing an exponential surge. Google's monthly token processing volume has increased from 480 trillion at the May I/O Conference to 980 trillion in June, doubling in just one month. In fact, Microsoft is not far behind. In the previous quarter, it disclosed a quarterly processing volume of 100 trillion, and this time it disclosed an annual processing volume of 500 trillion, which is equivalent to a 4-fold increase in this quarter. However, Google's statement focuses on the company's caliber, while Microsoft's disclosure is based on the Azure Foundry API caliber, excluding the total number of tokens consumed by its own business. Since the beginning of the year, OpenAI has no longer been restricted to only calling Azure services.

Value has begun to shift from infrastructure to AI applications. In the first 7 months of this year, OpenAI's monthly revenue was approximately $1 billion, doubling from about $500 million at the beginning of the year. There are also reports directly calling it an annualized revenue of $12 billion, almost 3 times that of last year. Its main competitor, Anthropic, was reported to have an annualized revenue of $4 billion in its recent financing, a 3-fold surge from the beginning of the year, and it has achieved a 10-fold growth rate for 3 consecutive years. Silicon Valley venture capital firm Menlo Ventures estimates that Anthropic's annualized revenue has actually exceeded $4.5 billion, surpassing OpenAI in the enterprise service market.

The speed of this value transfer is also accelerating. At the beginning of the year, payment giant Stripe revealed that the top 100 AI companies in terms of revenue can reach the threshold of $5 million in annualized revenue within 24 months, which is much faster than the 37 months of SaaS. For those AI-native startups established in the past three years, the average time taken is only 9 months.

As mentioned in the 36 氪 article about Anthropic, Anthropic's exponential growth stems from a well-designed "growth flywheel" consisting of three core components. The first is the "engine" - a consumption-based revenue model centered on APIs. Unlike OpenAI, which relies on ChatGPT subscription fees (accounting for about 73% of its revenue) through the consumer route, Anthropic firmly chooses enterprise APIs as its core engine, with 85% of its revenue coming from API calls by developers and enterprises, and payment based on the number of tokens processed. This model is subversive in two aspects: first, it can be used by any developer, shortening the value verification cycle from months to hours without going through a long sales cycle, effectively bypassing the bottleneck of traditional enterprise software sales; second, revenue is directly linked to the value created by customers. The more successful the customer's application, the more tokens consumed, and the higher Anthropic's revenue. As a result, the revenue contribution of a certain customer to Anthropic may increase by 10 times in a very short period.

5. "Curing" Baumol's Cost Disease

More importantly, this value transfer occurring in the digital world is actually promoting the transformation of the U.S. economic growth paradigm, and its impact will further directly or indirectly penetrate into the physical world and the real economy.

The U.S. economy has long been trapped in the structural "Baumol's Cost Disease", that is, industries with productivity growth lower than the overall average level of the economy (which can be called "stagnant sectors") often experience higher-than-average cost increases. The resulting "cost disease" may lead to faster-than-average price increases, declining service quality, and financial pressure in these stagnant sectors. In addition, due to the drag of stagnant sectors, the overall productivity growth rate and real output growth rate of the economy may also decline. This theory suggests that if consumers prefer labor-intensive services where productivity is inherently difficult to improve, this may lead to long-term economic stagnation and slow growth in real income.

American economist William Baumol, who pioneered this research, used this theory to explain the cost-price disease in multiple service industries, typically including higher education, urban services, healthcare and hospitals, and performing arts. This time, AI may truly trigger a productivity revolution in the above-mentioned service fields.

Recently, an empirical study by Microsoft has initially reached a conclusion roughly consistent with this. Large language models have the potential to release productivity in service-oriented tasks such as knowledge-based work and information transmission work.

One of the datasets of this study is Copilot-Uniform, which is based on 200,000 anonymized and desensitized interactions between users and Microsoft Copilot; it also has an auxiliary dataset, Copilot-Thumbs, which is user feedback on AI results. Microsoft studied the user goals and AI behaviors in the interactions, as well as the frequency of use, coverage, and success rate of AI behaviors in the work activities of the occupation, to evaluate the AI applicability score of different occupations. A higher score indicates a higher possibility of being affected by AI.

The results show that sales and related occupations, computer and mathematical occupations, and office and administrative support occupations have the highest scores, and they are often the occupations with the largest number of employees. Occupations that require more communication tasks, including community and social service occupations and educational guidance occupations, also have relatively high scores. These are typical examples of "Baumol's Cost Disease" in the U.S. economy.

In this study, Microsoft also found that in 40% of interactions, the tasks requested by users and the tasks actually performed by AI are completely non-overlapping (disjoint). That is, AI is not "replacing humans to complete tasks" but "helping humans understand and advance tasks". Moreover, the scope of impact of tasks has always been the best indicator for predicting user usage rate (activity share), rather than completion rate (completion rate) or satisfaction (positive feedback). This means that the motivation of users is not to find the best results but to save a lot of repetitive work.

It is worth noting that the average diagnostic accuracy rate of Microsoft's newly developed diagnostic agent in tests has exceeded the average level of human doctors, and its accuracy rate in the diagnosis of complex diseases has even exceeded that of top doctors.

It can be seen that AI is embedding and reconstructing traditional service industries at the "task" level, reducing the marginal cost of low-value-added and high-frequency "tasks". Currently, its impact on the economy mainly lies not in "replaceability" but in whether it changes the supply-demand balance, marginal cost, and process efficiency of a certain type of task.

6. The Last Industrial Revolution

This AI investment led by U.S. private sector giants is largely a bet that determines the "national fortune" of the U.S. The intensity and scale of investment have exceeded the communication and network investment during the dot-com bubble period, making it the largest-scale infrastructure investment in the U.S. since the railway investment in the 19th century, and it may not have yet peaked.

Humanity is thus moving towards the last industrial revolution. The demand for computing power and energy will be endless, and quantum computing and nuclear fusion may be its ultimate solutions, as well as the forefront of current global technological competition.

The integration of AI technology and applications with industries and the economy will not only improve labor productivity but also bring changes to the economic structure. Even if the efficiency of the service industry is significantly improved, it is impossible for the labor force to largely "flow back" to the manufacturing system. Robots are increasingly replacing front-line operational workers engaged in simple and repetitive work, such as most quality inspectors who "tighten screws" and identify quality with the naked eye, and eventually even drivers engaged in transportation.

A large number of currently "advanced" mental labor requiring high academic qualifications may be the first to be replaced and reduced. Although Microsoft's research tends to believe that the current enhancement effect of AI on the labor force is often greater than the substitution effect, it only analyzes the interactions between users and Copilot, and the samples of software programming users and intelligent agent users are insufficient. However, these two groups are precisely the directions where large models are implemented most quickly.

In Anthropic's user research on its own Claude model, the demand and effect feedback of programming users are far higher than those in other fields. Currently, in major tech companies in Silicon Valley, the "pass line" for the automation rate of software code has reached 30%, and it is increasing at a faster rate.

No company understands computers and software better than Microsoft. Last month, it laid off employees again, marking the fourth round of layoffs this year. This time, 9,000 people were laid off, second only to the scale of 10,000 layoffs at the beginning of 2023. At that time, the entire U.S. tech industry laid off more than 160,000 people in just one quarter. Currently, in the U.S., due to the popularization of AI, the demand for computer science graduates in many industries has decreased, and Silicon Valley is also experiencing rounds of layoffs. However, there is a huge shortage of manufacturing workers and engineers in the U.S.

If we say that AI is "eating" software, and software is "eating" the world, the hundreds of billions of dollars in capital expenditures of U.S. giants are not only improving productivity at the software and service industry levels but also driving changes in the U.S. labor structure in hardware and the real economy. In the short term, these AI infrastructures springing up across the U.S. are not only the computing power centers of Silicon Valley giants but also, together with the investment in supporting energy and power electrical equipment, serve as retraining centers for the U.S. manufacturing industry.

In a longer economic cycle, the reallocation of the labor force will also restart. If AI technology can transform sectors with low efficiency and high costs, such as education, healthcare, and government departments that require complex approval and administrative management, and redirect this part of the labor force to emerging manufacturing industries and related R&D and service industries, it may be another major transformation of the U.S. job market after entering the so-called "post-industrial society" and information society.

Moravec's Paradox tells us that the learning ability and skills combining human perception and action are the most difficult for AI and robots to replace. These related economic fields include a large number of service-oriented manual skills, technical blue-collar workers, AI maintenance jobs, on-site judgment and operation, the integration of service and technical skills, personalized customization craftsmanship, the delicate combination between humans and AI, and humans using AI to expand the frontiers of innovation. These will become the areas where the human labor force defends and creates value. If we say that AI is passing the "final human test", the U.S. AI investment boom and competition with China are also testing its last chance to lead the industrial revolution.

Conclusion

To sum up, the current U.S. AI industry is in a critical period of both opportunities and risks. On the one hand, the huge investment in AI is driving the upgrading of infrastructure, stimulating the development of the manufacturing industry, promoting the formation of the token economy, helping to solve Baumol's Cost Disease, and moving towards the goal of the last industrial revolution. On the other hand, there are also risks such as irrational investment, bubble formation, and labor structure imbalance.

Whether U.S. AI will eventually evolve into a bigger bubble or achieve an application outbreak depends on whether it can effectively convert technological advantages into real productivity, whether it can achieve a commercial closed-loop, and whether it can balance the relationship between innovation, development, and risk prevention. In the context of global AI competition, the U.S. AI industry still has certain advantages in technological research and development, capital investment, and talent accumulation, but it also faces challenges such as technological bottlenecks, market uncertainties, and international competition. Only by continuously promoting technological innovation, optimizing investment structure, improving application efficiency, and solving social problems such as labor employment can the U.S. AI industry achieve sustainable development and truly play a role in promoting economic growth and social progress.

Is U.S. AI Facing a Bigger Bubble or an Application Outbreak? An Analysis from Six Dimensions

As we enter the second half of 2025, the U.S. AI industry has reached an unprecedented climax. NVIDIA's market value has exceeded $4 trillion, and Microsoft's market value has also once broken through the $4 trillion mark. The annual capital expenditures of the four major U.S. AI giants are each approaching $100 billion. The valuations of U.S. AI unicorn enterprises have soared to hundreds of billions of dollars, and the annualized revenue of AI-native companies has exceeded $10 billion. The cost of poaching a top AI talent in Silicon Valley has even reached an astonishing $1 billion.

This round of AI revolution was triggered by OpenAI's successive releases of ChatGPT and GPT-4 from the end of 2022 to March 2023. People once assumed that GPT-5, GPT-6, and so on would be launched every one or two years, and that Artificial General Intelligence (AGI) would soon be realized. However, the launch of GPT-5 has been more difficult and delayed than expected. Large models are evolving towards multimodality, reasoning, and intelligent agents in the physical world. A true technological revolution is far more tortuous and "abnormal" than initially demonstrated, but its overall direction remains unchanged until a complete productivity solution is formed by combining with existing technologies on a newly built infrastructure.

The U.S. clearly stated in its newly released AI Action Plan that AI will bring an industrial revolution, a new information revolution, and even a "Renaissance"-like effect. Do these trends indicate that U.S. AI is becoming an even bigger bubble? Or is AI ushering in an outbreak of applications, driving more industries towards intelligent upgrading through innovation, improving labor productivity, bringing profound changes to the economic structure, and even achieving ultimate transformations in energy and computing? Under the banners of AGI and Artificial Super Intelligence (ASI), can we continue to expect Silicon Valley to expand the boundaries of AI technology, reduce costs, and promote the popularization of AI in the economy and society? To answer these questions, we need to observe from the following six aspects:

1. New Upsurge in AI Investment

In the second half of 2025, Silicon Valley giants have launched a new round of computing power arms race. Google, Microsoft, Amazon, and Meta have all invested in unprecedentedly large-scale capital expenditures, generally approaching the $100 billion mark.

Microsoft has concluded its 2025 fiscal year, and its guidance for capital expenditures in the next quarter has reached $30 billion, setting the tone for more than $120 billion in the full 2026 fiscal year. Meta has raised the lower limit of its capital expenditures for the 2025 fiscal year by $30 billion compared to the previous fiscal year. Since it has promised "similarly significant growth" in the 2026 fiscal year, this means that its expenditures will exceed $100 billion. Amazon had previously committed to $100 billion in capital expenditures this year.

The capital expenditures of these four tech giants (Big 4) are approaching $100 billion per quarter. These expenditures are driving U.S. economic growth. In 2025, capital expenditures related to AI computing power will account for approximately 2% of U.S. GDP, contributing 0.7 percentage points to GDP growth. The scale of AI investment is sufficient to stimulate U.S. economic growth.

However, behind this investment boom, there are also hidden risks of irrational expansion. As mentioned in the Oriental Finance Magazine report, the global investment in AI-related infrastructure such as chips and data centers has reached $1.5 trillion this year and is expected to exceed $2 trillion next year. In the first half of 2025, the average U.S. GDP growth was 1.6%, which was better than initially expected, but this growth was almost driven by AI investment. Data shows that AI-related capital expenditures contributed nearly half of U.S. GDP growth in the first half of the year, and even reached 92% under some statistical standards.

In Silicon Valley and Wall Street, the slogan "computing power is the future" has become an absolute consensus, and enterprises have fallen into a collective "fear of missing out" (FOMO), worrying that falling behind will lead to being eliminated by the times. However, the irrational expansion behind this investment boom is breeding multiple risks, causing the U.S. economy to accumulate increasing vulnerabilities under the superficial prosperity. For example, OpenAI's first-half revenue was only $4.3 billion, but its net loss was as high as $13.5 billion, and the net loss in the third quarter even exceeded $11.5 billion, falling into a strange cycle of "the more it develops, the more it loses money". A McKinsey survey revealed the industry's truth: nearly 80% of enterprises that have deployed AI have failed to achieve profit growth, and 95% of generative AI pilot projects have not brought direct financial returns.

2. Infrastructure Investment and Migration

The destination of these massive investments determines whether AI can move from technological breakthroughs to application revolutions, and the most important current focus is on AI infrastructure, where more and more new production factors - tokens - are being generated.

Microsoft and Google were once fierce competitors, but no one expected that before AI could disrupt Google's search business, it had already begun to disrupt Amazon's cloud services. In the second quarter, the revenue growth of Amazon AWS was only 17%, lower than that of its competitors Microsoft (39%) and Google (37%). If AWS does not make efforts, it will lose its position as the world's largest cloud giant to Microsoft by the end of next year. During the earnings call, Microsoft CEO Satya Nadella proudly announced that "Microsoft is leading in AI infrastructure construction", with a newly added scale of 2GW in the past year, surpassing any other cloud service provider.

In terms of the adoption speed of new-architecture chips, AWS is also slower than its main competitors. At Microsoft's Build Conference in May this year, Microsoft Vice President Scott Guthrie stated in his keynote speech that "some cloud providers, such as AWS, have not yet launched services based on GB200".

The weakness of Amazon Cloud Services is related to its unstructured vertical integration layout for AI. Compared with Google's video and search business, Meta's social network, and Microsoft's enterprise applications, Amazon's core business is e-commerce, and the breadth and depth of its current AI implementation are slightly lower. It also lacks self-developed cutting-edge large models.

As Nadella mentioned, Microsoft once took the lead in this AI revolution by virtue of its magical investment and cooperation model with OpenAI. Although OpenAI is no longer bound to Microsoft Azure since the beginning of the year, Microsoft Copilot still benefits from GPT; Google's Gemini has helped it get rid of the "innovator's dilemma". Due to the fact that the certainty of its own business and large models for cloud business is lower than that of other cloud giants, the intensity and growth rate of AWS's capital expenditures are the lowest among Silicon Valley cloud giants, even lower than Meta's. Moreover, Amazon is originally a heavy-asset tech giant, and a large part of its capital expenditures are used for logistics and warehousing.

Competition will intensify in the second half of the year, even though capital expenditures are eroding the giants' short-term profit margins and free cash flow. Both Microsoft and Amazon believe that AI-driven cloud services will continue to grow at a high speed. They also mentioned that the global IT stock is still migrating to the cloud, and this trend is mutually driven by the growth of AI workloads. Amazon stated that currently, 85% to 90% of global IT expenditures are still deployed on-premises, but this ratio will reverse in 10 or 15 years; and the migration may even be faster due to AI and other reasons. Microsoft's view is that "we are still far from the finish line".

This infrastructure investment and migration trend is not only changing the competitive landscape of cloud service providers but also profoundly affecting the entire IT industry chain. As pointed out in the research report of Microsoft Research (document 4), the migration of IT infrastructure to the cloud and the growth of AI workloads are mutually reinforcing. This not only drives the demand for cloud services but also promotes the upgrading of related hardware and software technologies, such as the development of high-performance chips, high-speed networks, and efficient data storage systems.

3. Stimulating the Manufacturing Industry

Investment in data centers has directly stimulated the demand for the construction and manufacturing industries and provided impetus for manufacturing investment itself and its upgrading.

The U.S. AI Action Plan hopes that AI will bring an industrial revolution. It clearly proposes to "empower U.S. workers" and "support the next generation of manufacturing" while expanding technology. At the same time, the U.S. has proposed to develop power grid and semiconductor manufacturing technologies that can match the speed of AI innovation and to cultivate a large number of skilled technical workers serving AI infrastructure.

Chips, drones, robots, and autonomous driving are the most strategically important manufacturing fields in the U.S. Currently, the "Silicon Valley Gang" has increasingly close ties with the Pentagon, setting off an upsurge in the intelligentization of defense equipment.

The expansion of this infrastructure not only brings a leap in computing power but also systematically reactivates the team of the new generation of blue-collar workers and engineers. Each data center project requires a large number of well-trained construction workers, electricians, electronics, and electrical engineers.

OpenAI's data center in Texas has a capacity of 1.2GW, xAI's data center in Tennessee also has a capacity of 1.2GW, Meta's data center in Louisiana has a capacity of 2GW, and Amazon's data center in Indiana has a capacity of up to 2.2GW. In the past year, all giants have reached milestone nuclear power agreements. In particular, the "recommissioning" of Microsoft's Three Mile Island Nuclear Power Plant is quite symbolic.

This set of skills can be replicated and migrated. Virginia is known by Americans as the "Global Data Center Capital" with an astonishing density of data centers. The regional manufacturing and engineering skills flywheel effect has taken shape and has spilled over to neighboring states such as Maryland. In the U.S. Rust Belt, former steel factories are being transformed into data centers.

This stimulation of the manufacturing industry is not limited to the construction of data centers. As mentioned in the Framework View article, the U.S. "Star Gate" plan involves a large number of manufacturing links, such as the production of high-end AI chips, the manufacturing of data center equipment, and the construction of supporting new energy facilities. These investments have not only driven the growth of the U.S. manufacturing industry but also promoted the upgrading of manufacturing technologies. For example, the production of high-end AI chips requires more advanced manufacturing processes, which has promoted the development of the U.S. semiconductor manufacturing industry.

4. Token Economy

In AI data centers, as electricity is continuously consumed, tokens are constantly being produced. These tokens will fill the digital space and profoundly change the software product and service industry.

The training of large models is still continuing to expand, and reasoning has just begun. The anxiety about computing power has not disappeared; it comes not only from the guarantee of total amount but also from the rapid reduction of computing power costs to achieve the commercial closed-loop of AI. Meta has just invested heavily to form an AI "dream team" with the ambitious goal of creating personal super intelligence. The guarantee of sufficient computing power was a promise when it was poaching talents. It is said that this team, which currently has only 50 people, can control 30,000 GPU cards.

Every business under the umbrella of these giants benefits from AI. The monthly active users of AI functions in all Microsoft products have exceeded 800 million. Amazon CEO Andy Jassy admitted that the demand for AI computing power exceeds supply, and AWS's order backlog has exceeded $195 billion, a year-on-year increase of 25%; Microsoft CTO Amy Hood disclosed that this figure has reached $368 billion, a year-on-year increase of 37%.

Reinforcement learning technology and reasoning models have ushered in the era of intelligent agents (Agentic). The application of Agentic will lead to an exponential growth in token consumption. The number of tokens consumed by a single agent is 4 times that of model dialogue, and when it comes to multi-agent systems, the consumption starts from 15 times! Currently, in terminal-side and embodied intelligence applications, the most bottlenecked area is the computing efficiency defined by tokens/joule.

The market's consumption of tokens is experiencing an exponential surge. Google's monthly token processing volume has increased from 480 trillion at the May I/O Conference to 980 trillion in June, doubling in just one month. In fact, Microsoft is not far behind. In the previous quarter, it disclosed a quarterly processing volume of 100 trillion, and this time it disclosed an annual processing volume of 500 trillion, which is equivalent to a 4-fold increase in this quarter. However, Google's statement focuses on the company's caliber, while Microsoft's disclosure is based on the Azure Foundry API caliber, excluding the total number of tokens consumed by its own business. Since the beginning of the year, OpenAI has no longer been restricted to only calling Azure services.

Value has begun to shift from infrastructure to AI applications. In the first 7 months of this year, OpenAI's monthly revenue was approximately $1 billion, doubling from about $500 million at the beginning of the year. There are also reports directly calling it an annualized revenue of $12 billion, almost 3 times that of last year. Its main competitor, Anthropic, was reported to have an annualized revenue of $4 billion in its recent financing, a 3-fold surge from the beginning of the year, and it has achieved a 10-fold growth rate for 3 consecutive years. Silicon Valley venture capital firm Menlo Ventures estimates that Anthropic's annualized revenue has actually exceeded $4.5 billion, surpassing OpenAI in the enterprise service market.

The speed of this value transfer is also accelerating. At the beginning of the year, payment giant Stripe revealed that the top 100 AI companies in terms of revenue can reach the threshold of $5 million in annualized revenue within 24 months, which is much faster than the 37 months of SaaS. For those AI-native startups established in the past three years, the average time taken is only 9 months.

As mentioned in the 36 氪 article about Anthropic, Anthropic's exponential growth stems from a well-designed "growth flywheel" consisting of three core components. The first is the "engine" - a consumption-based revenue model centered on APIs. Unlike OpenAI, which relies on ChatGPT subscription fees (accounting for about 73% of its revenue) through the consumer route, Anthropic firmly chooses enterprise APIs as its core engine, with 85% of its revenue coming from API calls by developers and enterprises, and payment based on the number of tokens processed. This model is subversive in two aspects: first, it can be used by any developer, shortening the value verification cycle from months to hours without going through a long sales cycle, effectively bypassing the bottleneck of traditional enterprise software sales; second, revenue is directly linked to the value created by customers. The more successful the customer's application, the more tokens consumed, and the higher Anthropic's revenue. As a result, the revenue contribution of a certain customer to Anthropic may increase by 10 times in a very short period.

5. "Curing" Baumol's Cost Disease

More importantly, this value transfer occurring in the digital world is actually promoting the transformation of the U.S. economic growth paradigm, and its impact will further directly or indirectly penetrate into the physical world and the real economy.

The U.S. economy has long been trapped in the structural "Baumol's Cost Disease", that is, industries with productivity growth lower than the overall average level of the economy (which can be called "stagnant sectors") often experience higher-than-average cost increases. The resulting "cost disease" may lead to faster-than-average price increases, declining service quality, and financial pressure in these stagnant sectors. In addition, due to the drag of stagnant sectors, the overall productivity growth rate and real output growth rate of the economy may also decline. This theory suggests that if consumers prefer labor-intensive services where productivity is inherently difficult to improve, this may lead to long-term economic stagnation and slow growth in real income.

American economist William Baumol, who pioneered this research, used this theory to explain the cost-price disease in multiple service industries, typically including higher education, urban services, healthcare and hospitals, and performing arts. This time, AI may truly trigger a productivity revolution in the above-mentioned service fields.

Recently, an empirical study by Microsoft has initially reached a conclusion roughly consistent with this. Large language models have the potential to release productivity in service-oriented tasks such as knowledge-based work and information transmission work.

One of the datasets of this study is Copilot-Uniform, which is based on 200,000 anonymized and desensitized interactions between users and Microsoft Copilot; it also has an auxiliary dataset, Copilot-Thumbs, which is user feedback on AI results. Microsoft studied the user goals and AI behaviors in the interactions, as well as the frequency of use, coverage, and success rate of AI behaviors in the work activities of the occupation, to evaluate the AI applicability score of different occupations. A higher score indicates a higher possibility of being affected by AI.

The results show that sales and related occupations, computer and mathematical occupations, and office and administrative support occupations have the highest scores, and they are often the occupations with the largest number of employees. Occupations that require more communication tasks, including community and social service occupations and educational guidance occupations, also have relatively high scores. These are typical examples of "Baumol's Cost Disease" in the U.S. economy.

In this study, Microsoft also found that in 40% of interactions, the tasks requested by users and the tasks actually performed by AI are completely non-overlapping (disjoint). That is, AI is not "replacing humans to complete tasks" but "helping humans understand and advance tasks". Moreover, the scope of impact of tasks has always been the best indicator for predicting user usage rate (activity share), rather than completion rate (completion rate) or satisfaction (positive feedback). This means that the motivation of users is not to find the best results but to save a lot of repetitive work.

It is worth noting that the average diagnostic accuracy rate of Microsoft's newly developed diagnostic agent in tests has exceeded the average level of human doctors, and its accuracy rate in the diagnosis of complex diseases has even exceeded that of top doctors.

It can be seen that AI is embedding and reconstructing traditional service industries at the "task" level, reducing the marginal cost of low-value-added and high-frequency "tasks". Currently, its impact on the economy mainly lies not in "replaceability" but in whether it changes the supply-demand balance, marginal cost, and process efficiency of a certain type of task.

6. The Last Industrial Revolution

This AI investment led by U.S. private sector giants is largely a bet that determines the "national fortune" of the U.S. The intensity and scale of investment have exceeded the communication and network investment during the dot-com bubble period, making it the largest-scale infrastructure investment in the U.S. since the railway investment in the 19th century, and it may not have yet peaked.

Humanity is thus moving towards the last industrial revolution. The demand for computing power and energy will be endless, and quantum computing and nuclear fusion may be its ultimate solutions, as well as the forefront of current global technological competition.

The integration of AI technology and applications with industries and the economy will not only improve labor productivity but also bring changes to the economic structure. Even if the efficiency of the service industry is significantly improved, it is impossible for the labor force to largely "flow back" to the manufacturing system. Robots are increasingly replacing front-line operational workers engaged in simple and repetitive work, such as most quality inspectors who "tighten screws" and identify quality with the naked eye, and eventually even drivers engaged in transportation.

A large number of currently "advanced" mental labor requiring high academic qualifications may be the first to be replaced and reduced. Although Microsoft's research tends to believe that the current enhancement effect of AI on the labor force is often greater than the substitution effect, it only analyzes the interactions between users and Copilot, and the samples of software programming users and intelligent agent users are insufficient. However, these two groups are precisely the directions where large models are implemented most quickly.

In Anthropic's user research on its own Claude model, the demand and effect feedback of programming users are far higher than those in other fields. Currently, in major tech companies in Silicon Valley, the "pass line" for the automation rate of software code has reached 30%, and it is increasing at a faster rate.

No company understands computers and software better than Microsoft. Last month, it laid off employees again, marking the fourth round of layoffs this year. This time, 9,000 people were laid off, second only to the scale of 10,000 layoffs at the beginning of 2023. At that time, the entire U.S. tech industry laid off more than 160,000 people in just one quarter. Currently, in the U.S., due to the popularization of AI, the demand for computer science graduates in many industries has decreased, and Silicon Valley is also experiencing rounds of layoffs. However, there is a huge shortage of manufacturing workers and engineers in the U.S.

If we say that AI is "eating" software, and software is "eating" the world, the hundreds of billions of dollars in capital expenditures of U.S. giants are not only improving productivity at the software and service industry levels but also driving changes in the U.S. labor structure in hardware and the real economy. In the short term, these AI infrastructures springing up across the U.S. are not only the computing power centers of Silicon Valley giants but also, together with the investment in supporting energy and power electrical equipment, serve as retraining centers for the U.S. manufacturing industry.

In a longer economic cycle, the reallocation of the labor force will also restart. If AI technology can transform sectors with low efficiency and high costs, such as education, healthcare, and government departments that require complex approval and administrative management, and redirect this part of the labor force to emerging manufacturing industries and related R&D and service industries, it may be another major transformation of the U.S. job market after entering the so-called "post-industrial society" and information society.

Moravec's Paradox tells us that the learning ability and skills combining human perception and action are the most difficult for AI and robots to replace. These related economic fields include a large number of service-oriented manual skills, technical blue-collar workers, AI maintenance jobs, on-site judgment and operation, the integration of service and technical skills, personalized customization craftsmanship, the delicate combination between humans and AI, and humans using AI to expand the frontiers of innovation. These will become the areas where the human labor force defends and creates value. If we say that AI is passing the "final human test", the U.S. AI investment boom and competition with China are also testing its last chance to lead the industrial revolution.

Conclusion

To sum up, the current U.S. AI industry is in a critical period of both opportunities and risks. On the one hand, the huge investment in AI is driving the upgrading of infrastructure, stimulating the development of the manufacturing industry, promoting the formation of the token economy, helping to solve Baumol's Cost Disease, and moving towards the goal of the last industrial revolution. On the other hand, there are also risks such as irrational investment, bubble formation, and labor structure imbalance.

Whether U.S. AI will eventually evolve into a bigger bubble or achieve an application outbreak depends on whether it can effectively convert technological advantages into real productivity, whether it can achieve a commercial closed-loop, and whether it can balance the relationship between innovation, development, and risk prevention. In the context of global AI competition, the U.S. AI industry still has certain advantages in technological research and development, capital investment, and talent accumulation, but it also faces challenges such as technological bottlenecks, market uncertainties, and international competition. Only by continuously promoting technological innovation, optimizing investment structure, improving application efficiency, and solving social problems such as labor employment can the U.S. AI industry achieve sustainable development and truly play a role in promoting economic growth and social progress.