Author: 137Labs
Introduction:
In the past two years, discussions in the market surrounding artificial intelligence have primarily focused on whether model capabilities can continue to improve, whether generative AI will become a ubiquitous technology similar to the Internet and smartphones, and how large models will change industries such as search, software, advertising, and enterprise services. As model capabilities continue to enhance, user scale rapidly expands, and global computing power infrastructure enters an intensive construction phase, the focus of the market has clearly shifted to commercial returns, specifically whether the massive capital investment from large tech companies in AI can ultimately be converted into stable income, profit, and free cash flow.
Goldman Sachs points out in "Why AI Companies May Invest More than $500 Billion in 2026" that Wall Street's consensus expectation for capital expenditure among major AI hyperscale cloud companies in 2026 has been raised from $465 billion to $527 billion, while broader estimates indicate that global AI infrastructure investment could continue to expand rapidly in the coming years. As funding flows from chips and servers further into data centers, electricity, cooling, and network facilities, AI is no longer just a phase of software and chip innovation, but is evolving into a global capital reconstruction covering energy, industry, finance, and infrastructure systems.
1. Why AI capital expenditure is still being continuously revised upwards
The reason why large tech companies continue to increase AI investments despite pressure on free cash flow, skepticism about valuations, and even poor stock performance in stages, fundamentally lies in their judgment of AI not as a normal product investment but as a strategic investment that could determine the dominance of technology platforms in the next decade. Therefore, what they fear the most is not overspending tens of billions of dollars in a given year, but losing competitive position in the process of forming the next generation of computing platforms, cloud service systems, and software ecosystems.
Enterprises like Microsoft, Amazon, Alphabet, and Meta are not faced with a simple cost-benefit choice but rather a typical strategic competition. If any company is the first to slow down its investment, its cloud computing clients may turn to competitors with more ample computing power, developer ecosystems may migrate to other platforms, and the enterprise itself may gradually fall behind in model training, reasoning capabilities, and product iteration.
Thus, a significant portion of current AI capital expenditure has obvious defensive characteristics. Even if tech companies have not yet proven that every dollar invested yields ideal returns, they generally believe that the long-term losses from not investing could be greater. When all leading companies reach a similar judgment, the entire industry can fall into a competitive state akin to a "prisoner's dilemma," where each company hopes its rivals will reduce investment, but no leading company is willing to be the first to pull back.
At the same time, the market's understanding of computing power demand is also changing. Originally, AI infrastructure investments primarily served large model training, but as AI Agents gradually emerge, reasoning demand is becoming a more important source of growth. Traditional chatbots typically only need to perform one or a few model calls around user questions, but AI Agents often need to decompose objectives, search for information, compare options, invoke tools, modify plans, and repeatedly verify results to complete a genuinely complex task, resulting in token and computing power consumption that may be multiple times or even several dozen times of regular Q&A scenarios.
Goldman Sachs forecasts that driven by consumer and enterprise adoption of AI Agents, global token usage could grow 24 times by 2030 compared to 2026, reaching 120 quadrillion per month, or 12 x 10^16 tokens. The core logic behind this prediction is that AI may no longer just be a chat tool that users occasionally activate, but will gradually become a persistent execution system embedded in office software, customer service, advertising, financial analysis, e-commerce, software development, and industrial processes.
If this change occurs, the main driver of future computing power demand will shift from a few enterprises training large models periodically to hundreds of millions of users and businesses continuously running intelligent agents. Reasoning computations will also gradually exhibit a sustained consumption characteristic similar to that of cloud computing, electricity, and communication traffic. Therefore, even if training efficiency improves or the training frequency of individual models decreases, the overall societal demand for computing resources may still maintain rapid growth.
Another study by Goldman Sachs indicates that as chip performance improves, model compression technology advances, and data center architecture continues to optimize, the unit token computing cost for AI reasoning may fall by 60% to 70% annually. However, a decrease in unit cost does not necessarily mean that total industry expenditure will decrease, as a common phenomenon in the history of technological development is that the cheaper a resource becomes, the larger its usage scope and total consumption tend to be, and AI is likely to follow this rule.
With the continuous decline in reasoning costs, applications that previously lacked economic feasibility, such as automated customer service, real-time video generation, personalized education, software development agents, and enterprise process automation, will find it easier to enter mass deployment stages. Therefore, the key variable determining the scale of AI capital expenditure in the future will not be the price of a single token, but whether the growth in usage can continuously exceed the rate of decline in unit costs.
2. $527 billion and $765 billion represent two different statistical calibers
Understanding Goldman Sachs' judgment on AI investment scale requires distinguishing between the two figures of $527 billion and $765 billion. The former mainly refers to the consensus expectation of capital expenditure for leading AI cloud computing enterprises such as Microsoft, Amazon, Alphabet, and Meta in 2026, covering servers, data centers, network devices, and other capital projects; the latter derives from Goldman Sachs' broader AI infrastructure model, which includes data center buildings, power supply, cooling facilities, and other supporting infrastructure needed for computing systems beyond direct investments from tech companies. Thus, $527 billion is closer to the annual budget of leading enterprises, while $765 billion corresponds more to the annual construction scale of the entire AI infrastructure system, and $7.6 trillion reflects the cumulative investment that may occur from 2026 to 2031.
Although chips are the most market-attended part of AI infrastructure, they are actually just the starting point of the entire system because a large-scale AI computing cluster requires, besides GPUs and other accelerators, server racks, high-speed networks, optical modules, liquid cooling equipment, uninterruptible power supplies, transformers, transmission lines, backup power generation, and data center buildings capable of supporting these systems. Therefore, every new batch of AI chips often corresponds to systemic investments several times that of the chips themselves.
Goldman Sachs uses several baseline assumptions in related models, including an AI accelerator packaging power of about 3,000 watts, a data center PUE of about 1.2, data center construction costs of about $15 million per megawatt, and new electricity infrastructure costs of about $2,500 per kilowatt. These parameters indicate that the competition in the AI industry is no longer just a competition of chip design capability and model algorithm capability, but is gradually evolving into a comprehensive competition involving land, electricity, financing, supply chain, engineering construction, and long-term operational capacity.
3. Why chip companies are already profitable while the entire AI industry has not yet proven returns
To date, the most clear and concentrated economic value in the AI industry mainly emerges in semiconductor companies and related equipment suppliers. This is because large tech companies need to purchase chips, build data centers, and complete network and power support in advance before their AI businesses can generate stable revenue. Chip suppliers can confirm revenue and profits in a relatively short time once deliveries are completed. In contrast, cloud computing companies and model companies purchasing these chips need to recoup costs gradually over the coming years through cloud services, software subscriptions, enhanced advertising efficiency, or enterprise AI products.
The result is that the current profits in the AI industry chain are mainly concentrated among upstream chip and equipment suppliers, while cloud platforms, model companies, and enterprise clients are still in the investment and commercialization verification stage, and the entire industry has not formed a widespread and stable cash flow loop.
Goldman Sachs points out that while semiconductor companies are achieving record revenue and profits, many other companies in the AI ecosystem have not yet obtained returns that match their inputs. This structure, where upstream suppliers earn first and downstream clients continue to invest, is hard to maintain in the long term. A healthy industrial chain must be based on the ability of end customers to utilize AI to increase revenue, reduce costs, or improve efficiency. Only when profits gradually spread from chip manufacturers to cloud platforms, software companies, and end applications can customers consistently increase purchases and drive the entire industry into a stable positive cycle.
While the consumer market has provided rapid user growth for AI, the number of users cannot be directly equated to commercial value. Relevant studies cited by Goldman Sachs show that the adoption rate for generative AI reached about 53% within three years of the launch of the first widely available product, a pace significantly faster than the early adoption paths of personal computers and the Internet. However, a large number of users are utilizing free products, and even if some consumers are willing to pay subscription fees, individual subscription revenue may not be enough to cover costs associated with model training, reasoning, electricity, data centers, and R&D.
The real support for the tens of trillions of dollars in AI infrastructure investment still comes from enterprise clients, as enterprises have a higher payment capacity and large-scale customer service, sales, R&D, financial, and supply chain processes. However, deploying AI in enterprises is significantly more challenging than typical users using chatbots. Many enterprises initially believe that by purchasing the most advanced large models, they will naturally achieve productivity improvements. The reality is that internal data is often scattered across different systems, suffering from issues such as unified formats, unclear permissions, and varying quality. If inventory, membership, order, and recommendation algorithm data are disjointed, even the most powerful model will struggle to consistently produce reliable commercial results.
Thus, the key issue restraining the implementation of enterprise AI is no longer just model capability but whether enterprises can complete data governance, model orchestration, and business process restructuring. In the future, enterprises will likely call upon different models based on task complexity, cost, data security, and risk levels while establishing permission control, manual review, and result tracking mechanisms. This means that enterprise AI is not as simple as purchasing a software account but involves a long-term project involving system transformation, compliance reviews, and organizational change. Goldman Sachs forecasts that by 2030, only about 12% of knowledge workers may use AI Agents, rising to 37% by 2040, indicating that the pace of infrastructure investment may outstrip the pace at which enterprises realize commercial returns.
4. Is AI investment evolving into a new round of bubbles?
Comparing today's AI boom with the internet bubble of the late 1990s is one of the most common analytical frameworks in the market, as both are accompanied by rapid capital expenditure expansion, optimistic investor expectations about future demand, companies worried about missing the technological revolution and building infrastructure in advance, and many company valuations relying on revenue and profit many years down the line. However, judging that AI will inevitably repeat the internet bubble solely based on these similarities is also overly simplistic.
Today, the primary AI investments are undertaken by global tech giants with large operating cash flows, mature business models, and robust balance sheets, whereas during the internet bubble, many telecom and internet companies relied on high-leverage financing, with numerous companies yet to establish stable revenue and profits. Therefore, the current cycle of AI investment may not end with a comprehensive collapse, but this does not mean that capital misallocation and structural excess will not occur.
The first risk is that the speed of investment may exceed the speed of commercialization. Data centers, chips, and power facilities can be concentrated in construction over a few years, but enterprise AI revenue, process transformations, and organizational changes often take much longer to mature. Once supply is massively formed in advance, while effective paid demand fails to grow in sync, capital returns could be under long-term pressure.
The second risk derives from the uncertainty of the economic lifespan of chips. Goldman Sachs points out that the economic lifespan of AI chips may be one of the most crucial variables determining the cumulative scale of AI investments, as AI accelerators are generally expected to last four to six years, while next-generation chips can often swiftly replace old products with higher performance at a lower unit cost. In contrast, data center buildings typically have a depreciation period of around 20 years, and power infrastructure usage cycles may even reach 25 years or longer. Therefore, if chips need to be frequently updated, tech companies not only have to shoulder ongoing new investments but also face increased depreciation pressure and equipment obsolescence risks.
The third risk arises from power and engineering bottlenecks. Even if enterprises have procured a large number of chips, if data centers cannot obtain power connections on time, or if transformers, transmission lines, generation equipment, and cooling systems cannot be constructed in sync, these chips cannot be fully operational. Thus, in the future, key factors limiting AI development may gradually shift from inadequate chip supply to insufficient power, land, approvals, and engineering capabilities.
Aside from investment risks at the industrial level, the capital market also faces issues of overly rapid valuation expansion. Goldman Sachs research indicates that its AI infrastructure stock basket once achieved around 44% average return within the year, while the consensus expectation for the earnings per share of related companies has only risen about 9% over the next two years, meaning that a significant portion of the stock price increase comes from valuation multiple expansion rather than synchronized earnings forecasts. If stock prices are to continue rising, they must rely on earnings significantly exceeding expectations, further valuation enhancements, or companies demonstrating stronger monopoly positions and pricing power. Once capital expenditure growth slows or downstream clients cut orders, those companies whose stock prices have risen markedly faster than profit expectations may face greater valuation contraction pressures.
5. AI investment is transitioning from thematic trading to fundamental differentiation
Goldman Sachs finds that the average stock price correlation among major listed AI hyperscale companies has dropped from about 80% to 20%, indicating that the market no longer views all AI enterprises as a single thematic trade but is beginning to distinguish differences among companies regarding capital expenditure, capital structure, revenue realization capabilities, and cash flow quality. In the early stages of the AI market, investors were more concerned about whether a company belonged to the AI industry chain. However, as valuations and investment scales continued to rise, the market began to further probe whether capital expenditure has already yielded revenue, whether companies rely on debt financing, whether customer demand is stable, and whether business models have clear profit realization paths. Consequently, AI investment has transitioned from thematic trading to a stricter fundamental screening phase.
In the future, the market may重点关注 the alignment of capital expenditure with AI revenue. Because investment growth itself is not scary; what truly matters is whether AI revenue and profit can grow at a similar or even faster pace. The market will also continue to observe free cash flow, as a company may see profits still growing, but as long as the growth rate of capital expenditure is faster, free cash flow may still decline. Financing structure will also become a vital variable, as investments financed by operational cash flow differ entirely from those relying on debt financing to build data centers.
Moreover, computing power utilization will directly determine whether data center investments can achieve reasonable returns, as the number of servers does not equate to effective demand. Only when computing power can maintain a high utilization rate in the long term can enterprises hope to cover depreciation, electricity, and maintenance costs. Companies with cloud platforms, developer ecosystems, proprietary data, and stable enterprise customer relationships are more likely to convert infrastructure advantages into long-term revenue and customer retention.
6. The next stage of winners may no longer be limited to chip companies
Goldman Sachs believes that after semiconductor companies first capture most AI profits, the next phase of value may gradually spread to hyperscale cloud computing companies, AI platforms, and productivity-benefiting enterprises, as the market has fully recognized the capital expenditure pressures faced by tech giants but may underestimate the profit margins offered by future AI cloud service revenue growth and unit cost reduction.
As reasoning efficiency continues to improve and usage expands, cloud computing platforms have the opportunity to achieve revenue expansion and unit cost reductions simultaneously, gradually forming economies of scale similar to traditional cloud computing businesses. This also means that truly capable platform companies with computing power resources, customer bases, and software ecosystems may gain more stable earnings once AI commercialization enters a mature stage.
At the same time, the AI orchestration layer connecting enterprise data, business processes, and different models may also become a new value center. Platforms able to provide data access, model routing, cost control, permission management, and compliance audits have the opportunity to form fundamental positions similar to databases and middleware in the cloud computing era, securing stable income due to high customer migration costs.
The ultimate profits generated by AI do not necessarily all accrue to AI suppliers because when advertising companies use AI to improve conversion rates, logistics companies reduce transportation costs through AI, and software companies shorten R&D cycles with AI, the economic value created by AI may reflect directly in the profit statements of these users. Therefore, in the next investment phase, investors need to pay attention not only to which companies are selling AI but also to which enterprises can redistribute industry profits through AI, expand market shares, or improve cost structures.
AI infrastructure investment will also create opportunities for many non-traditional tech industries. Goldman Sachs forecasts that relevant capital expenditures by large technology companies may cumulatively reach $5.3 trillion between 2025 and 2030, surpassing the previously predicted $4.5 trillion. This massive construction scale cannot rely solely on the cash of tech companies. Data center developers, power companies, equipment suppliers, infrastructure funds, real estate capital, and credit markets will all become important participants. As of September 2025, global infrastructure funds had managed assets exceeding $1.7 trillion, with about $400 billion awaiting investment. By 2030, the managed asset scale may exceed $3 trillion. This indicates that AI could also become a significant growth driver for private markets, infrastructure finance, and corporate bond issuance in the coming years.
7. How to judge whether AI capital expenditure will ultimately be successful
To determine whether this round of AI capital expenditure is sustainable, one cannot only look at chip sales, server orders, or the number of data centers but should observe from multiple levels, including demand, revenue, unit economic models, capital returns, and profit distribution across the industry chain.
Investors first need to observe whether the number of AI users, enterprise adoption rates, and token usage continue to grow, and further assess whether cloud computing platforms, model companies, and software enterprises can convert usage into paid income. Additionally, they should evaluate whether the revenue generated per reasoning task or AI task is sufficient to cover costs such as chip depreciation, electricity, networking, maintenance, and R&D. Only when real demand, commercial revenue, and reasonable unit economic models are all in place can infrastructure investments sustain long-term support.
On this basis, investors also need to assess whether the newly generated operating profits from AI businesses can exceed the capital costs required for infrastructure construction, as revenue growth does not necessarily equate to value creation. If a company needs to invest more capital to obtain one dollar of AI revenue, then the relevant business may still destroy shareholder value.
Ultimately, to determine whether the AI industry is mature, one needs to observe whether the entire value chain has formed a stable commercial closed loop and not just whether chip suppliers continue to earn high profits. If model companies, cloud platforms, and enterprise clients can create sustainable revenue and improve cash flow through AI, then the current capital expenditure may convert into long-term economic value; conversely, if downstream enterprises rely on continuous investment without forming returns over the long term, the sustainability of the entire investment cycle will be questioned.
8. Conclusion
Goldman Sachs’ judgment about $500 billion in AI capital expenditure reveals not just how much technology companies will invest but indicates that the global economy is entering a new investment cycle driven by computing power, data centers, and electricity infrastructure. Historical experience shows that significant technological revolutions often bring about genuine productivity increases while concurrently causing temporary capital misallocations. Railroads, the Internet, and optical fiber networks have all altered the world while costing many investors lacking commercial returns. Therefore, the greatest suspense around AI in the future is not whether technology can continue to advance, but which companies can convert technological advantages into sustainable revenue, profit, and cash flow.
In this sense, $500 billion could signify the starting point of a new cycle or become the first watershed in examining the commercial realization capabilities of the AI industry. Ultimately, what will determine the success or failure of this investment wave is not the scale of capital expenditure itself, but whether these投入 can genuinely create long-term economic value sufficient to cover capital costs.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
