AI CapEx, Token Costs, and Meta’s Compute Monetization: Revaluing the Three Cloud Giants — Microsoft, Amazon, and Google

Abstract

Since 2026, capital expenditures around artificial intelligence infrastructure have continued to rise, and the market's focus on large technology companies has shifted accordingly. In the past, investors were more inclined to view AI investment as a long-term growth asset; now, the core question is gradually turning to: can these capital expenditures be translated into sufficiently high revenue, gross profit, free cash flow, and return on capital?

The recent share price pressure on the three major cloud providers—Microsoft, Amazon, and Alphabet/Google—reflects surface-level concerns about AI capital expenditures, free cash flow, and token costs. However, a deeper industrial shift may be that enterprise AI is moving from "extensive trial" to an "epoch of precise usage calculation." In this stage, enterprises no longer call the most expensive models without limit, but rather manage models, computing power, data, security, and billing more finely based on task value, cost, success rate, latency, and compliance requirements.

This does not necessarily mean a collapse in AI demand. Instead, it may mean that AI is moving from the experimental phase to the formal business deployment phase. For single-model companies, enterprise control of token costs may bring pressure; but for multi-model cloud platforms such as Azure, AWS, and Google Cloud, precise usage calculation may instead strengthen their value as an enterprise AI cost control layer, model routing layer, data governance layer, and billing layer.

I. Core reasons for the market's repricing of the three major clouds

The recent pressure on stocks related to the three major clouds cannot be simply understood as the market rejecting long-term AI demand. More accurately, the market is shifting its focus from "is there demand for AI" to "how much capital is needed to realize AI demand." This is a change in valuation logic. Previously, AI capital expenditures were often viewed as growth investments; but as investments in data centers, GPUs, networking, power, cooling, and storage expand rapidly, investors are beginning to question whether these expenditures are high-return growth assets or a defensive arms race that tech giants are forced to participate in.

Microsoft's data provides a visual representation of this capital expenditure acceleration. Microsoft's property and equipment expenditures in the third quarter of fiscal year 2026 were $30.876 billion, compared with $16.745 billion in the same period last year; for the first nine months of fiscal year 2026, this expenditure was $80.146 billion, compared with $47.472 billion in the same period last year. This change illustrates that AI infrastructure construction has upgraded from normal cloud capital expenditures to a higher-intensity investment cycle.

Amazon's cash flow structure also shows capital expenditure pressure. Amazon's AWS sales in the first quarter of 2026 increased by 28% year-over-year to $37.6 billion, and AWS operating income was $14.2 billion, indicating that the cloud business itself still maintains strong growth; however, the company's operating cash flow for the trailing twelve months was $148.5 billion, and free cash flow fell to $1.2 billion, mainly due to a significant increase in property and equipment purchases, which the company explicitly stated mainly reflected AI investments.

Alphabet/Google also exhibits similar characteristics. Alphabet's Google Cloud revenue in the first quarter of 2026 grew 63% to $20 billion, a significant acceleration in cloud business growth; however, in the same quarter, expenditures for the purchase of property and equipment reached $35.674 billion, and free cash flow was $10.116 billion. Google's cash flow has not deteriorated to an uncontrollable state, but the slope of capital expenditures has clearly steepened.

Therefore, what the market is truly worried about is not that the three major clouds have no demand, but the capital constraints brought by the strong demand itself. If AI infrastructure investment continues to rise, but the corresponding revenue and cash flow conversion speed is less than expected, valuation multiples will naturally be compressed.

II. AI costs have entered financial reports, not just staying at the narrative layer

One of the biggest differences between AI products and traditional software subscription services lies in the cost structure. The marginal cost of traditional Office or enterprise software is low; a user writing an extra email, making an extra spreadsheet, or generating an extra document does not significantly increase the software company's unit cost. But AI products are different. Every model call, every inference, and every AI agent execution corresponds to computing power, storage, networking, and model service costs.

Microsoft has already disclosed this change in its financial reports. In the third quarter of fiscal year 2026, Microsoft Cloud's gross margin fell to 66%, driven by continued investments in AI infrastructure and growth in AI product usage, partially offset by efficiency improvements in Azure and Microsoft 365 Commercial Cloud. In other words, AI usage is no longer just a story on the revenue side, but has also entered the cost side.

In the Productivity and Business Processes segment, Microsoft disclosed a $680 million increase in cost of revenue, primarily driven by AI infrastructure investments supporting Microsoft 365 Copilot seats and usage growth. This does not mean Copilot has damaged Microsoft's fundamentals; on the contrary, it illustrates that the gross margin structure of AI software may be lower than traditional Office SaaS, but if the company can redesign the billing model through "seats + usage," there is still room to manage gross margin pressure.

The implications of this change for investors are important. AI products cannot be valued solely based on traditional software logic, but must also be observed through revenue growth, usage intensity, inference cost, model routing efficiency, caching ratios, and usage billing capabilities. What truly deserves tracking in the future is not just the number of users of AI products, but whether unit usage can form a sustainable gross margin.

III. Enterprises begin to control tokens: collapse in demand or precise usage calculation?

Enterprises starting to control AI usage is another main line of recent market concern. However, this phenomenon should not be directly equated with the disappearance of demand. A more reasonable explanation is that enterprises are moving from "use AI whenever possible" to a precise usage calculation phase of "is every AI use worth it."

The case of Uber is representative. According to media reports, after encouraging employees to use AI programming tools, Uber exhausted its full-year AI budget within four months, subsequently setting a cap of $1,500 per employee, per tool, per month for agentic coding tools like Claude Code and Cursor. This case does not show that AI tools have no value, but rather illustrates that when tools are useful enough and diffuse too quickly, enterprise budgets can rapidly spiral out of control.

The Gemini computing power constraint between Google and Meta also shows that high-end computing power remains in tight supply. According to reports cited by the media, Google informed Meta that it could not meet all of the Gemini computing capacity Meta required, and Meta consequently encouraged its employees to use AI tokens more efficiently. This report has not been independently verified by Reuters and should therefore be treated as a media report rather than an officially confirmed fact; however, the trend it reflects is clear: even as big tech companies continue to build AI infrastructure in large volumes, the supply of high-quality computing power may still fail to keep up with demand.

Microsoft's internal adjustment of Claude Code licenses also reflects the same logic. The Verge reported that Microsoft plans to remove most Claude Code licenses and push many developers to turn to GitHub Copilot CLI; the report stated that this adjustment was related to both tool integration and financial factors. This case illustrates that even within large technology companies, the use of AI tools has begun to be constrained by costs, product strategy, and ecosystem control.

Some of AWS's pricing changes reflect the scarcity of AI computing power from the supply side. According to Business Insider, AWS raised prices for EC2 Capacity Blocks for ML by approximately 20%, following an approximate 15% price increase in January. It should be emphasized that this is not a comprehensive price hike across AWS, but rather a price adjustment for specific AI/ML capacity reservation services. Its implication is that high-end AI computing power still possesses scarcity, and leading cloud providers maintain certain pricing power on some capacity products.

These cases together point to one judgment: enterprises are not stopping the use of AI, but are starting to manage AI usage. Token costs, computing budgets, task values, and usage permissions are becoming common concerns for enterprise IT and finance departments. This is precisely the industrial foundation of the "AI precise usage calculation epoch."

IV. Reports of Meta selling computing power: the supply side also begins to calculate ROI

Recent reports of Meta planning a cloud business and selling excess AI computing power add supply-side evidence to this logic. Reuters, citing Bloomberg, reported that Meta is establishing a cloud business to sell excess AI computing capacity; the plan is still developing and strategies may change, with Reuters also stating it could not independently verify the report. Therefore, this information should be viewed as a media report and a signal to be observed, rather than a completed official strategic disclosure.

The significance of this event lies not in whether Meta will immediately become a direct replacement for AWS, Azure, or Google Cloud, but in that it illustrates that large tech companies are also starting to consider how to turn computing power into an externally billable asset after huge investments in AI infrastructure. The Reuters report also mentioned that analysts believe Meta's increased supply of computing power may have a greater impact on neocloud companies like CoreWeave and Nebius than on major hyperscalers, because these companies partially rely on Meta's demand growth; the report also pointed out that Meta's AI infrastructure expenditures in 2026 could be as high as $145 billion.

From an industrial perspective, Meta's case illustrates that AI computing power is being assetized. Computing power is no longer just an internal R&D investment, but may also become an infrastructure asset that is leased, priced, and managed. However, those truly capable of capturing value long-term are not just "who owns computing power," but those who can integrate computing power, models, data, security, billing, and enterprise workflows together. This is also a key difference between the three major clouds and single computing power leasing providers.

V. Precise AI usage calculation epoch: from extensive trial to formal business deployment

The "AI precise usage calculation epoch" can be understood as the second phase of enterprise AI commercialization. The first phase is the extensive trial phase, where the core question for enterprises is whether they can connect AI to products and processes. In this stage, AI features are quickly embedded into software, search, office, customer service, and programming tools, and the market focuses more on usage, narrative, and technical feasibility.

The second phase is the precise usage calculation phase. Enterprises begin to ask: is every AI call worth it? Do all tasks require frontier models? Which tasks can use cheaper models? What content can be cached? Which tasks are suitable for batch processing? Which AI uses should be included in seat fees, and which should be billed based on usage? This indicates that AI is moving from product demonstrations to formal business deployments, and enterprises are starting to incorporate it into budgets, compliance, and cost-allocation systems.

The third phase is cost-per-task optimization. In the AI era, one cannot just look at the price per million tokens, because what enterprises are actually buying is not the tokens themselves, but the effective results within a certain workflow. For example, in programming, a task can be fixing a code bug or completing a code modification accepted by the team; in customer service, a task can be resolving a ticket; in financial analysis, a task can be completing a summary that can be adopted by analysts. What truly matters is the total cost to achieve an effective result—that is, model costs, tool costs, retry costs, and manual review costs divided by the success rate.

The fourth phase is the pay-for-outcome phase. Enterprises are ultimately more likely to want to pay for outcomes rather than purely for tokens. For example, billing based on resolved customer service tickets, completed compliance reviews, qualified sales leads generated, or completed software features. If AI commercialization enters this stage, value will shift from "consuming more tokens" to "generating more effective results at lower costs."

VI. Why precise usage calculation may strengthen the value of the three major cloud platforms

If enterprises only pursue the strongest models, value may concentrate in the hands of single-model companies. But when enterprises start to pursue completing the most effective tasks at the lowest cost, value will flow more to multi-model cloud platforms. The reason is that enterprises formally deploying AI do not just need the model itself; they also need permission management, data boundaries, model selection, cost monitoring, billing allocation, latency control, compliance reviews, and business workflow integration.

The common advantage of Microsoft, Amazon, and Google is that they do not just lease GPUs or sell a single model, but provide an enterprise AI runtime environment. Azure, AWS, and Google Cloud respectively possess computing, storage, networking, database, security, identity management, model marketplaces, model routing, cost monitoring, billing, compliance, and enterprise sales channels. These capabilities make them look more like enterprise AI platforms rather than pure computing power suppliers in the AI precise usage calculation epoch.

Cloud providers are productizing this capability. Microsoft Azure AI Foundry's model router is designed to optimize cost and latency, routing simple tasks to smaller, cheaper models and complex tasks to stronger models while maintaining similar quality. AWS Bedrock Intelligent Prompt Routing can select models within the same model family based on each request, which AWS officially claims can reduce costs by up to approximately 30% without sacrificing accuracy. Google Cloud's Model Garden offers more than 200 models from Google and partners, supporting customers to discover, customize, and deploy models on the same platform.

This is also why reports of Meta selling AI computing power do not necessarily constitute a direct negative for the three major clouds. In the short term, it may increase the supply in the AI computing power market, posing competition especially to some companies that only do computing power leasing; but in the longer term, it instead proves that computing power is becoming a leasable and priceable infrastructure asset. What truly matters in the long run is not single computing power supply, but whether the platform can integrate computing power, models, data, and enterprise workflows together.

VII. Different paths of the three major clouds

Microsoft's advantage lies in its entry point into enterprise workflows. Office, Teams, Outlook, Excel, PowerPoint, GitHub, Dynamics, Security, and Azure together constitute the daily working environment of enterprises. It is difficult for Anthropic or other model companies to directly disrupt Microsoft's core fundamentals in the short term, because enterprise workflows, identity systems, permissions, files, collaboration, and compliance systems have high migration costs. What Microsoft truly needs to manage is Copilot's gross margin structure. If Copilot relies entirely on fixed seat fees and allows unlimited high-cost calls, gross margin pressure will be significant; but if Microsoft can gradually advance the "seat + usage" model, passing high-cost AI usage to higher-value scenarios, its gross margin pressure may be manageable.

Amazon's advantage lies in AWS's positioning as a neutral, multi-model platform. AWS does not need to bet entirely on a single model; customers can use Claude, Amazon's self-developed Nova, Meta, Mistral, or other models in Bedrock. Even if some tasks are diverted from Claude to cheaper models, as long as the workload remains in the AWS ecosystem, Amazon can still capture infrastructure, data access, security governance, and platform service revenues. Media reports state that some terms of Amazon's partnership with Anthropic may shift from settling by computing time to something closer to settling by token usage; this information is not yet an official full disclosure and can only serve as an observational signal. If this direction holds, the cost of using Claude will rise more linearly with usage, which may prompt AWS to more actively promote model routing and the use of its self-developed models.

Google's path emphasizes the integration of models, chips, and cloud. Google possesses resources such as Gemini, TPUs, Google Cloud, Model Garden, search, and multimodal data. Alphabet's Google Cloud revenue in the first quarter of 2026 grew 63% to $20 billion, with cloud business growth accelerating significantly; the company also disclosed that Google Cloud backlog exceeded $460 billion. If the cost and performance advantages of Gemini and TPUs can translate into Google Cloud's enterprise share and high-quality cash flow, Google Cloud may also benefit in the AI precise usage calculation epoch.

VIII. Why Oracle is not the same kind of story

Oracle also benefits from the shortage of AI computing power, but its investment story is different from those of Microsoft, Amazon, and Google. The core of the three major clouds is the integration of models, computing power, data, security, billing, and enterprise workflows into a platform; Oracle's story is more like a massive AI backlog accompanied by higher capital expenditures, financing needs, and free cash flow pressures.

Oracle officially disclosed that its remaining performance obligations in fiscal year 2026 reached $638 billion, up 363% year-over-year; however, free cash flow for the same fiscal year was negative $23.7 billion as the company continued to invest to support the growth of its Cloud Infrastructure business. Oracle also disclosed that it had raised $43 billion in debt and $5 billion in equity in fiscal year 2026, and expects to raise approximately $40 billion more through debt and equity in fiscal year 2027.

Therefore, Oracle's existence reminds investors that market concerns about AI capital expenditures are not without foundation. It is just that this risk is more acute in AI infrastructure stories with high capital intensity, high financing needs, and high customer concentration, whereas in platform-type cloud companies like Microsoft, Amazon, and Google, risk and platform value need to be evaluated separately.

IX. Valuation implications: Are the three major clouds excessively discounted?

At the valuation level, the forward P/E ratio can serve as a rough but useful observation tool. Microsoft currently trades at about 19 times forward P/E, compared with a five-year average of nearly 30 times; Amazon currently trades at about 27 times, compared with a five-year average of about 45 times; Alphabet currently trades at about 24–25 times, compared with a five-year average of about 22 times. If AI capital expenditures ultimately turn out to be low-return defensive spending, then even if the three major clouds return to forward P/E ratios in the low twenties or even teens, they may not necessarily be cheap enough. Conversely, if precise AI usage calculation strengthens the cloud platforms' irreplaceable nature in enterprise AI systems, then current valuations may excessively reflect short-term cash flow pressures while underestimating future platform cash flows.

For investors, the key is not simply to judge whether the three major clouds are "cheap" or "expensive," but to judge whether their capital expenditures can be translated into long-term platform revenue. This requires continuous tracking of cloud revenue growth, AI product gross margins, capital expenditure growth rates, depreciation pressures, free cash flow recovery speeds, model routing efficiency, enterprise willingness to pay for AI, and usage billing structures.

X. Risk factors

First, the payback period for AI capital expenditures may be longer than the market expects. Data centers, GPUs, networks, power, and cooling facilities are all highly capital-intensive assets; if AI revenue growth falls short of expectations, depreciation and cash flow pressures may continue to suppress valuations.

Second, a decline in token prices may have a dual impact on unit economics. Lower token costs help diffuse AI usage, but if prices fall faster than usage growth and cost efficiency improvements, the gross profit generated per unit of computing power by cloud providers may come under pressure.

Third, enterprise AI adoption may fall short of optimistic assumptions. Enterprises controlling token usage, setting budget caps, and tightening permission management may imply that some AI applications have yet to demonstrate a sufficiently clear ROI. If AI agents fall short in success rates, stability, compliance, and controllability, the pace of formal business deployment may slow down.

Fourth, model companies and emerging compute providers may shift value distribution. OpenAI, Anthropic, Google, Meta, and emerging neoclouds are all competing for AI infrastructure and model API value. Rumors of Meta selling compute capacity suggest that the compute supply landscape may be changing, though its long-term impact remains to be seen.

Conclusion: The question is not whether there is demand for AI, but whether capital expenditure can be converted into platform cash flow.

The core issue currently facing the big three cloud providers is not whether AI demand exists, but whether AI capital expenditures can be translated into sustainable platform cash flow. The market's attitude toward AI investment is shifting from "the more investment, the better" to "can investment generate returns." This explains why the valuations of Microsoft, Amazon, and Google are beginning to be collectively influenced by AI capital expenditures, free cash flow, and token costs.

However, enterprises starting to control AI usage does not necessarily equate to a collapse in AI demand. It may also indicate that AI is transitioning from broad-based experimentation into formal business deployment. The formal deployment phase requires more complex model selection, data governance, permission controls, cost monitoring, billing systems, and business process integration, which are precisely the strengths of Azure, AWS, and Google Cloud.

Rumors of Meta selling compute capacity indicate that the supply side is also starting to seek returns on AI capital expenditures; the example of Oracle reminds the market that strong demand for AI infrastructure does not automatically translate into high-quality free cash flow. For investors, what truly needs to be understood is that different AI infrastructure companies bear capital expenditure risks in different ways. The investment value of Microsoft, Amazon, and Google ultimately depends on their ability to convert AI infrastructure investments into long-term platform cash flows in the enterprise AI era, rather than simply on the scale of short-term capital expenditures.