Graphic and Text | Sister Tang

From July 24 to August 3, Microsoft rose by 27.76%, Amazon by 22.36%, Google by 16.82%, Meta fell by 0.83%, and the iShares Semiconductor ETF (SOXX) corrected by 3.67%; representative stocks in memory and flash memory still declined significantly, while stocks in optical communications had mixed performance. The most abnormal trend remains the continued rise of the three cloud providers procuring AI infrastructure, while the semiconductor ETF remains below its July 24 close.

According to the trading logic of the past two years, when cloud providers increase capital expenditures, it raises order and revenue expectations for semiconductor companies while suppressing their own free cash flow and profits. Microsoft simultaneously lowered its capital expenditure expectations, but all four companies indicated that investments in underlying computing power are still expanding. Under this premise, the three cloud providers continued to rise, while the semiconductor ETF still fell; although Meta remains slightly below its July 24 close, it has recovered most of its previous losses.

In our view, this divergence at least indicates that the market is placing higher weight on the monetization progress of newly added computing power. Who pays for this computing power and how they pay determines when investments can be converted into revenue and cash flow.

However, the four companies rarely disclose AI revenue separately, and existing metrics cannot distinguish whether a dollar of computing power serves in-house operations, cloud clients, or AI applications. Given the limited disclosure rules and transparency, earnings calls, financial report footnotes, and management commentary have become more important windows for observation. How newly added computing power is prepared for use and how returns are generated often first appear in these written disclosures.

In Microsoft's Q4 FY2026, as well as in the official financial reports, earnings calls, and supplementary materials for Q2 2026 from Google, Amazon, and Meta, Meta divided the returns on AI infrastructure into in-house use, external sales of computing power, and sales of intelligence. Google announced that its TPU system has been delivered to customer data centers, Amazon began discussing selling Trainium chips or racks outside of AWS, and Microsoft kept its self-developed chips within Azure, charging through cloud services and applications.

Some have already generated revenue, while others remain in management discussions.

Another layer of change comes from Microsoft's emphasis on CPUs. Nadella stated that the importance of CPUs in agent runtime is no less than that of GPUs; after agents enter production environments, AI infrastructure demand is no longer limited to accelerators for training but also includes CPUs, databases, storage, and networking. This is not about CPUs replacing GPUs but about AI workloads expanding the range of infrastructure that needs to be procured.

This also raises two questions. Why does a one-dollar AI investment generate different revenue and cash flow across the four companies? And why doesn't the new demand brought by agents fall evenly across all hardware companies?

01 Three Paths to Monetize Computing Power

Zuckerberg divided the returns on AI infrastructure into three categories during the earnings call: in-house use by the company, external sales of computing power, and sales of intelligence on top of computing power. When selling computing power externally, customers purchase computing resources and choose their own models and applications; when selling intelligence, customers purchase model calls, agent services, or business outcomes. The former is usually charged by capacity or usage duration, while the latter is charged by calls, seats, or results.

For Meta, the returns from in-house use of computing power are first reflected in existing businesses like advertising. Q2 advertising revenue reached $59.4 billion, up 27% year-over-year; AI advertising tool Advantage+'s annualized revenue exceeded $75 billion, with model trials showing an 8.3% increase in Facebook click-through rates and a 15.7% increase in conversion rates.

However, the company did not disclose the number of paying enterprises, external sales contracts, or corresponding revenue, so Meta is currently validating internal advertising returns, with external commercialization remaining an optional direction proposed in earnings calls. The Business Agent product for enterprises has over 1 million weekly active businesses, and management also discussed APIs, external sales of computing power, and results-based charging, stating that external computing power quotes are significantly higher than costs and that profit margins from selling intelligence may be higher.

Google has confirmed hardware external sales revenue from its self-developed chips. In Q1, management only said that hardware agreements for its self-developed AI chip TPU had entered contract reserves; by Q2, TPU systems had been delivered to customers' own data centers, with the delivered portion beginning to be counted in Google Cloud revenue, while related unfulfilled agreements remained in contract reserves, with most related revenue expected to be recognized in 2027.

Changes in cloud revenue and inventory in the financial reports provide further validation. Q2 Google Cloud revenue reached $24.8 billion, up 82% year-over-year; inventory rose from $2.439 billion at the end of 2025 to $9.991 billion at the end of June this year, with management explaining that inventory is being prepared for subsequent TPU system deliveries. However, the company did not disclose TPU revenue or profit margins, only stating that cloud revenue excluding TPUs still accelerated significantly.

Although Amazon also has a self-developed chip business, it has not yet reached Google's stage. Trainium is Amazon's self-developed AI training chip, currently provided to customers primarily through AWS cloud services. During the Q2 earnings call, management stated that customers want to deploy Trainium outside of AWS, leading the company to begin discussing selling chips or entire racks separately.

Amazon's currently confirmable revenue remains from cloud services formed by its self-developed chips through AWS. Q2 AWS revenue reached $42.2 billion, up 36.7% year-over-year; the annualized revenue of its self-developed chip business, including Trainium, calculated at the then-current operating speed, increased from over $20 billion in Q1 to over $25 billion in Q2.

Microsoft's approach differs; it has both backend cloud infrastructure like Azure and frontend entry points like Windows and Microsoft 365, allowing it to directly connect computing power with applications. Its self-developed AI chip Maia and self-developed CPU Cobalt are both deployed within Azure, with no independent external sales signals; financial reports show Azure revenue up 43% year-over-year, with backend computing power already generating revenue through cloud services.

Microsoft can also charge subscription and usage fees through frontend applications. Microsoft 365 Copilot has over 30 million paid seats, Dynamics consumer revenue grew 4x quarter-over-quarter, and GitHub Copilot saw revenue grow 60% quarter-over-quarter after switching to usage-based billing.

Thus, the difference among the four companies lies in how far their computing power revenue has progressed. Meta is currently realizing advertising returns, with external sales of computing power and intelligence remaining optional directions; Google has confirmed TPU hardware revenue; Amazon's self-developed chips still primarily generate revenue through AWS, with external cloud sales not yet confirmed; Microsoft generates revenue from both Azure cloud services and frontend applications.

The same batch of earnings calls also revealed another common change: when discussing agents, the four companies repeatedly mentioned CPUs, databases, storage, and networking.

02 Agents Expand Infrastructure Demand

The reason the four companies repeatedly mention these elements is that after agents enter production environments, a single task typically involves model inference, tool calls, enterprise data read/write, and cross-system transmission, each of which requires different infrastructure.

Among these, tool calls increase CPU load. All four companies are adding general-purpose computing capabilities for agents. Microsoft stated during the earnings call that agent runtime demand for CPUs is no less than that for GPUs, with racks equipped with its self-developed server CPU Cobalt 200 already deployed in over 25 data centers. Google directly launched its self-developed CPU Axion optimized for agents.

Amazon further clarified which tasks would use CPUs, stating that model training optimization, reinforcement learning, and some agent tool calls would run on CPUs; its Q1 self-developed chip announcement also disclosed that Meta committed to using tens of millions of AWS self-developed server CPU Graviton cores.

When agents call tools, they also need to read data from enterprise systems and write results back, increasing database and storage usage. In Amazon's Q2 AWS management commentary, Andy Jassy stated that customers also need storage and vector databases for semantic retrieval and want inference services close to existing applications and data.

Microsoft's earnings call further disclosed that customers using both its AI development platform Foundry and PostgreSQL database services grew 80% year-over-year.

As tasks flow among CPUs, enterprise data, and accelerators, networking becomes a condition for expansion. Google's Q2 earnings call disclosed that Virgo is a network system connecting AI accelerators distributed across different data centers, aiming to connect up to 1 million accelerators; model API call volume increased from 16 billion tokens per minute last quarter to 22 billion.

New demand has extended from chips to data center connectivity and capacity sources. The increase in API calls indicates growing model workloads, while Virgo's planning shows that Google is simultaneously expanding network capacity; the more model calls there are, the more cross-data center computing power scheduling and data transmission will increase. Meta also described the same change, stating that when arranging capacity, it considers computing, storage, networking, and energy together and configures them among self-built, leased, and third-party clouds.

Viewing the two quarters together, the focus of the four companies' discussions has shifted from capacity shortages and delivery plans to CPU deployment, database consumption, API usage, and storage and networking procurement.

These demands have already begun to translate into specific procurement, generating revenue for some upstream vendors. Amazon cited storage price increases as one reason for raising capital expenditures and mentioned price hikes for mechanical hard drives (HDDs) and solid-state drives (SSDs) during the earnings call; the company also signed a multi-year, billion-dollar agreement with Corning to procure fiber optic cables and connectivity products.

The landing points for new revenue vary; mechanical hard drives, solid-state drives, NAND flash memory, and DRAM memory occupy different positions, and Corning's contract corresponds to fiber optic cables, not high-speed optical transceivers. From July 24 to August 3, Micron and SanDisk fell by 9.93% and 10.34%, respectively; among the two HDD manufacturers, Western Digital rose by 1.43%, while Seagate fell by 2.42%; representative U.S. optical communication stocks had mixed performance.

Comparing the four companies, capital expenditures must be broken down into at least six items: equipment quantity, procurement price, accounting classification, asset lifespan, financing method, and computing power usage. Quantity and price affect upstream orders, accounting classification and asset lifespan affect reported amounts, financing methods affect cash flow, and computing power usage determines whether returns come from internal operations or external sales.

Microsoft's capital expenditure expectations have indeed decreased. The company extended the estimated useful life of data centers and office buildings from 15 to 25 years, causing more future data center leases to shift from finance leases to operating leases; the former are counted in capital expenditures, while the latter are not, reducing the 2026 calendar year expectation from about $190 billion to about $175 billion. Management simultaneously emphasized that excluding this classification change, the original investment expectations remain unchanged.

Amazon raised its full-year cash capital expenditures to about $220 billion, with increased equipment procurement and storage price hikes jointly driving up the total, corresponding to equipment quantity and procurement price, respectively. Google's Q2 capital expenditures were $44.9 billion, with about 60% for servers and 40% for data centers and networking; the company remains capacity-constrained, so it plans to expand the use of third-party capacity in Q3.

In terms of capacity construction and financing methods, Meta simultaneously uses self-built data centers, third-party clouds, debt, and joint ventures, while the other three companies also adopt cash purchases, leasing, or third-party capacity.

03 Conclusion

The focus of competition among the four companies in the next stage has shifted from how much computing power they possess to how newly added computing power is allocated. It can serve internal operations, cloud clients, hardware clients, or agent applications, and companies that can adjust capacity across different revenue paths are more likely to improve utilization and generate revenue faster.

New gaps will emerge in demand forecasting, capacity scheduling, and pricing capabilities. The first two determine how much to buy and where computing power is used; pricing methods determine whether revenue ultimately appears as hardware, cloud services, seats, or business outcomes.

In the coming quarters, external chip clients, CPU deployment, database consumption, storage procurement, and network usage will validate whether these paths continue to expand; Microsoft's seat and usage-based charging and Meta's results-based charging will test whether application layers can generate independent revenue.

All revenue must ultimately cover depreciation, leasing, debt, and procurement commitments. Contract terms and booking levels affect revenue visibility, while charging methods determine how computing power usage enters revenue. Capital expenditures will continue to expand, but what truly differentiates the four companies is the speed at which they can convert computing power into revenue and then into free cash flow.

Disclaimer: This article is for learning and communication purposes only and does not constitute investment advice.

Welcome to like, share, and repost. Your support is what keeps us updating!