Buying more GPUs is no longer enough: monetization progress diverges among the four major cloud providers

From July 24 to August 3, Microsoft rose 27.76%, Amazon gained 22.36%, Google climbed 16.82%, Meta fell 0.83%, and the iShares Semiconductor ETF (SOXX) pulled back 3.67%. Memory and flash memory representative stocks continued to decline notably, while optical communication stocks showed mixed performance. Most unusually, the three cloud providers investing heavily in AI infrastructure kept rising, yet the semiconductor ETF remained below its July 24 closing level.

Under the trading logic of the past two years, increased capital expenditures by cloud providers would boost order and revenue expectations for semiconductor companies while compressing their own free cash flow and profits. Microsoft simultaneously lowered its capital expenditure forecast, but all four companies stated that underlying compute investments are still expanding. Against this backdrop, the three cloud providers continued to rise while the semiconductor ETF kept falling; although Meta remains slightly below its July 24 closing price, it has already recovered most of its earlier losses.

In our view, this divergence at least indicates that the market is assigning greater weight to the pace at which new compute capacity is monetized.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-related revenue separately, and current reporting standards cannot distinguish whether a dollar of computing power serves their own operations, cloud customers, or AI applications. With disclosure rules and transparency still limited, earnings calls, financial statement footnotes, and management commentary have become increasingly important windows for observation. Details about who new computing capacity is intended for and how it will generate returns often appear first in these textual disclosures.

In Microsoft’s fourth quarter of fiscal year 2026, and the second quarters of 2026 for Google, Amazon, and Meta—as reflected in their official earnings reports, earnings calls, and accompanying commentary—Meta categorized returns from AI infrastructure into self-use, selling computing power externally, and selling intelligence; Google announced that its TPU systems had already been delivered to customer data centers; Amazon began discussing selling Trainium chips or racks outside AWS; and Microsoft kept its custom-designed chips within Azure, monetizing them through cloud services and applications.

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

Another shift comes from Microsoft’s emphasis on CPUs. Nadella stated that during Agent runtime, CPUs are no less important than GPUs; once Agents enter production environments, AI infrastructure demand expands beyond just training accelerators to include CPUs, databases, storage, and networking. This isn’t about CPUs replacing GPUs, but rather AI workloads broadening the range of infrastructure that needs to be procured.

This raises two questions: Why does the same dollar of AI investment generate different revenue and cash flows across the four companies? And why doesn’t the incremental demand driven by Agents distribute evenly across all hardware vendors?

01 Three monetization pathways for computing power

During the earnings call, Zuckerberg categorized returns from AI infrastructure into three types:namely, internal use by the company, selling computing power externally, and selling intelligence built on top of that computing power. When selling computing power externally, customers purchase compute resources and choose their own models and use cases; when selling intelligence, customers purchase model invocations, Agent services, or business outcomes.The former is typically billed based on capacity or usage duration, while the latter is charged per invocation, per seat, or based on results delivered.

For Meta, the returns from using its own computing power are first reflected in its existing businesses, such as advertising. Advertising revenue reached $59.4 billion in the second quarter, up 27% year-over-year; its AI-powered ad tool Advantage+ generated annualized revenue exceeding $75 billion, with model tests showing an 8.3% increase in Facebook click-through rates and a 15.7% increase in conversion rates.

However, the company has not disclosed the number of paying enterprises, external sales contracts, or related revenue. Thus, Meta has so far validated only internal advertising returns, while external commercialization remains an optional direction mentioned during earnings calls. Its business-facing agent product, Business Agent, now has over 1 million weekly active businesses. Management also discussed APIs, selling computing capacity externally, and outcome-based pricing, noting that external pricing for computing power is significantly above cost and that monetizing intelligence could yield even higher margins.

Google has already confirmed hardware sales revenue from its self-developed chips.In the first quarter, management only stated that hardware agreements for its self-developed AI chip, the TPU, had entered the contract backlog; by the second quarter, TPU systems had already been delivered to customers’ own data centers, with the delivered portion now included in Google Cloud revenue, while the remaining undelivered commitments stayed in the contract backlog. Most of the related revenue is expected to be recognized in 2027.

Cloud revenue and inventory changes reported in the financial statements provide further validation. Google Cloud revenue reached $24.8 billion in the second quarter, up 82% year-over-year; inventory rose from $2.439 billion at the end of 2025 to $9.991 billion as of the end of June this year. Management explained that it is building inventory in preparation for upcoming TPU system deliveries. However, the company has not disclosed TPU-specific revenue or margins, only noting that cloud revenue growth accelerated significantly even after excluding TPU-related contributions.

Although Amazon also has a self-developed chip business, it has not yet reached the stage Google has. Trainium is Amazon’s self-developed AI training chip, currently offered primarily through its AWS cloud services. During the second-quarter earnings call, management noted that customers expressed interest in deploying Trainium outside of AWS, prompting the company to begin discussing standalone chip or full-rack sales.

What Amazon can confirm so far is still the cloud service revenue generated via AWS from its self-developed chips. AWS revenue totaled $42.2 billion in the second quarter, up 36.7% year-over-year; annualized revenue from its self-developed chip business—including Trainium—calculated based on utilization rates during the quarter, increased from over $20 billion in Q1 to over $25 billion in Q2.

Microsoft takes a different approach,as it possesses both backend cloud infrastructure like Azure and front-end entry points such as Windows and Microsoft 365, enabling it to directly link computing power with applications.Its self-developed AI chip Maia and CPU Cobalt are both deployed internally within Azure, with no signals of independent external sales; financial results show Azure revenue grew 43% year-over-year, indicating that backend computing power is already generating revenue through cloud services.

Microsoft can also generate subscription and usage-based revenue through front-end applications. Paid seats for Microsoft 365 Copilot have exceeded 30 million, Dynamics consumption revenue grew fourfold quarter-over-quarter, and GitHub Copilot’s revenue rose 60% quarter-over-quarter after shifting to usage-based billing.

Viewed this way, the key difference among the four companies lies in how far they’ve progressed in monetizing computing power.Meta is currently realizing returns primarily from advertising, while selling computing capacity externally and offering AI models remain optional paths. Google has already confirmed TPU hardware revenue. Amazon’s in-house chips are still primarily monetized through AWS, with external sales outside the cloud yet to be confirmed. Microsoft, meanwhile, generates revenue both from Azure cloud services and front-end applications.

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

02 Agents drive increased infrastructure demand

The reason all four companies emphasized these components is that once Agents enter production environments, a single task typically involves model inference, tool invocation, reading from and writing to enterprise data systems, and cross-system data transmission—each step requiring different infrastructure resources.

Tool invocation increases CPU load. All four companies are enhancing general-purpose computing capacity for Agents. Microsoft stated during its earnings call that CPU demand during Agent execution is no less than GPU demand, and server racks equipped with its in-house Cobalt 200 CPUs have already been deployed across more than 25 data centers. Google, meanwhile, directly launched Axion, its in-house CPU optimized specifically for Agents.

Amazon further clarified which tasks run on CPUs: post-training model optimization, reinforcement learning, and certain Agent tool invocations operate on CPUs. Additionally, Amazon’s Q1 disclosure on in-house chips revealed that Meta has committed to using tens of millions of AWS Graviton CPU cores.

When Agents invoke tools, they also read data from enterprise systems and write results back, increasing usage of databases and storage. In AWS’s Q2 management commentary, Andy Jassy noted that customers also need storage and vector databases for semantic retrieval, and prefer inference services located close to their existing applications and data.

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

Tasks flow among CPUs, enterprise data, and accelerators, making networking a key factor for scaling. At its Q2 earnings call, Google disclosed that Virgo is a networking system designed to connect AI accelerators distributed across different data centers, with the goal of linking up to one million accelerators. Meanwhile, model API token usage rose from 16 billion per minute last quarter to 22 billion.

New demand has expanded beyond chips to include data center interconnectivity and capacity sourcing. Rising API usage indicates increasing model workloads, while Virgo’s roadmap shows Google is simultaneously scaling network capacity—as model calls grow, so too will cross-data-center compute orchestration and data transfers. Meta described a similar shift, noting that when planning capacity, it considers compute, storage, networking, and power together, and allocates resources across self-built facilities, leased space, and third-party cloud providers.

Viewing the two quarters together, the focus of discussions among the four companies has evolved from concerns about insufficient capacity and delivery timelines to CPU deployments, database consumption, API usage, and procurement of storage and networking infrastructure.

These requirements are already translating into concrete procurement orders, generating revenue for some upstream suppliers. Amazon cited memory price increases as one reason for raising its capital expenditure guidance and mentioned rising prices for both hard disk drives (HDDs) and solid-state drives (SSDs) during its earnings call. The company also signed a multi-year, multi-billion-dollar agreement with Corning to purchase fiber optics, cables, and connectivity products.

The revenue impact varies across segments: HDDs, SSDs, NAND flash, and DRAM operate at different points in the supply chain, and Corning’s contract pertains specifically to fiber optic cabling rather than high-speed optical transceivers. Between July 24 and August 3, Micron and SanDisk fell by 9.93% and 10.34%, respectively; among HDD makers, Western Digital rose 1.43%, while Seagate declined 2.42%; U.S. optical communications stocks showed mixed performance.

Comparing the four companies, capital expenditures must be broken down into at least six components: equipment volume, procurement pricing, accounting classification, asset useful life, financing method, and intended compute application.Volume and pricing affect upstream orders; accounting classification and asset life impact reported financial figures; financing methods influence cash flow; and the intended use of compute capacity determines whether returns come from internal operations or external sales.

Microsoft has indeed lowered its projected capital expenditure figure. The company extended the estimated useful life of its data centers and office buildings from 15 to 25 years, reclassifying more future data center leases from finance leases to operating leases—only the former counts toward capital expenditures. As a result, its fiscal year 2026 capital expenditure outlook was reduced from approximately $190 billion to around $175 billion. Management emphasized that excluding this accounting reclassification, its original investment plans remain unchanged.

Amazon raised its full-year cash capital expenditures to approximately $220 billion, driven by increased equipment purchases and higher memory prices—reflecting changes in both equipment volume and procurement pricing. Google reported $44.9 billion in capital expenditures for the quarter, with roughly 60% allocated to servers and 40% to data centers and networking. Still constrained by capacity limits, the company plans to expand its use of third-party capacity in the third quarter.

In terms of capacity deployment and financing approaches, Meta utilizes a mix of self-built data centers, third-party cloud services, debt financing, and joint ventures. The other three companies also employ varying combinations of cash purchases, leasing, and third-party capacity.

03 Conclusion

In the next phase of competition among the four companies, the focus has shifted from how much computing power they possess to how newly added computing capacity is allocated.It can serve internal business needs, cloud customers, hardware clients, or agent applications. Companies that can flexibly reallocate capacity across different revenue streams are better positioned to improve utilization and generate revenue more quickly.

New gaps will emerge in demand forecasting, capacity scheduling, and pricing capabilities. The first two determine how much to purchase and where to deploy computing power; pricing models determine whether revenue ultimately materializes as hardware sales, cloud services, seat licenses, or business outcomes.

Over the coming quarters, external chip customers, CPU deployments, database consumption, storage procurement, and network usage will test whether these pathways continue to expand. Microsoft’s seat-and-usage-based billing and Meta’s outcome-based pricing will validate whether the application layer can generate standalone revenue.

All revenue must ultimately cover depreciation, leases, debt, and purchase commitments. Contract duration and booking levels affect revenue visibility, while billing models determine how computing utilization translates into revenue.Capital expenditures will continue to expand. What will truly differentiate the four companies is how quickly they can convert computing power into revenue—and then turn that revenue into free cash flow. $Microsoft (MSFT.US)$$Alphabet-C (GOOG.US)$$Amazon (AMZN.US)$$Meta Platforms (META.US)$

Disclaimer: This article is for educational and discussion purposes only and does not constitute investment advice.

156K Views