Chief Perspective: In-depth Research, Technology Tracing, Landscape Analysis, Policy Interpretation, Trend Forecasting
Sister Na's Observation: Sharp Comments on Industry Hotspots, Frontline Observations, Application Interpretation Optoelectronics Focus: Daily News Headlines, Disseminating Highly Influential Optoelectronic Industry Information Optoelectronics Date: Tuesday Night 8:00 PM (20:00-21:30) Video Channel Live Broadcast, Sharing Optoelectronic Communication Industry Trends
ICC News - As the AI investment race enters its third year, the market is no longer satisfied with the "arms race" narrative. Investors are beginning to ask the same question: Of these massive expenditures, how much can truly be converted into sustainable revenue?
We selected two key metrics—annual capital expenditure budget (CAPEX budget) and Remaining Performance Obligations (RPO)—to observe the four core cloud providers within the same coordinate system. RPO represents "committed future revenue," while the CAPEX budget represents "planned funds to fulfill these commitments." The latest RPO details disclosed by the four providers—including time distribution, contract duration, and customer composition—provide us with a more precise analytical dimension. The quality of RPO depends not only on the total scale but also on "when it is recognized, who drives it, and how long the duration is."
01. RPO Time Distribution: Who Faces Immediate Pressure and Who Has Long-Term Exposure
Data Source: Financial reports of respective cloud providers
Note 1: Google uses the "revenue backlog" metric; for Q226, it stands at $519.5 billion, of which $513.9 billion comes from Google Cloud.
Note 2: Oracle's fiscal year ends on May 31; therefore, the corresponding Q226 data in this table is actually as of May 31.
The time distribution reveals a significant difference: the "recognition pace" of RPO varies drastically among the providers.
Over 50% of Google's RPO is scheduled to be recognized within 24 months. Starting this quarter, it also includes short-term contracts of less than one year in its statistics, making it the highest in short-term conversion efficiency among the four. The average duration for Microsoft's commercial segment is only 2.3 years, with about 30% recognized within 12 months. The order turnover is fast—but this also means continuous signing of new orders is required to replenish the RPO pool.
Only 12% of Oracle's RPO is scheduled to be recognized in the next 12 months (approximately $76.6 billion), with 68% concentrated in years 2 to 5, and 20% beyond five years. This structure indicates that Oracle's RPO is a "long-term contract pool." While short-term reporting pressure is limited, the volume of long-term delivery commitments far exceeds what its current CAPEX budget can support.
Amazon's weighted average contract duration is as long as 6.4 years, the longest among the four. This is directly related to two mega long-term contracts recently signed: an expansion of $100 billion (8-year term) on top of the original $38 billion with OpenAI, and an expansion of over $100 billion (10-year term) with Anthropic. The ultra-long duration implies a slower revenue recognition pace, but it also represents extremely high customer stickiness and long-term revenue lock-in.
02. RPO/CAPEX Budget Matching: A Re-evaluation with the Introduction of the Time Dimension
With the introduction of time distribution, the degree of "commitment mismatch" previously measured by the total RPO/CAPEX budget ratio needs to be revised:
Oracle's short-term matching is better than what the total ratio indicates. Only about $76.6 billion needs to be recognized within 12 months, which is about 1.5 times the $50 billion CAPEX budget. Coupled with external leasing and chip procurement, the short-term delivery pressure is controllable. However, the real challenge lies in the long term: an average of about $108.5 billion needs to be recognized annually from year 2 to year 5, far exceeding the current CAPEX budget level. Oracle's RPO structure is characterized by "low short-term pressure and high long-term exposure."
Microsoft's duration structure determines a relatively high "turnover rate" for its RPO. The average 2.3-year duration for the commercial segment means fast order renewal, but it also implies that if the pace of new signings slows down, the RPO will decline relatively quickly. Recently, the RPO growth rate has remained in the single digits for three consecutive quarters, requiring attention to its subsequent signing pace.
Google's short-term matching is likely the best among the four. Over 50% of the RPO is recognized within 24 months. Including short-term contracts expands the RPO base, cloud business revenue grew 82% year-over-year, and TPU sales have begun to contribute revenue—these factors collectively point to strong order conversion capabilities.
Amazon's ultra-long duration deserves special attention. Behind the 6.4-year average duration are two mega contracts totaling over $200 billion with OpenAI and Anthropic. These contracts are deeply tied to AWS's self-developed Trainium chips (the contracts explicitly mention "includes contractual obligations related to the performance of AWS chips"), resulting in high customer migration costs and strong stickiness. However, the ultra-long duration also means a slower revenue recognition pace, and there is a possibility of contract term adjustments during execution.
03. The "Driving Engine" of RPO: Whose Orders Are More Sustainable
The driving forces behind the RPO growth of the four providers differ fundamentally:
Google's driving force is the most diversified. RPO growth stems more from demand for its own cloud services, and external TPU sales have opened up new revenue sources. The customer structure is dispersed, with the lowest reliance on a single customer among the four.
Amazon's growth is driven by mega long-term contracts, but the binding method is different. Contracts with Anthropic and OpenAI are explicitly tied to the performance of self-developed Trainium chips, meaning the core of the customer relationship is the "chip ecosystem" rather than mere "computing power leasing." This results in higher switching costs and stronger contract stability.
The growth of Microsoft and Oracle remains highly dependent on OpenAI. Approximately 45% of Microsoft's and 55% of Oracle's RPO is related to OpenAI. OpenAI's annual revenue is about $25 billion, and it recently announced an investment of $750 billion by 2030 to build its own AI infrastructure, with the first $20 billion data center project already landed in Georgia. The OpenAI-related portion of Microsoft's RPO faces potential risks of customer self-built substitution or order transfer; Oracle's risk exposure is even greater—once OpenAI's self-built computing power is gradually put into operation, its external procurement demand may marginally decline.
04. The "Fulfillment Cost" of CAPEX Budgets Is Rising
Even if budgets are executed as planned, the cost of fulfilling orders is climbing. Google's Q2 CAPEX reached $44.9 billion, with free cash flow turning negative to -$5.9 billion. Bill of Materials (BOM) inflation is becoming a theme for the second half of 2026; Amazon's multi-billion-dollar long-term fiber optic agreement with Corning is a typical case of locking in the supply chain. Whether CAPEX budgets can be executed at planned prices and paces is itself a variable that requires continuous verification.
05. Revised Core Conclusions
First, the quality of RPO varies fundamentally among the providers and cannot be generalized. Google has the highest short-term recognition ratio and the most dispersed customer structure; Amazon's ultra-long-term contracts are deeply tied to self-developed chips, featuring strong stickiness but a slow recognition pace; Microsoft has a short duration and fast turnover but is highly dependent on OpenAI; Oracle has controllable short-term pressure but massive long-term exposure.
Second, "commitment mismatch" needs to be viewed by duration. In the short term (12 months), the matching degree of RPO and CAPEX budgets for all four providers is within a controllable range. The real mismatch lies in the long term: Oracle needs to recognize an average of about $108.5 billion annually from year 2 to year 5, while its long-term CAPEX planning remains unclear.
Third, OpenAI remains the biggest variable, but the impact paths are diverging. For Microsoft and Oracle, OpenAI's self-built data centers pose a direct substitution risk; for Amazon, OpenAI is a customer for its Trainium chips, and self-building and procurement may coexist. Amazon's contract with OpenAI explicitly ties in chip performance, giving the customer relationship stronger technical stickiness.
Fourth, the core issue of the industrial chain has shifted from "how much is planned to be invested" to "whether and when commitments can be fulfilled." For the upstream optical communication supply chain, more attention should be paid to customers with high short-term RPO conversion rates, dispersed customer structures, and ties to the self-developed chip ecosystem (Google, Amazon). The path from budget to actual procurement for these customers is more direct and predictable. Orders highly dependent on a single highly leveraged customer require prudent assessment of the uncertainty in their long-term fulfillment.
Note: The core data and analytical framework of this article are sourced from the "Q2 2026 Global Optical Communication Market Analysis Report" by the ICC Stone Industry Research Institute. The report will be distributed in electronic format to member units of the ICC Stone Enterprise Alliance. Non-member units should contact ICC Stone staff for consultation and ordering. Contact: Ms. Feng, 15989420950 (WeChat same number).
-END-