Is the AI infrastructure investment cycle peaking? Oracle RPO of $638 billion reveals that demand is still accelerating.

On June 10, 2026, Oracle released a financial report that caused Wall Street to re-examine the logic behind the AI investment cycle. In FY2026 Q4, this traditional database giant achieved total revenue of $19.2 billion, up 21% year over year, and cloud infrastructure revenue surged 93% year over year to approximately $5.2 billion. But the impact of these numbers pales in comparison to another metric—remaining performance obligations (RPO). Oracle’s RPO reached $638 billion at the end of Q4, up 363% year over year. In this quarter alone, new AI infrastructure contracts signed totaled as much as $67 billion.

Just a year ago, Oracle’s RPO was still at $138 billion. In less than 12 months, it grew nearly fourfold. This figure far exceeded analysts’ expectations in the range of about $180 billion, forcing the market to reassess the true sustainability of the AI infrastructure investment cycle and the depth of its transmission.

This is not an isolated case. The world’s five largest hyperscale cloud providers—Amazon, Microsoft, Alphabet, Meta, and Oracle—had combined capital expenditures of about $750 billion in 2026, up about 67% year over year and representing the third consecutive year in which growth exceeded 60%. CreditSights noted in its latest report that the capital expenditure-to-revenue ratios of these companies have climbed to unprecedented levels: Oracle at about 86%, Meta at about 54%, Microsoft at about 47%, Alphabet at about 46%, and Amazon at about 25%.

The relationship between capital expenditure and the AI infrastructure investment cycle has been discussed repeatedly over the past year. One core controversy has always remained: will this huge spending translate into recurring revenue, or will it turn into a bubble of excess capacity? Oracle’s data provides a key validation signal—the sharp rise in RPO indicates that the actual pace of contract signing on the demand side has not slowed down and is even accelerating.

Cloud Infrastructure: The First Conversion Layer of AI Compute Power Demand

The most noteworthy detail in Oracle’s financial report is not only the RPO figure itself, but its composition. Of the $638 billion RPO, about 12% is expected to recognize revenue within 12 months, and about 34% between 13 and 36 months. This means that over the next 1 to 3 years, Oracle will gradually recognize roughly $220 billion to $290 billion in committed additional revenue in its income statement—substantial growth support for a company with annual revenue of about $67 billion.

Meanwhile, Oracle’s global GPU utilization rate reached 97.5% in Q4. This figure suggests that, at least at the current stage, the assumption that “overbuilding has not yet been put to use” is not valid. Between supply-side expansion and demand-side consumption, there is a structural time lag—this is a typical feature of the AI infrastructure investment cycle: physical assets are put in place first, and then capacity is gradually converted into revenue through contracts.

Oracle is not the only beneficiary. Alphabet’s Google Cloud revenue surged 63% in Q1, and its cloud backlog orders have exceeded $460 billion. Microsoft’s AI business achieved annualized revenue of $37 billion, up 123%. Amazon AWS’s annualized revenue is about $150 billion, with growth of 28%. Taken together, these numbers point to a judgment: AI cloud services are undergoing a critical transition from pilot phase to large-scale deployment, and Oracle’s RPO data is one of the most direct confirmations of this trend.

From the perspective of investment transmission, cloud infrastructure service providers are the first layer to benefit from AI capital expenditure. They convert capital investment into externally delivered service capabilities, and then lock in future revenue streams through contracts. The smoothness of this chain will directly affect the sustainability of the entire AI infrastructure investment cycle.

Chips and Advanced Packaging: The Core Capturing Layer of Capital Expenditure

The most direct recipient of physical construction of AI data centers is the semiconductor industry chain. In a report released in June 2026, Omdia expects that global AI infrastructure spending will exceed $600 billion that year, with a significant portion flowing to GPU clusters, custom accelerators, and data center compute core components.

Nvidia is the most representative player in this chain. In Q4 2026, Nvidia’s data center revenue reached $62.3 billion, up 75% year over year, and its networking business revenue surged 263%. But more importantly, Nvidia’s ecosystem penetration is extending into broader application scenarios. At WWDC in June 2026, Apple, together with Nvidia and Google, announced that Apple will use Nvidia Blackwell GPUs in Google Cloud to support server-side inference for Apple Intelligence—marking a key customer breakthrough for Nvidia in the incremental market for secure AI inference.

The benefits in the chip industry chain extend well beyond Nvidia. TSMC, which dominates about 70% of the global chip foundry market, benefits from every stage of physical manufacturing of AI chips. Advanced packaging is another critical bottleneck. BE Semiconductor (Besi) and ASMPT hold the core equipment for advanced packaging of AI chips. ASMPT’s research report shows that as long as the long-term trend for global AI data centers and AI PCs/phones remains unchanged, demand for high-end equipment will have rigid support. UBS expects that in 2026, memory semiconductor revenue will reach about $961 billion, and the DRAM market will continue to expand driven by AI training and inference demand.

A noteworthy signal is that benefits from the AI chip industry chain are spreading from Nvidia to a broader supply chain. However, because upstream design and materials are highly concentrated, risks are also concentrated at a small number of key nodes—including supply bottlenecks, changes in geopolitical policy, and overly high customer concentration.

Power and Infrastructure Bottlenecks: Transmission From GPU Shortage to Energy Constraints

The AI infrastructure investment cycle is entering a new stage: bottlenecks are shifting from GPU supply to power and infrastructure. The U.S. Department of Energy expects that by 2028, data centers will account for 12% of U.S. electricity demand. This structural shift means that the upper limit of the AI infrastructure investment cycle is moving from chip production capacity to grid capacity and power supply stability.

There are clear cases of this transmission already in place. In June 2026, Fluence Energy announced a collaboration with Siemens and Nvidia to jointly develop power and electrical architecture solutions for AI data centers. Fluence will integrate its products to meet AI workloads’ requirements for grid stability, covering voltage and frequency fluctuation management, restart capabilities without external grid support, and AI load smoothing. After the news was released, Fluence’s stock price jumped nearly 44% that day.

Bloom Energy also holds a key position in AI power infrastructure. Oracle has signed contracts with Bloom to procure a total of 2.8 gigawatts of fuel cell capacity for multiple data center projects.

Hydropower infrastructure and power distribution equipment suppliers are becoming beneficiary areas in the AI investment cycle that cannot be ignored. However, it must be noted that the project execution cycles in these areas are far longer than chip manufacturing. Whether grid upgrades can keep pace with the schedule of data center construction is a key physical constraint facing AI infrastructure investment.

Humanoid Robots: Application-Layer Extension After AI Compute Power Scales Up

AI infrastructure investment is not limited to data centers themselves. Once compute power reaches a certain scale, the horizontal expansion of application scenarios accelerates—rising investment in humanoid robots is validating this logic.

In June 2026, German humanoid robot company Neura Robotics completed Series C financing, with the funding amount reaching up to $1.4 billion, with participation from multiple institutions including Nvidia, Amazon, Qualcomm, Tether, Bosch, Schaeffler, and the European Investment Bank. The company’s valuation is about $7 billion. This round of financing includes milestone-based payment terms, meaning all funds will be disbursed after the company achieves specific stage goals. This indicates that investors are participating in long-term planning in a more prudent manner.

A notable macro backdrop is that Dealroom data shows that total global funding in the robotics sector reached a record $55.8 billion in 2026, nearly double the full-year record for 2025. Neura Robotics’ funding amount may look astonishing, but it accounts for less than 3% of the overall global robotics investment market. This means that capital investment into humanoid robots has expanded rapidly in the early stage of industrialization, but the industry is still in an early phase with highly uncertain competitive dynamics.

These kinds of application-layer investments are both an extension of AI infrastructure development and a long-term window for realizing AI’s commercial value. However, humanoid robots have not yet formed a clear large-scale business model or a profitable path, and the payback cycle remains highly uncertain.

Validation Signals and Risk Assessment in the Investment Cycle

The core data supporting the current AI infrastructure investment cycle can be summarized into several key points. Oracle’s FY2026 Q4 RPO reached $638 billion, up 363%. More than one-third of the contracts are expected to be recognized as revenue within 13 to 36 months—this is the most direct forward-looking demand evidence in today’s market. CreditSights expects that the combined capital expenditures of the world’s five major hyperscale cloud providers in 2026 will be about $750 billion, up 67% year over year, covering GPU clusters, custom accelerators, data centers, and supporting power and cooling systems. Macquarie’s Viktor Shvets recently said that AI infrastructure investment has formed a global bubble, but it is not expected to burst in 2026 or 2027. At the same time, JPMorgan noted that as of the end of May 2026, AI-related debt accounted for about 15% of the entire corporate debt market.

Under the optimistic data, risk signals are also clear. Sequoia’s David Cahn calculated that there is an annualized gap of about $600 billion between hyperscalers’ AI capital expenditure and the actual AI ecosystem revenue, and that this gap is still widening in 2026. Allianz Research pointed out that the deviation between AI capital expenditure and revenue growth has reached about 46%, exceeding the 32% deviation seen during the telecom bubble cycle in 2001.

Another signal to pay attention to comes from CoreWeave. After its IPO, the AI data center operator’s stock price rose by more than 150%. But since the lock-up period ended, the company’s three co-founders have cumulatively cashed out about $2.3 billion, with Chief Strategy Officer Brian Venturo personally selling more than $1.1 billion. Major institutional investor Magnetar Financial has also sold about $5.5 billion worth of shares, cutting its holdings by half. Founder and early investor selling at high stock prices does not necessarily indicate deterioration of fundamentals—CoreWeave’s revenue still grew 111% year over year in the most recent quarter—but it is an early warning sign that market sentiment may shift at elevated valuations.

Conclusion

Overall, the current AI infrastructure investment cycle can be summarized into three relatively certain judgments.

First, the expansion of capital expenditure scale is still accelerating. The approximately $750 billion in capital expenditures by the five hyperscalers in 2026 is quantitative proof that the AI infrastructure investment cycle has not peaked. Oracle’s $638 billion RPO indicates that contract signing speed on the demand side has not slowed down. There is roughly a three-year conversion window from AI compute power moving from physical construction to contract revenue. This means that even in the most aggressive phase of supply expansion, revenue support is being built in parallel.

Second, the distribution of benefiting assets is shifting from concentration toward dispersion. From revenue growth among cloud service providers to capital inflows in chip design, advanced packaging, power supply, and even application-layer areas such as humanoid robots, the benefit chain of the AI infrastructure investment cycle has formed a multi-layered transmission. The chip industry chain remains the most certain beneficiary layer, but attention to capital flows into power equipment, infrastructure hardware, and application-layer companies is rising quickly.

Third, the risks stemming from supply-demand gaps cannot be ignored. Behind the current scale of investment, there is indeed a structural concern: capital expenditure growth is far outpacing revenue growth, and the share of AI-related corporate debt is rapidly climbing. Investors participating in this cycle need to closely track demand-side validation indicators such as RPO conversion rates and GPU utilization, rather than focusing only on the absolute value of capital expenditure.

For investors hoping to participate in this round of AI infrastructure investment cycle, Gate’s recently launched U.S. stock trading feature provides a new entry point. Through a strategic partnership with Alpaca, users can invest in more than 10,000 stocks and ETFs listed on the NYSE and Nasdaq using USDT directly on the Gate platform. From chip leaders like Nvidia and TSMC to power infrastructure companies like Fluence and Bloom Energy, to cloud service providers like Oracle, and even publicly listed humanoid robot companies that have not been fully commercialized yet, all can be configured on the same interface.

The underlying logic of the AI infrastructure investment cycle is based on the real progress of physical buildouts, not purely on emotional narratives. The answer to whether the validation cycle is real does not lie in the numbers in financial reports, but in every data center that is being connected to the grid.