Trillion-dollar AI infrastructure race: After Oracle's surge in capital expenditure, who has a higher ROI—Microsoft, Amazon, or Google?

2026 is becoming the most aggressive year in the history of AI data center capital expenditures. According to the latest forecast from Dell'Oro Group, global data center CapEx will surpass $1 trillion in 2026, with the data centers of the four major US cloud providers having already increased their CapEx by 76% in 2025. Based on estimates from Morgan Stanley, just four companies—Amazon, Microsoft, Alphabet, and Meta—are expected to spend about $630 billion on data centers and AI chips in 2026, more than quadrupling the 2023 expenditure level, accounting for approximately 2.2% of US GDP.

In such a massive capital race, how can we evaluate the actual efficiency of these investments? The anxiety this question triggers among investors is becoming a more urgent real-world concern than whether AI demand has peaked.

On June 10, 2026, Oracle delivered an almost impeccable-looking financial report—yet its stock price plummeted over 13% the next day, with a market value evaporating more than $70 billion. This stark contrast provides an excellent entry point for discussions on the return on investment in AI infrastructure.

Oracle: Exceptionally Beautiful Earnings, Exceptionally Harsh Market Reaction

From any traditional financial metric, Oracle’s FY2026 Q4 report (ending May 31, 2026) reads like a “battle record.” Quarterly total revenue reached $19.2 billion, up 21% year-over-year, beating the market expectation of $19.08 billion. Non-GAAP EPS was $2.11, about 7.7% higher than the consensus estimate of $1.96.

But what truly caught the market’s attention were two data points.

First, cloud infrastructure (OCI) revenue grew 93% YoY to $5.8 billion, with total cloud revenue for the year reaching $9.9 billion, up 47%. This is Oracle’s fastest-growing business segment ever, indicating its rapidly expanding share in the AI compute leasing market.

Second, remaining performance obligations (RPO) surged to a record $638 billion, a 363% YoY increase. Of this, $12 billion is expected to be recognized as revenue within the next 12 months, with about $34 billion gradually converting into actual revenue over the following two years. More critically, in just FY2026 Q4 alone, Oracle signed AI infrastructure contracts worth $67 billion. In comparison, Oracle’s total revenue for FY2026 is only about $67 billion—meaning that a single quarter’s new orders are nearly equal to the entire year’s revenue.

These figures point to a clear conclusion: Oracle’s demand for AI compute power is not “just talk”—it has already been locked in through prepaid contracts and long-term agreements.

However, the market did not buy this. After the earnings release, Oracle’s stock fell about 7% in after-hours trading, and during the next day’s trading, it dropped over 11%. As of June 12, 2026, the stock price had fallen from about $201 before the earnings to around $183.

The issue lies in capital expenditure. Oracle’s full-year CapEx for FY2026 reached $55.7 billion, far exceeding the management’s previous forecast of $50 billion. More critically, free cash flow turned negative—operating cash flow for the year was $32 billion, while net CapEx outflows reached $48 billion, creating a roughly $16 billion gap. This means Oracle is not only investing all its operating cash flow into infrastructure but also relying on external financing to fill the gap. Data confirms this: in FY2026, Oracle raised about $48 billion through debt and equity financing, and plans to refinance approximately $40 billion in FY2027 to continue funding data center construction.

From a financial structure perspective, Oracle faces a classic “growth versus profit” dilemma. Gross margin has declined due to accelerated data center investments, and management expects FY2027 gross margin to come under further pressure as the construction pace consumes substantial upfront capital.

The core issue is not whether Oracle can raise funds—backed by a backlog of $638 billion in orders, financing channels are not the obstacle. What truly alarms the market is: a company with annual revenue of about $67 billion needs to invest over $50 billion each year—what does this imply? If AI infrastructure spending remains at this high level long-term, when will gross margins recover? Are marginal returns on unit capital diminishing? These questions have no definitive answers, but the market has already made a preliminary judgment through the stock price—a 13% decline is a vote of distrust in capital efficiency.

Microsoft: Largest Scale, But What About Return on Investment?

In contrast to Oracle’s “aggressive expansion,” Microsoft exemplifies a “scaled-up advancement” approach.

Microsoft’s FY2026 Q3 (ending March 31, 2026) earnings show quarterly revenue of $82.9 billion, up 18% YoY, beating the consensus of $81.46 billion. The Intelligent Cloud segment revenue was $34.7 billion, up 30%, with Azure and other cloud services growing 40%, surpassing the management guidance of 37-38%, ending several quarters of slowing growth. The acceleration in Azure’s growth signals that the previous slowdown was mainly supply-side, not demand-side—new GPU capacity coming online suggests the growth ceiling is not yet reached.

Microsoft’s annualized AI revenue run rate has exceeded $37 billion, up 123% YoY. Microsoft 365 Copilot paid seats surpassed 20 million, a 250% annual increase, with Accenture alone purchasing over 740k seats. Commercial RPO reached $627 billion, up 99%.

From a revenue perspective, Microsoft’s progress in AI commercialization is the most solid among the four giants.

But CapEx is also high. FY2026 Q3 CapEx was $31.9 billion, up 49%, below the market expectation of $35.3 billion. However, Microsoft has raised its FY2026 CapEx guidance to about $190 billion, with roughly $25 billion attributed to component price increases rather than actual capacity expansion. CFO Amy Hood stated during the earnings call that Q4 CapEx is expected to exceed $40 billion.

A noteworthy detail: about two-thirds of Microsoft’s short-term CapEx is concentrated in “short-term assets”—mainly GPUs and CPUs. This indicates a relatively flexible CapEx structure, with shorter depreciation cycles, allowing timely adjustments based on demand. In contrast, most of Oracle’s infrastructure investments are in long-term fixed assets, with less flexibility for adjustment.

Microsoft’s gross margin has declined from about 70% last year to 68%, mainly due to ongoing investments in AI infrastructure and increased AI product usage.

Unlike Oracle, Microsoft’s AI spending generates three cash flow streams: Azure compute leasing revenue, SaaS subscription revenue from products like Copilot, and model training and inference revenue through its partnership with OpenAI. This diversified revenue structure provides more avenues for unit capital return, partially hedging the risk of uncertain returns from any single business line.

After the earnings release, Microsoft’s stock dipped about 3.5% in after-hours trading but stabilized, reflecting a divided market tolerance for high CapEx—while Azure’s growth rebound is bullish, the $190 billion annual CapEx guidance still keeps some investors cautious.

Amazon: Fastening the AWS Pace, But CapEx Already Tops Industry

Among the four giants, Amazon’s CapEx scale is the largest.

Amazon’s Q1 FY2026 report shows quarterly net sales of $181.5 billion, up 17%, significantly exceeding the consensus of $177.2 billion. AWS revenue reached $37.6 billion, up 28%, the fastest growth in 15 quarters. AWS operating profit was $14.2 billion, also surpassing expectations. However, this quarter’s net profit includes about $16.8 billion of non-operating gains from valuation of investments in Anthropic; excluding this, operating profit is about $23.9 billion. AWS’s customer backlog continues to grow, with enterprise clients accelerating multi-year cloud and AI contracts.

In CapEx, Amazon’s cash CapEx for Q1 FY2026 hit $43.2 billion, and including financing leases, $44.2 billion—mainly for data centers, network equipment, custom chips, and AI infrastructure. Amazon has committed to about $200 billion in CapEx for 2026, making it the industry’s largest among hyperscale cloud providers.

But such massive CapEx exerts clear pressure on cash flow. Over the past 12 months, Amazon’s free cash flow plummeted from about $25 billion to roughly $1.2 billion—a 95% decline.

From a capital efficiency perspective, Amazon’s strategy differs structurally from the others. It invests heavily in self-developed chips like Trainium2, Trainium3, and Graviton5, which serve both external AWS customers and large tech firms like Meta. Amazon’s CEO Andy Jassy has said that if its chip business is sold externally, it could grow into a $50 billion annual revenue line. Additionally, Amazon’s $25 billion investment in Anthropic has already appreciated significantly, providing extra capital returns.

Amazon’s AI CapEx model can be summarized as “vertical integration + strategic investment.” Self-developed chips reduce reliance on external suppliers and long-term procurement costs, while investments in Anthropic give priority access to advanced AI models. This approach results in the highest capital density among the four, with potential for deeper long-term moat.

Post-earnings, Amazon’s stock rose about 2.74% to $270.25 in after-hours trading, reflecting market optimism about AWS’s return to 28% growth, though high CapEx remains a key investor concern.

Google Cloud: Leading Growth, Still in the Catch-up Phase

If growth rate is the metric, Google Cloud is the most dynamic among the four.

Alphabet’s FY2026 Q1 report shows quarterly revenue of $109.9 billion, up 22% YoY—the fastest in recent two years—and exceeding the consensus of about $107 billion. Google Cloud revenue was $20 billion, up 63%, further accelerating from about 48% in the previous quarter. Cloud operating profit tripled to about $6.6 billion, with an operating margin approaching 33%.

The most notable data is Google Cloud’s backlog (RPO)—nearly doubling to about $462 billion. While still below Microsoft’s $627 billion, considering Google Cloud’s smaller absolute revenue (~$20 billion vs. Azure’s ~$34.7 billion), the RPO-to-revenue multiple is higher, indicating stronger future revenue visibility among the four.

In CapEx, Alphabet raised its 2026 CapEx guidance to the $180–$190 billion range, higher than the previous $175–$185 billion. Q1 CapEx was $35.7 billion, slightly below the expected $36.4 billion. Management indicated that 2027 CapEx will continue to grow significantly.

Google Cloud’s approach involves TPU chips developed in-house alongside NVIDIA GPUs. TPUs are already widely deployed internally and among some enterprise clients. Starting in 2027, Google plans to supply TPUs directly to some customers for use in its own data centers, opening new commercial avenues for AI infrastructure.

From a capital efficiency perspective, Google Cloud is at a “transition from economies of scale to profit realization” stage. Cloud is shifting from a “strategic investment” to a “profit center,” with operating margins near 33%, one of the highest among the four. However, cloud accounts for only about 18% of Alphabet’s total revenue, so while rapid growth boosts overall valuation, the reshaping of the group’s profitability structure still takes time.

After the earnings, Alphabet’s stock dipped slightly by 0.6% to about $345, reflecting investor weighing of high CapEx expectations against strong performance.

Who Spends the Most Efficiently: Comparing Four Models

Each of the four companies has a distinct path in AI infrastructure spending. To answer “who spends most efficiently,” a unified efficiency evaluation framework is needed.

From the output side, capital returns can be broken down into three dimensions: current revenue capital return rate (CapEx / cloud revenue growth), future revenue visibility (RPO / revenue), and gross margin pressure transmission (margin change magnitude).

Oracle’s cloud revenue growth is the fastest (OCI +93%), but FY2026’s $55.7 billion CapEx corresponds to about $25 billion in annualized cloud revenue (Q4 cloud revenue of $9.9 billion annualized to ~$39.6 billion; Q4 OCI of $5.8 billion annualized to ~$23.2 billion). CapEx is roughly 1.4 times cloud revenue. Considering OCI’s growth rate above 90% and the huge backlog (RPO/revenue ratio of about 9.5), it has significant long-term growth potential. But the current problem is: free cash flow has turned negative, and infrastructure investments are consuming all operating cash flow, requiring ongoing external financing. This means Oracle must sustain financial pressure until the infrastructure’s return on scale is realized.

Microsoft’s capital efficiency is the most balanced among the four. Its annualized CapEx of about $190 billion corresponds to roughly $34.7 billion in quarterly cloud revenue, or nearly $140 billion annualized. RPO of $627 billion is about 2.9 times the annualized cloud revenue (~$218 billion). Gross margin has only slightly declined from about 70% to 68%, indicating that AI spending has not yet caused systemic profitability issues. Microsoft’s advantage lies in its diversified AI revenue streams—from Azure compute leasing, SaaS subscriptions, to model services—dispersing capital return volatility and providing a longer buffer between investment and revenue realization.

Amazon’s strategy is the most “long-term.” Its $200 billion annual CapEx is the industry’s largest, but a significant portion is invested in self-developed chips and strategic equity investments. Chips like Trainium and Graviton have formed over $20 billion in annualized revenue and are growing at triple-digit rates. If these vertical integration paths succeed, they could significantly reduce reliance on external suppliers and improve margins long-term. The short-term cost is evident: free cash flow plunged from about $25 billion to roughly $1.2 billion.

Google Cloud leads in growth, with an operating margin (~33%) already showing scale profitability. Its CapEx of about $180–$190 billion annually corresponds to roughly $20 billion quarterly cloud revenue, or about $80 billion annualized. RPO of about $462 billion is roughly 5.8 times the annualized cloud revenue, the largest among the four in relative scale. Google’s self-developed TPU chips are already operational (Gemini API calls over 10k tokens/min), but the fact that cloud accounts for only about 18% of Alphabet’s total revenue means the “group-level” return on AI infrastructure investments remains to be seen. Google also needs to watch for potential cannibalization of search ad revenue by AI search, though current AI search ad growth remains strong.

The Trillion-Dollar AI Infrastructure Race: Supply-Demand Mismatch Becomes a Bottleneck

It’s important to note that whether these four companies’ CapEx plans can be implemented as scheduled is itself a variable worth monitoring. The expansion of AI data centers is encountering physical world constraints.

According to S&P Global Energy Horizons data, these four tech giants currently operate about 600 data center facilities, with another 544 in planning or under construction. A modern AI data center costing over $4 billion, with 100 MW capacity, typically spends about 70% of that on servers and GPUs. But the bottleneck isn’t funding—it’s power supply, transformers, and construction permits. Delivery times for transformers in Europe have extended to 100 weeks, and about one-third of US data centers under construction rely on on-site gas turbines, which are nearly sold out until 2029.

This means that even if the four companies are willing to invest hundreds of billions, the actual pace of capacity addition may lag significantly behind the release of capital. This may be the deeper logic behind the market’s reaction after the Oracle, Microsoft, Amazon, and Google earnings—investors are not worried about “how much they spend,” but whether “these funds can be used efficiently.”

Conclusion

The competition among these four tech giants in AI data center CapEx is fundamentally a “time window” game. AI compute power is evolving from a differentiator to an infrastructure-level “threshold condition”—the companies that can build a global AI compute network first will establish structural advantages in model training, inference costs, and customer lock-in. But if supply growth outpaces demand growth systematically, excessive early investments could drag down shareholder returns.

From current efficiency assessments, Microsoft leads in risk diversification and short-term visibility; Google excels in cloud growth and backlog scaling, with cloud already beginning to scale profitably; Amazon is building a long-term moat through self-developed chips and strategic investments, but faces the most immediate cash flow pressure; Oracle is betting heavily on growth prospects with the largest leverage—its backlog and cloud growth are very strong, but its financial fragility is also the most apparent—using the highest capital consumption rate to chase the fastest market share expansion.

For investors, these paths are not simply “which is better,” but strategic choices involving “different risks at different time scales.” Who spends most efficiently may only be truly revealed around 2028 or even 2030, when these hundreds-of-billion-dollar capital expenditures start to flow back in large-scale cash flows.