725 billion USD AI Capex: Where is it going? An analysis of the three main beneficiary chains: chips, networks, and electricity

On June 18, 2026, the Federal Reserve maintained the federal funds rate for the fourth consecutive time, and the three major U.S. stock indices all closed lower. However, amidst macro disruptions, one underlying trend line has remained unshaken—capital expenditures (Capex) on AI infrastructure are expanding at an unprecedented scale in tech history. Nvidia (NVDA) closed at $204.65 that day, demonstrating relative resilience amid market pressure.

The root of this resilience lies in the fundamental demand support. According to Goldman Sachs’s updated forecast released in June 2026, the four largest hyperscale data center operators—Alphabet (Google), Amazon, Microsoft, and Meta—will have a total Capex of $725 billion in 2026, a 77% increase from $410 billion in 2025. S&P Global’s estimates also point to a figure above $700 billion. Including Oracle and others, the combined capital expenditure of the five major cloud service providers (CSPs) is expected to reach $760 billion, a year-over-year increase of 102.56%. Meanwhile, the Magnificent Seven’s combined spending on AI and data center infrastructure is projected at $527 billion.

When such a massive flow of capital converges in one direction, understanding its allocation becomes the prerequisite for comprehending the entire industry landscape.

First Beneficiary Layer: Chips—The Largest Absorber of Capex

In this flood of capital, chips are the most direct and concentrated beneficiaries. A significant portion of the $725 billion Capex from the four hyperscalers flows directly into GPU, ASIC, and related semiconductor procurement. Every dollar spent on AI infrastructure results in a substantial portion returning to chip suppliers.

The expenditure guidance from these giants confirms this trend. Amazon (AWS) projects about $200 billion in Capex for 2026; Microsoft around $190 billion; Meta between $115 billion and $145 billion; Alphabet raises its guidance to $180–$190 billion. The combined total exceeds $700 billion, with most directed toward AI data centers, in-house chip development, and computing infrastructure.

Notably, this spending race is already altering the financing structure of tech companies. Alphabet completed a $84.75 billion equity raise in early June 2026— the largest single equity issuance in history. Even though Alphabet’s operating cash flow over the past 12 months reached approximately $174 billion, facing an annual Capex of $180–$190 billion, internal cash flow alone is insufficient. Previously, Alphabet had raised over $85 billion through bonds in six currencies, even issuing rare 100-year GBP bonds. Meta is also considering raising several hundred billion dollars via stock issuance.

This shift from a “light asset, high margin” model to a “heavy asset, physical infrastructure” model extends the visibility of chip demand. Dell’Oro Group expects Capex growth to accelerate further in the second half of 2026, driven by NVIDIA’s Rubin system deployment and the update cycles of custom accelerators for hyperscale cloud providers. Even if short-term chip stocks face sell-offs, the demand fundamentals remain robust—this is the core safety margin provided by the Capex arms race for upstream suppliers.

Second Beneficiary Layer: Network Infrastructure—The “Vascular System” of Computing Power

Beyond chips, network infrastructure is the second major beneficiary. The expansion of AI clusters directly boosts demand for network bandwidth within and between data centers. Segments like optical modules, switches, and high-speed interconnect chips are experiencing exponential growth.

Research indicates that the global AI computing industry mainly presents investment opportunities in CPUs, optical interconnects, and AI PCs, involving TSMC, Nvidia, Intel, ASML, and Chinese industry chain companies like Zhongji Xuchuang. The rapid growth in token call volume drives persistent shortages of core computing resources in AI centers. The proliferation of inference workloads demands lower latency and higher bandwidth, pushing data center network architectures from 100G toward 400G, 800G, and even higher speeds.

Morgan Stanley estimates that a 350% surge in token demand is a key factor behind the upward revision of Capex forecasts for hyperscale cloud providers from $450 billion to over $800 billion. This data reveals the pull of inference-side demand on network infrastructure—each model inference requires data transmission across the network, and the growth in token volume directly translates into increased demand for network equipment.

Third Beneficiary Layer: Power and Cooling—The Invisible “Shovels”

Once computing power reaches a certain scale, power and cooling become “core bottlenecks” rather than just “support costs.” Gartner forecasts global data center electricity consumption will reach 565 TWh in 2026, up 26% from 447 TWh in 2025. Power demand will rise from 104 GW to 132 GW, a 27% increase. By 2030, this figure is expected to hit 290 GW. AI-optimized servers will account for 31% of data center power consumption in 2026, surpassing traditional servers by 2027.

Gartner’s director Linglan Wang notes: “The surge in compute-intensive AI workloads is driving unprecedented growth in data center power use. Today, AI computing capacity is constrained by power supply, making data center power assurance a new battleground in the global AI race for scale and profitability.”

On the cooling side, the high power density of AI servers is pushing thermal architectures from air cooling to liquid cooling. Liquid cooling has evolved from a niche technology to the mainstream infrastructure for high-power intelligent computing centers. JPMorgan predicts the global AI server liquid cooling market will jump from about $8.9 billion in 2025 to over $17 billion in 2026. The chilled water system market is expected to grow from $1.6 billion in 2026 to $12.7 billion by 2030. Companies like Delta Electronics, Chicony, Shuanghong, and Wistron are positioning themselves at this upgrade window.

Morgan Stanley estimates a 55 GW power shortfall in data centers. This supply-demand imbalance makes companies involved in power equipment, nuclear, and renewable energy projects beneficiaries of Capex spillovers. The logic of benefits here differs from chips—it’s not direct Capex absorption but “rigid derivative demand” driven by the expansion scale, closely tied to data center construction progress.

Global Data Center Capex: A Trillion-Dollar Macro Narrative

Summarizing these segments, the overall picture of global data center Capex becomes clearer. Dell’Oro Group has raised its 2026 global data center Capex outlook to over $1 trillion. The top four U.S. cloud providers have increased their data center Capex by 78%.

Looking further ahead, Goldman Sachs forecasts that by the end of 2030, the Capex of just these four hyperscalers could reach $53 trillion. Nvidia CEO Jensen Huang predicted in May that global AI data center investments will surge to $3–$4 trillion by 2030. While these projections are subject to uncertainty, the direction is clear: AI infrastructure investment shows no signs of slowing.

ROI Challenges: When Will Profits Materialize?

However, the flip side of massive Capex is the uncertainty of ROI (Return on Investment). This remains the core market debate.

The Capex-to-Revenue ratio has hit historic highs. Meta’s 2026 Capex is expected to be about 54% of sales; Microsoft around 47%; Alphabet approximately 46%. Goldman Sachs analysts note that consensus estimates project hyperscalers’ Capex at $770 billion in 2026, roughly 100% of their operating cash flow.

Such capital intensity means that relying solely on cloud revenue growth to quickly amortize Capex, depreciation, and financial pressures is no longer feasible. The four hyperscalers are shifting from a traditional “self-sustaining, cash flow-driven expansion” model to a leveraged “heavy asset, physical infrastructure” approach relying on capital markets’ debt/equity tools.

CITIC Securities still identifies three key mechanisms expected to improve cloud ROI: accelerated YoY growth of cloud revenue, increased AI adoption, and sustained demand for cloud services. But the timing of ROI improvement remains uncertain. Gartner data shows global AI spending will reach $17.6 trillion in 2025, rising to $26 trillion in 2026. Infrastructure investments account for the largest share (55%), but the speed of ROI realization depends heavily on the commercialization pace of inference-side applications.

Conclusion

The 2026 AI Capex arms race is no longer just a “money-burning contest,” but a comprehensive industry chain reconstruction—from chips to power, from computing to energy. Annual expenditures of $725–$750 billion constitute an undeniable economic fact.

The transmission logic of this benefitting chain is clear: chip manufacturers (especially Nvidia) as the primary direct absorbers benefit first; network infrastructure providers gain incremental orders as clusters expand; power and cooling suppliers, as “rigid derivative demand,” benefit later during construction cycles. The timing and elasticity of these three layers differ, but together they form a complete industry narrative.

For markets, the real test may not be whether Capex continues to grow—current guidance suggests that’s quite certain—but when and how efficiently these investments translate into sustainable profits. Validating ROI will determine whether this arms race ushers in a new technological cycle or repeats some chapters of the internet bubble. Until then, every link in this benefitting chain will continue to absorb the largest tech capital influx in history.