Where is the $650 billion flowing? A comprehensive analysis of AI capital expenditure landscape and key beneficiary stocks by 2026

The first quarter of 2026 earnings season has just wrapped up, and a set of numbers has sparked ongoing discussion in the investment community: Amazon locked in its full-year capital expenditure at around $200 billion, Alphabet raised it to $180–$190 billion, Microsoft maintained its expectation of $190 billion, and Meta adjusted again to $125–$145 billion. The combined capital expenditure of the four mega cloud providers in 2026 has surpassed the $650 billion mark, a growth of over 60% compared to approximately $410 billion in 2025. If we include Nvidia, Apple, Tesla, and others into the “Magnificent Seven” category, this figure approaches nearly $750 billion.

Meanwhile, signals such as squeezed free cash flow, declining gross margins, and stock price divergence after earnings reports are also indicating to the market: the capital expenditure cycle for AI infrastructure is shifting from “cost-no-object land grab” to a new phase of “return-on-investment calculation.”

The Four Major Supercomputing Centers’ 2026 AI Capex Panorama

In 2026, investment in AI infrastructure is shifting from “experimental deployment” to “large-scale implementation.” The annual capital expenditure of the four major CSPs is expected to be between $650 billion and $700 billion, accounting for roughly 40% of the total capital expenditure of the Russell 1000 index, doubling the level of 2024.

Amazon: $200 billion investment “shakes the scene.” Amazon has locked in its 2026 capital expenditure at about $200 billion, a nearly 60% increase from the estimated $125 billion in 2025, far exceeding analyst expectations of $144.7 billion. The funds are mainly allocated to AI data center construction, in-house Trainium/Graviton chip R&D, and the “Kuiper” low-earth orbit satellite internet infrastructure. The core driver comes from the continuous expansion of AWS—Q1 AWS revenue hit $37.6 billion, up 28% YoY, the fastest growth in nearly four years, but free cash flow sharply shrank from $25.9 billion to just $1.2 billion.

Alphabet: The most aggressive infrastructure attacker. Google raised its 2026 capital expenditure guidance to $175–$185 billion, nearly double the actual $91.4 billion spent in 2025. CFO Anat Ashkenazi revealed during the earnings call that about 60–65% of this is allocated to short-cycle assets like servers, with the rest going to data centers, energy infrastructure, and related facilities. Google Cloud’s revenue in Q1 reached $20 billion, up 63% YoY, with backlog orders approaching $462 billion, over half of which will be recognized as revenue within the next 24 months. This data has boosted market confidence in Google’s “invest while earning big profits” capability.

Microsoft: Demand continues to outpace supply. Microsoft’s full-year 2026 capital expenditure is expected to be around $190 billion, up 61% YoY. In Q1, quarterly capex was $31.9 billion, with about two-thirds invested in GPUs, CPUs, and other compute assets, roughly $25 billion driven by component price increases. AI business annualized revenue has surpassed $37 billion, up 123% YoY, with Azure growth maintaining at 40%. The core bottleneck is not demand but power supply and chip delivery times—Microsoft expects supply constraints to persist throughout 2026.

Meta: Advertising cash cow fuels AI. Meta has further raised its 2026 capital expenditure guidance from $115–$135 billion to $125–$145 billion, an increase of 73–100% over 2025’s $72.2 billion. Spending covers large-scale GPU procurement for Meta Superintelligence Labs, data center expansion, and deployment of self-developed chips exceeding 1 GW. Q1 advertising revenue continues to grow, with global daily active users reaching 3.56 billion, providing stable cash flow support for AI investments.

While the spending pace of these four giants is aligned, their strategic logic differs significantly: Amazon is in the “left-side heavy position, betting on supply”—using cash flow to secure future compute share; Microsoft is in the “post-expansion supply catch-up”—facing the challenge that capacity rollout cannot keep pace with customer consumption; Google is pursuing a “platform path of infrastructure + ecosystem reinforcement,” self-developing TPU systems to reduce reliance on single-GPU solutions; Meta, as a cash cow from advertising, is fully reinvesting into AI infrastructure, with the unique aspect that it is not a traditional cloud service provider—its expenditure and returns are more closely tied to advertising effectiveness rather than direct cloud revenue monetization.

AI Capital Expenditure Flows: From GPU to HBM to Optical Networks

To better illustrate the allocation logic of funds, the diagram below dissects approximately 75–80% of the $650 billion into hardware and infrastructure components, following a roughly “infrastructure → compute chips → storage → networking → power” hierarchy.

$200 billion (AWS data centers and AI chip procurement) + $185 billion (Google server and data center construction) + $190 billion (Microsoft GPU/CPU short-cycle asset procurement) + $135 billion (Meta compute clusters and self-developed chip deployment) = approximately $710 billion total capital expenditure of the four giants.

Layer 1: Infrastructure

• Servers, data centers, self-built power plants, land, supply chain: about 65–70%

Layer 2: Core compute chips

• Nvidia GPUs (B200/GB200/H200 series, NVLink interconnect architecture, Spectrum-X Ethernet): about 25–30%
• Self-developed AI chips (Amazon Trainium/Graviton, Google TPU, Meta self-developed + AMD procurement): about 10–15%

Layer 3: Storage and Memory

• HBM high-bandwidth memory (led by Micron, SK Hynix), data center DRAM: about 5–8%

Layer 4: Networking and Interconnects

• Data center optical networking equipment (Ciena, etc.), routers, switches, InfiniBand: about 3–5%

Layer 5: Power and Cooling

• High-voltage distribution (800V architecture), BBU backup batteries, liquid cooling: about 2–3%

Layer 1: Core compute chips—Nvidia’s absolute dominance

In FY2026, Nvidia achieved revenue of $180B, up 65% YoY, with data center revenue at $125B, accounting for 89.6% of total revenue. In Q4, data center revenue hit $62.3 billion, up 75% YoY, representing over 91% of total revenue. The main driver is the dominance of the Blackwell computing platform (B200/GB200 systems) in generative AI, large model training, and inference scenarios. Network business performed even better—Q4 revenue nearly $11 billion, up 263% YoY, indicating cloud giants’ procurement has expanded from “buy GPUs” to “buy complete system-level solutions including NVLink interconnect architecture and Spectrum-X Ethernet.”

It’s noteworthy that self-developed chips are diverting some expenditure. Amazon’s Trainium 2 chips have already generated over $10 billion in annual revenue, aiming for 30% of AI compute tasks to be handled by self-developed chips by the end of 2026. Google continues to push TPU systems, forming a parallel procurement structure with Nvidia GPUs. Meta, while deploying large-scale self-developed chips, also procures a significant number of AMD chips to mitigate supplier risk.

Layer 2: Storage and Memory—HBM capacity bottleneck

AI training clusters require loading massive parameters into memory in real-time, making HBM (high-bandwidth memory) an indispensable core component. Micron disclosed that its HBM capacity in 2026 is fully sold out, with the market expected to grow from $35 billion in 2025 to $100 billion in 2028, a CAGR of 40%. In Q2 FY2026 (ending February), Micron’s revenue was $23.86 billion, up 196%, with a gross margin of 75%. Data center-related revenue rose to $5.78 billion, up 57%. The demand signals are clear: as cloud providers pour hundreds of billions into data centers, expanding storage capacity and bandwidth is a rigid expense.

Layer 3: Networking and Interconnects—Optical networks upgrade from “supporting role” to “necessity”

As individual AI clusters expand from thousands to tens or hundreds of thousands of GPUs, bottlenecks in intra-rack and inter-data center communication become apparent. Ciena’s revenue in Q2 FY2026 was $1.57 billion, up 40%, with adjusted EPS of $1.64, nearly tripling YoY. The core driver is the shift in cloud service providers’ AI investments from purely “compute” to “network infrastructure”—the demand for data center interconnects (DCI) and intra-cluster optical switching is exploding. Ciena has raised its full-year revenue guidance to $6.3 billion, up about 32%, and its CEO stated that the optical network market could double to $50 billion by 2029.

Layer 4: Power and Cooling—Invisible but essential

With single rack power consumption surpassing 1 MW, traditional power distribution systems are insufficient. Amazon and Google have explicitly required the next-generation data centers to adopt 800V architectures. Meanwhile, BBU backup systems have shifted from optional to mandatory. Power infrastructure investments account for about 2–3% of hardware spending, but if power shortages or distribution bottlenecks occur, trillions of dollars in AI infrastructure investments could fail to deliver expected outputs.

Wall Street Divergence: AI ROI Debate

On the other side of capital expenditure, the debate over “when will these investments generate returns” is heating up on Wall Street.

Bullish logic: demand is translating into revenue, monetization is imminent

Google Cloud’s Q1 2026 revenue hit $20 billion, up 63% YoY, with backlog orders nearly doubling to over $460 billion—clear evidence of “spending → demand → monetization.” Microsoft’s AI business annualized revenue exceeds $37 billion, up 123%, with Azure’s remaining performance obligations (RPO) at $627 billion. Nvidia CEO Jensen Huang stated during the earnings call that intelligent agents are rapidly being adopted by enterprises worldwide—“compute equals revenue”—without compute, tokens cannot be generated, and without tokens, no revenue.

From a macro perspective, Charles Schwab’s mid-term outlook indicates the S&P 500’s earnings could grow about 25% this year, but this growth is highly concentrated among a few AI supply chain companies like Alphabet, Micron, Intel, and Broadcom. This suggests that while positive spillovers from AI are not yet fully widespread, they are strongly supporting index-level earnings growth.

Bearish logic: costs are front-loaded, returns are delayed, cash flows are heavily consumed

The core argument among skeptics centers on the sharp contraction of free cash flow. Morgan Stanley predicts Amazon’s free cash flow will be negative $17 billion in 2026; Pivotal Research expects Alphabet’s free cash flow to plummet from $73.3 billion in 2025 to just $8.2 billion. Microsoft’s gross margin has fallen to 67.6%, the lowest since 2022, mainly due to accelerated depreciation from AI infrastructure investments.

Goldman Sachs research chief Covello’s cautious stance reflects a view that about 95% of companies’ AI-related returns are near zero, and the high profit concentration in semiconductors is unsustainable. Another industry report estimates that AI infrastructure capex by major US tech giants from 2025 to 2027 could reach $1.4 trillion, but with returns far below market expectations, and significant risks of technological obsolescence leading to sunk costs.

Balanced view: The core of the ROI debate is a mismatch in timing

From an industry logic perspective, the core of this debate isn’t whether AI investments are effective but rather the timing of return realization. Early-stage investments in GPUs, data centers, and power incur depreciation costs, while AI revenue streams—incremental SaaS subscriptions, advertising efficiency gains, cloud consumption—materialize gradually with a typical lag of 6–12 months. The market’s current pricing reflects a transitional state: cloud providers need to demonstrate that the marginal ROI of capital expenditure can turn positive before 2027.

Gate’s Real Stock Trading: A New Bridge Connecting Crypto Assets and Traditional Markets

While traditional financial markets continue to witness the evolution of AI investment themes, Gate officially launched its real US stock trading service on June 1, 2026, creating a compliant channel for crypto users to directly participate in the US stock market.

Core advantage: direct USDT purchase of US stocks. Unlike common tokenized stocks or RWA-mapped products, Gate’s service connects with a licensed US broker-dealer and clearing firm Alpaca, enabling users to buy actual US stock shares (via a non-depository brokerage account structure) within the Gate platform, not just on-chain derivatives. This means users can flow liquidity directly from crypto assets into Nasdaq and NYSE stocks, covering over 10,000 US stocks and ETFs. On June 5, 2026, Gate further added pre-market and after-hours trading, extending trading hours from the original 6.5 hours to 16 hours, covering more market volatility windows.

Crypto users’ adaptation: fragmentation and zero implicit costs. Gate’s real stock trading supports fractional trading as low as 0.01 shares, allowing participation in top US stocks like Apple, Nvidia, Tesla with as little as $1. The platform imposes zero holding costs—no swap fees, no overnight fees, dividends and dividends paid in USDT automatically credited.

Direct mapping to AI themes. For readers interested in AI capital expenditure and supply chain structure, Gate’s US stock trading channel offers a direct participation gateway—whether it’s Amazon, Microsoft, Google, Meta in their capital expansion phase, or Nvidia, Micron, Ciena as core supply chain players—all can be traded in one platform using USDT. The disconnect between crypto assets and traditional securities is thus eliminated.

Conclusion

2026 marks a critical turning point where AI infrastructure shifts from “arms race” to “business validation.” The hundreds of billions of dollars in capital expenditure by the four cloud giants are no longer about “whether to invest,” but about “how to invest more effectively and realize faster returns.” Nvidia’s dominant position at the compute layer is unlikely to be challenged in the short term, but self-developed ASICs, HBM storage, optical networking, and power infrastructure are forming new growth poles, with supply chain profits spreading from GPU alone to upstream and downstream networks.

For investors, the key window of observation will open in the second half of 2026 through 2027—if AI revenue growth continues to outpace depreciation costs, the ROI inflection point of capital expenditure will become clearer, and the market’s current “burn money” skepticism could be reassessed. The launch of Gate’s real US stock trading provides a bridge for crypto users to operate between traditional and digital assets.