AI Infrastructure Enters the Second Phase: Why Is the Capital Market Starting to Focus Again on Profitability?

June 30, 2026 (Beijing Time), the Nasdaq Composite Index surged by 522.52 points, or 2.07%, to close at 25,820.14. NVIDIA rose 1.27% on the day, closing at $194.97, with a market cap of approximately $4.72 trillion. But just five trading days prior, the world’s most valuable semiconductor company had only just gone through a round of consecutive declines.

Behind the short-term stock-price volatility is a deeper structural shift: AI investment is moving from the “storytelling” phase into the “accounting” phase. The market is no longer only asking “who is participating in AI,” but is starting to ask “who can truly make money from AI.”

The $725 Billion “Bill”: Why Capital Expenditure Is Still Rushing Ahead

To understand this shift, you first need to see where the money is flowing.

Goldman Sachs’ updated forecasts released in June 2026 show that the four major hyperscale data center operators—Alphabet (Google), Amazon, Microsoft, and Meta—will reach total capital expenditure of $725 billion in 2026. Specifically by company: Amazon at about $200 billion, Microsoft at about $190 billion, Google at about $175 billion to $185 billion, and Meta at about $115 billion to $135 billion. Compared with 2025’s $410 billion, this figure is up 77% year over year.

More noteworthy is the pace of the revisions. In just the past six months, market expectations for 2026 cloud vendors’ capital expenditure have been raised by nearly 80%. Barclays expects major cloud vendors’ capital expenditure to reach $919 billion in 2027 and further rise to about $1.16 trillion in 2028.

What does $725 billion amount to? The number exceeds the total size of the global semiconductor market in 2025: the World Semiconductor Trade Statistics (WSTS) forecasts the global semiconductor market size in 2026 to be $1.5112 trillion—meaning the four companies’ AI capital expenditure is already approaching nearly half of the global semiconductor market total.

The flow of this funding can roughly be divided into three tiers: the top tier is chip procurement (NVIDIA GPUs, AMD accelerators, custom ASICs, etc.); the middle tier is data center infrastructure (land, buildings, power, cooling systems); and the bottom tier is network equipment and the software ecosystem (InfiniBand, Ethernet, the CUDA ecosystem, etc.). Bernstein Research notes that even just increases in HBM (High Bandwidth Memory) prices could raise hyperscale cloud vendors’ overall AI capital expenditure by about 30%.

At the same time, total cumulative global data center investment is expected to reach around $1.6 trillion by 2030. In the U.S., annualized data center construction spending reached a rate of $50.7 billion in April 2026. Samsung announced total investment of 2,655 trillion won (about 11.68 trillion yuan), while SK Group plans to invest more than 100 trillion won per year in South Korea for the next 10 years. Blackstone plans to invest $30 billion to build AI data centers in Japan.

But the scale of spending itself is no longer the only focus for the market.

From “Compute Scarcity” to “Return Verification”: A Deep Switch in Investment Logic

Over the past three years, the AI industry has followed a clear and powerful logic chain: the scarcer the compute, the more justified the capital expenditure; the larger the capital expenditure, the higher the valuation. This self-reinforcing cycle was hardly questioned. However, entering 2026, every link in this logic chain is undergoing stress testing.

In a June research report, Goldman Sachs clearly pointed out that the core contradiction in the AI market is intensifying: fundamentals remain strong, but the market has already priced in too much of the future upside. The share of GDP accounted for by U.S. tech investment has risen to about 4.9%, exceeding the peak around the dot-com bubble era of 2000. The market’s pace of pricing in future AI gains is plainly faster than the pace at which productivity dividends actually materialize.

Goldman Sachs further noted that since November 2022, the market value of AI-related companies has surged by $27 trillion, far exceeding the $9 trillion estimated by macro benchmark calculations. Upgrades to earnings have temporarily kept valuation concerns in check, but stock volatility could rise further.

It is in this context that the AI industry is reaching a key threshold. According to an Exponential View report, as of the first quarter of 2026, global generative AI industry revenue (excluding China) exceeded the quarterly depreciation expenses of AI infrastructure for the first time. In 2026, annual depreciation expenses for AI infrastructure are expected to approach $111 billion. In other words, the cash flow generated by AI businesses is already able to cover the accounting depreciation costs formed by servers, GPUs, and data centers—meaning the industry has crossed the first “threshold of being able to support itself.”

However, there is still a considerable gap before proving that the entire capital cycle can deliver reasonable returns. The report estimates that by the end of 2026, cumulative AI-related capital expenditure by global hyperscale cloud vendors and emerging AI cloud platforms will reach approximately $2 trillion. The market is shifting from a “compute scarcity belief” to a systematic review of return on investment.

Jim Chanos, the legendary Wall Street short-seller, gave a specific number at a seminar in June 2026: the expected pre-tax return on invested capital (ROIC) for compute infrastructure is currently only 5% to 8%. Chanos’ logic is not complicated. He points out that the current AI industry chain has a huge “financial mismatch”: companies selling chips and data center equipment recognize revenue and profit immediately, while the cloud vendors spending money to purchase these assets are capitalizing the costs. Once these assets go online and begin depreciating, the impact on profits will be enormous.

He drew an analogy between current AI infrastructure investment and the dot-com bubble period from 1998 to 2000. At that time, S&P 500 operating profit grew by 30% in two years, but when the order book collapsed in 2001 and depreciation costs continued to show up, S&P 500 profits plunged 40%.

Real Signals in the ROI Debate: Compute Rental Prices Are Coming Down

Market doubts are not without basis. The hourly rental price of the NVIDIA B200 chip fell from $6.11 on May 30 to $4.22 on June 22, a drop of 31% in less than a month. AI server rental costs overall are showing a sustained downward trend.

If scarcity disappears, the logical foundation for sustaining capital spending will be weakened. In its latest report, JPMorgan summarized this shift as “from faith to cost.” This assessment is key to understanding the current cycle: the same fact supports the certainty of capital expenditure while also planting fragility on the demand side.

But the short side’s logic also has a counterpoint. In response to market skepticism, NVIDIA founder and CEO Jensen Huang gave a positive reply at the annual shareholder meeting on June 24. He backed it with financial data: NVIDIA’s fiscal 2026 revenue grew 65% to $216 billion, and operating cash flow reached $103 billion. Among this, data center revenue grew 68% to $194 billion.

From a market share perspective, NVIDIA’s dominant position in the AI accelerator market remains solid. As of early 2026, NVIDIA controlled about 81% to 90% of the AI accelerator and data center chip market. In the core area of AI training, its share is even higher, at around 85% to 90%. Although overall market share is expected to decline to about 75% by 2026 as AMD expands in scale and hyperscale cloud service providers deploy custom chips (ASICs), absolute revenue figures are still rising—because the expansion speed of the overall available market far outpaces any single competitor’s ability to capture it.

Monetization Coming Through: AI Is Turning From a Cost Center Into a Revenue Source

If AI infrastructure is a “shovel-selling” business, then what the market is now watching is whether the “gold miners” have truly mined gold.

In the first half of 2026, the global AI industry is moving from a “technology boom period” into a “rational implementation period.” Compute supply is diversifying, model capabilities are continuously iterating, application-layer offerings are starting to realize revenue and profits, and the token economy is shifting from an implicit cost into a visible operational variable.

Some companies are already able to integrate AI deeply into their own product ecosystems and form relatively stable revenue sources. For example, AI Copilot in enterprise office software, AI advertising recommendation systems, AI automation operations tools, intelligent customer service, and data analytics systems.

Microsoft’s annualized AI business revenue has exceeded $37 billion, up 123% year over year, and Azure’s growth has been maintained at 40%. In its first quarter, Google Cloud revenue reached $20 billion, up 63% year over year; backlog orders are nearing $462 billion, with more than half expected to be recognized as revenue within the next 24 months. Amazon’s AWS revenue in the first quarter was $37.6 billion, up 28% year over year, the fastest growth rate in nearly four years.

However, commercialization results show clear differentiation across companies. Some companies, despite having strong AI narratives, still find it difficult in the short term to prove their commercialization capabilities. The market is starting to shift from “who is laying out AI” to “who can truly make money using AI.”

A key investment conclusion in a Goldman Sachs report is that improving profit margins among hyperscale cloud vendors will make the currently high infrastructure investments more sustainable, thereby alleviating the market’s core concerns about the returns on AI capital expenditure. The report points out that operators are still constrained by supply in meeting current and future compute demand.

The Spread of the Investment Map: From Chips to the Entire Industry Chain

As the AI investment logic shifts from a compute race to commercial delivery, attention in capital markets is also expanding from single chip companies to the entire industry chain.

By observing U.S. stock fund flows from 2023 to 2026, a clear pattern of sector rotation can be identified. In the first phase (early 2023 to mid-2024), funds were highly concentrated in the compute infrastructure layer, and chip companies such as NVIDIA saw gains far outpace other tracks. Entering 2026, funds are spreading further from AI core chip stocks such as NVIDIA, Broadcom, and TSMC to a “second-layer supply chain for AI infrastructure,” including fiber optics, glass, ceramics, and advanced materials.

Goldman Sachs’ Chief Global Equity Strategist Peter Oppenheimer said that capital expenditure by hyperscale cloud service providers is expected to continue growing at a high rate. Driven by strong capital expenditure, “Phase 2” AI infrastructure stocks have surged 40% since the beginning of the second quarter. Most of the price gains in the AI infrastructure sector have been driven by earnings. However, recent valuation expansion and positioning dynamics suggest that market volatility will increase going forward. The median P/E of stocks in the AI infrastructure sector has expanded to 26 times.

In terms of specific beneficiary directions, market focus includes: GPU and AI chips, cloud computing platforms, data center construction, high-speed networking equipment, and enterprise-level AI compute services.

It is worth noting that the targets span multiple tiers. Chip tier: NVIDIA (NVDA), AMD (AMD), Broadcom (AVGO); storage and equipment tier: Micron (MU), ASML (ASML), Applied Materials (AMAT); infrastructure and server tier: Dell Technologies (DELL); energy infrastructure tier: as electricity demand for data centers surges, power companies become direct beneficiaries, including Vistra (VST), GE Vernova (GEV), and Constellation Energy (CEG).

In addition, AI application-layer companies are also attracting increasing attention. Directions such as enterprise AI, AI advertising and recommendation systems, AI agents, and content generation are all considered important tracks for the next phase.

One-Stop Allocation of AI Track Assets on Gate

For investors looking to participate in the second phase of AI infrastructure investment, Gate offers a unique entry point. On June 1, 2026, Gate officially launched real stock trading services, becoming one of the first exchanges in the industry to directly connect to the U.S. stock market within a crypto platform. Users do not need to exchange currency, make cross-border remittances, or open an additional brokerage account. They only need to use the USDT liquidity within their Gate account to trade U.S. stocks directly.

As of June 2026, Gate has supported more than 12,500 stocks (U.S. stocks, Hong Kong stocks, and Korean stocks) and ETF assets. The asset range covers major tech stocks such as Apple (AAPL), NVIDIA (NVDA), and Tesla (TSLA). The platform covers major U.S. securities trading venues such as NYSE (New York Stock Exchange) and Nasdaq, and supports fractional-share trading with a minimum order size of 0.01 shares. Minimum fee rate is 0.023%, with settlement in USDT.

This means investors can allocate crypto assets and AI concept stocks on the same platform, enabling cross-asset class portfolio management.

Conclusion

AI infrastructure investment is entering a delicate phase. Capital expenditure is still growing at high speed—annual spending of $725 billion, a 77% year-over-year increase, and nearly half the size of the global semiconductor market—these figures alone show the intensity of industry expansion. But the market’s focus is shifting from “how much is being spent” to “how much is being earned.”

The industry has already crossed the first threshold—AI revenue covers infrastructure depreciation. However, judging by the expected ROIC of 5% to 8%, the continuously falling compute rental prices, and a sector median P/E ratio as high as 26 times, the market’s scrutiny of returns has only just begun.

This does not mean the end of AI infrastructure investment. On the contrary, it may be the beginning of a healthier phase—when the market shifts from concept-driven to delivery-driven, companies with real business moats and profitability will receive more sustainable valuation support. For investors, understanding this structural shift may be more important than chasing short-term volatility.

FAQ

Q1: What are the core characteristics of the second phase of AI infrastructure investment?

The core characteristic of the second phase is the shift from “compute expansion” to “commercial delivery.” The market is no longer focusing only on the scale of capital expenditure and model parameters, but is beginning to systematically examine return on investment, profitability, and progress in commercialization. In 2026, annual depreciation expenses for AI infrastructure are expected to approach $111 billion, and the industry needs to prove that revenue can cover these costs.

Q2: What are the specific capital expenditure figures for the four cloud vendors in 2026?

Alphabet (Google), Amazon, Microsoft, and Meta’s combined capital expenditure in 2026 is approximately $725 billion. Of this, Amazon is about $200 billion, Microsoft about $190 billion, Google about $175 billion to $185 billion, and Meta about $115 billion to $135 billion. This figure represents a 77% increase from $410 billion in 2025.

Q3: What risks does AI infrastructure investment currently face?

Major risks include: compute rental prices continuing to decline (NVIDIA B200 chip rental prices fell 31% within one month); expected investment returns on ROIC of only 5% to 8%; the market having already priced in too much future upside (U.S. tech investment as a share of GDP has exceeded the dot-com bubble period around 2000); and the potential impact of large-scale depreciation expenses on profits.

Q4: How can investors participate in AI infrastructure investment through Gate?

Gate has officially launched real stock trading services. Users can trade directly with USDT more than 12,500 stocks (U.S., Hong Kong, and Korean stocks) and ETFs. Covering core AI industry chain targets such as NVIDIA (NVDA), AMD (AMD), Micron (MU), etc., it supports fractional shares with a minimum of 0.01 shares and requires no additional brokerage account.