In early May, NVIDIA announced two major investments: one in Corning, a fiber optic manufacturer, and another in IREN, a data center power operator. This global leader in GPUs has invested over $9 billion in the fiber optics and power sectors this year alone.

Around the same time, the four cloud giants—Google, Microsoft, Amazon, and Meta—released their financial results. Their combined capital expenditures for 2026 are projected to reach up to $725 billion, a 77% increase from last year's $410 billion, surpassing market expectations of $670 billion. Financial reports from TSMC, NVIDIA, and AMD also exceeded expectations across the board. IDC predicted in April that the global semiconductor market would reach $1.29 trillion in 2026, a 52.8% year-over-year increase.

However, beyond these impressive revenue figures, less obvious numbers are scattered throughout the financial reports. During Microsoft's earnings call, the CFO revealed that rising memory chip prices alone consumed $25 billion of the capital budget. Research data shows a sharp increase in fiber optic prices, while the power supply gap, water consumption, and land costs for U.S. data centers have reached significant levels.

These phenomena collectively indicate that the semiconductor industry is undergoing structural changes, with the boundaries of this transformation extending far beyond chips themselves.

01

Physical World Capacity Becomes a Bottleneck

The rapid growth of cloud revenue underpins the massive investments by these tech giants. In the first quarter of 2026, Google Cloud's revenue surpassed $20 billion for the first time, with a year-over-year increase of 63.4%. Microsoft Azure's cloud computing business grew by 40%, while Amazon AWS reported revenue of approximately $37.6 billion, a 28% increase—marking its fastest growth in 15 quarters. Combined, the cloud businesses of these four companies generated over $70 billion in quarterly revenue, a year-over-year increase exceeding 40%.

However, during their earnings calls, these companies repeatedly emphasized not a lack of demand but constraints in physical world capacity. Google's order backlog for the quarter exceeded $462 billion, nearly doubling sequentially. Microsoft's backlog reached $627 billion, nearly doubling year-over-year, while Amazon's stood at $364 billion. Microsoft CEO Satya Nadella stated bluntly that cloud computing capacity "will remain constrained at least through the end of this year."

To fulfill these orders, the four companies have significantly increased their capital expenditures. Microsoft's capital spending for the 2026 calendar year is projected to reach $190 billion, exceeding analysts' expectations of over $150 billion and representing a 61% increase from last year. Amazon's annual capital expenditures are expected to reach approximately $200 billion, with $43.2 billion spent in the first quarter alone. This led to a decline in its free cash flow over the past 12 months from $25.9 billion year-over-year to $1.2 billion, as property and equipment purchases increased by $59.3 billion year-over-year. Google raised its full-year capital expenditure guidance to $180-190 billion, including $4.75 billion from the acquisition of clean energy developer Intersect Power. Meta increased its full-year capital spending projection to $125-145 billion, a $10 billion increase from previous expectations.

From a quarterly perspective, the combined capital expenditures of these four companies exceeded $130 billion in Q1, a 73% year-over-year increase. Based on full-year guidance, just Microsoft and Amazon's capital spending will approach $400 billion, exceeding the annual defense budgets of most countries worldwide. Goldman Sachs noted in a late-April report that the baseline model for AI capital expenditures in 2026 is $765 billion, projected to rise to $1.6 trillion by 2031. Bank of America, in an early-May report, offered an even more aggressive forecast: global hyperscale capital expenditures will exceed $800 billion in 2026 and surpass $1 trillion in 2027. Cumulative global AI infrastructure investments from 2026 to 2031 are expected to reach $7.6 trillion.

02

Hidden Costs in Capital Budgets

Amid the surge in capital expenditures, a critical detail emerges: not all of this money translates into increased computing power. A significant portion is being absorbed by rising prices of semiconductor components.

Microsoft CFO Amy Hood explicitly stated during the earnings call that of the $190 billion in capital expenditures, $25 billion was attributable to price increases in components like memory chips. She further revealed that for the fourth fiscal quarter, capital expenditures are projected at $40 billion, with $5 billion driven by chip price hikes. Meta also cited rising component prices as a key reason for increasing its capital expenditure forecast. Bank of America's report noted that the current environment grants semiconductor suppliers significant pricing power, with cost increases for wafers, memory, and substrates being passed on to downstream customers.

TrendForce data indicates that in the second quarter of 2026, DRAM contract prices are expected to rise by 58-63%, while NAND contract prices will increase by 70-75%. The contract price for 64GB RDIMM server memory surged from $450 in Q4 2025 to over $900 in Q1 2026, doubling in six months. Samsung and SK Hynix have warned that AI-driven memory shortages could persist beyond 2027. Due to soaring HBM demand, customers are booking supplies years in advance, and even by the end of next year, global memory capacity will only meet 60% of demand. NVIDIA's latest Vera CPU, equipped with eight SOC AMM modules, triples maximum memory capacity, with each new GPU generation amplifying demand for memory chips exponentially.

AI demand is severely squeezing the traditional consumer electronics market. A report from the Center for a New American Security notes that by 2026, data centers are expected to consume approximately 70% of global memory chip output, with memory shortages causing the global PC market to shrink by 11% and the smartphone market by 13%. Apple CEO Tim Cook also warned in a recent earnings call about rising memory costs potentially impacting business. Harvard Business School professor Willy Shih, in an interview with Marketplace, pointed out that AI servers typically use 10 times more memory than traditional data center servers, stating, "The scale of the AI impact is completely different from previous cycles." IDC market research analyst Ryan Reith warned that significant costs will be passed on to consumers, "and this will only intensify in the second half of this year."

03

The Rise of Self-Developed Chips

While NVIDIA continues to dominate AI computing power—with its FY2027 Q1 revenue guidance reaching $78 billion, a 77% year-over-year increase—the self-developed chip strategies of the four cloud giants are undergoing a qualitative transformation. They no longer view self-developed chips merely as internal tools to reduce reliance on NVIDIA but are now commercializing them as independent products.

Google announced a landmark shift in its earnings report: it will begin selling TPU hardware to external customers. On May 5, Anthropic committed to spending approximately $200 billion on Google Cloud over the next five years, accounting for over 40% of Google Cloud's total order backlog of over $460 billion. The computing power Google provides to Anthropic is primarily based on its self-developed TPUs rather than NVIDIA GPUs, offering higher profit margins for Google. Its latest-generation TPU 8i delivers 80% better cost-performance than its predecessor, and Google expects to begin recognizing some hardware sales revenue later this year, with the majority realized in 2027. Notably, Anthropic's and OpenAI's cloud computing contracts combined account for nearly half of the approximately $2 trillion in long-term contracts held by AWS, Azure, Google Cloud, and Oracle—two cash-burning AI companies are now supporting the core narrative of cloud giants' revenue growth.

Amazon's chip business expansion is equally noteworthy. Its chip portfolio, including Graviton CPUs, Trainium AI chips, and Nitro, has surpassed an annualized revenue run rate of $20 billion, maintaining triple-digit year-over-year growth. Amazon CEO Andy Jassy revealed in the earnings call that OpenAI has committed to consuming approximately 2GW of Trainium computing power. If Amazon's chip business were calculated independently, its ARR would reach $50 billion, with over $225 billion in revenue commitments. The Graviton CPU offers 40% better cost-performance than traditional x86 architectures, attracting interest from external clients like Meta.

Microsoft's self-developed chip deployments are relatively low-key but advancing, with its Maia 200 AI chip and Cobalt CPU already deployed internally. Meta announced it has deployed over 1GW of self-developed chip computing power while also purchasing significant hardware from AMD and Broadcom.

TrendForce data provides a quantitative perspective on this trend: in 2026, GPU-based AI servers will account for 69.7% of shipments, while ASIC-based (including TPUs, Trainium, and other self-developed chips) servers will rise to 27.8%, up nearly 10 percentage points from 18% in 2024. Self-developed chips are moving from the periphery to the mainstream. Meanwhile, in the evolution of AI infrastructure, CPUs are experiencing a resurgence. Intel CEO Chen Liwu noted that the traditional CPU-to-GPU ratio of 1:8 has now risen to approximately 1:1. ARM predicts that in the age of intelligent agents, demand for CPUs per GW of data center capacity will increase fourfold. AI is not only driving demand for GPUs and memory but also redefining the component ratios across entire server architectures.

04

Infrastructure Under Extreme Pressure

As the semiconductor industry ramps up production, physical world infrastructure is becoming a constraining factor, beginning to influence chip deployment timelines and investment returns.

First, consider fiber optics. The massive east-west traffic generated by AI training and inference (data exchange between GPU clusters) has increased fiber demand in modern AI facilities to 10-36 times that of traditional data centers. Reports indicate that G.657.A2 fiber prices have surged from 32 yuan per core-kilometer last year to 240 yuan, a 650% increase. Analysts project global fiber demand will exceed 800 million core-kilometers this year, with a supply gap of 5-10%. AI-driven fiber demand will rise from less than 5% in 2024 to over 30% by 2027.

The severity of the fiber bottleneck has prompted chip giants to intervene directly. On May 6, NVIDIA announced an investment of up to $3.2 billion in Corning to support the construction of three new optical manufacturing facilities in the U.S., increasing capacity tenfold. NVIDIA aims to replace 5,000 copper cables in its AI rack systems with fiber optics, achieving co-packaged optics technology with 5-20 times lower power consumption than copper transmission. In March, NVIDIA had already invested $4 billion in Coherent and Lumentum. Combined with Meta's $6 billion expansion agreement with Corning and Microsoft's locked-in dark fiber contracts exceeding $8 billion, investment commitments in the fiber sector alone have surpassed $20 billion. However, the expansion cycle for fiber preforms is 18-24 months, with delivery lead times now reaching 60 weeks—the longest since the dot-com bubble.

Next, consider power supply. U.S. Department of Energy data shows that data centers consumed approximately 4.4% of total U.S. electricity in 2023, projected to rise to 6.7-12% by 2028. Goldman Sachs research indicates that in 2025, U.S. data center capacity demand will exceed supply by approximately 11.4GW, a 43% gap. PJM, the largest U.S. grid operator, warned in early May that substantial power shortages could emerge as early as 2027 and is now considering reforms to its power market structure. Tech giants are beginning to build energy assets directly: in March 2026, Alphabet acquired clean energy developer Intersect Power for $4.75 billion; on May 7, NVIDIA invested up to $2.1 billion in data center operator IREN to deploy up to 5GW of AI infrastructure. Rabobank described this trend as "data centers building a parallel energy system in the U.S."

Water resources and land costs are equally critical. Mordor Intelligence data shows that North American data centers consumed nearly 1 trillion liters of water in 2025, equivalent to New York City's annual water consumption. Over a dozen investment firms are pressuring Amazon, Microsoft, and Google to disclose more detailed water usage data. Regarding land, prices in hotspots like Ashburn, Virginia, have soared to $4 million per acre. A Redfin survey found that 47% of Americans oppose building AI data centers nearby, with community opposition forcing the abandonment of multiple multibillion-dollar projects.

05

Conclusion

The financial reports of the four cloud giants reveal a clear reality: the AI computing boom has moved beyond mere GPU procurement, now encompassing memory, hard drives, CPUs, and underlying physical infrastructure. Google's first-party model APIs consume 16 billion tokens per minute, a 60% sequential increase; Amazon's Q1 token consumption exceeded the total of all previous years combined by 10 times. This rapidly growing inference demand is driving the semiconductor industry into a sustained expansion cycle.

However, rapid investment growth comes with structural concerns. Goldman Sachs notes that the economic lifespan of AI chips (typically 4-6 years) is the single largest variable affecting cumulative investments. The core customers supporting cloud giants' order backlogs—Anthropic and OpenAI—remain mired in losses, with their $200 billion-level commitments representing the greatest uncertainty. Oracle's stock has fallen 45% since announcing its $300 billion OpenAI agreement. When fiber deliveries are delayed by 60 weeks, memory prices double in six months, and power gaps reach 43%, the semiconductor industry faces not just technical and capacity challenges but systemic issues involving energy, the environment, and social consensus. Resolving these problems will become the next target for cloud giants' massive chip industry investments.