The Scale of the Commitment

The numbers circulating from Q1 2026 earnings season are striking. Amazon, Alphabet, Meta, and Microsoft collectively guided to approximately $725 billion in combined capex for calendar year 2026—a 77% increase from the ~$410 billion they spent in 2025, itself a record year. Amazon guided to ~$200 billion, Alphabet to as much as $190 billion (raised mid-season), Microsoft to ~$190 billion, and Meta to $115–135 billion (Financial Times / Tom's Hardware, April 2026). Including Oracle at ~$50 billion, the five largest US cloud and AI infrastructure providers collectively committed $660–690 billion (Futurum Group, February 2026).

Goldman Sachs has revised its cumulative FY2025–FY2030 capex estimate for the Big Four upward to $5.3 trillion from a prior $4.5 trillion figure, with a broader 2026–2031 AI infrastructure aggregate—covering compute, data centres, and power—of $7.6 trillion (Yahoo Finance / Goldman Sachs, mid-2026). Capital intensity for these firms has reached 45–57% of revenue, ratios previously unthinkable for software-and-services businesses.

Nearly all of the incremental spend is directed at AI compute infrastructure: GPU clusters, custom silicon design and procurement, and data centre construction and power delivery.

What Is Being Bought, and Who Sells It

The primary beneficiary of this spending wave has been NVIDIA. Its data centre segment generated $51.2 billion in fiscal Q3 2026, up 66% year-on-year, and now represents approximately 90% of total company revenue of $57.0 billion (MLQ.ai; Data Center Frontier, September 2025). For full fiscal year 2025, NVIDIA reported $115.2 billion in data centre revenue (Persistence Market Research).

NVIDIA holds an estimated 70–85% of AI accelerator revenue in 2026 (Tom's Hardware / Persistence Market Research). AMD's Instinct series is the nearest merchant GPU alternative at approximately 6% revenue share. The top five semiconductor companies collectively control roughly 70% of accelerator shipments, making this one of the most concentrated high-growth markets in the technology sector.

However, the composition of AI chip procurement is shifting. Custom ASIC shipments from cloud service providers are growing at 44.6% year-on-year in 2026, compared to 16.1% for merchant GPUs (TrendForce, reported by Tom's Hardware, May 2026). ASIC-based AI server shipments now represent 27.8% of the market—the highest share since 2023. TrendForce projects that custom ASICs will surpass GPUs in total shipment volume by 2028. The underlying driver is workload composition: inference tasks (running trained models at scale) favour architecturally specialised chips over general-purpose GPUs.

The Inference Inflection

This workload shift has significant structural implications. Inference workloads accounted for approximately 50% of all AI compute in 2025, rising to an estimated two-thirds in 2026 (Deloitte), and are projected to reach 75% by 2030 (Brookfield). JLL's 2026 Global Data Center Outlook identifies 2027 as the critical inflection point when inference overtakes training as the dominant data-centre AI requirement.

Inference economics differ materially from training economics: workloads run continuously at massive scale, are latency-sensitive, and reward energy-efficient, workload-specific architectures. The shift validates the case for custom silicon and changes the competitive calculus for every chip vendor in the ecosystem. Google Cloud's contract backlog reaching approximately $460 billion in Q1 2026—roughly double the ~$240 billion at end-Q4 2025—alongside Q1 2026 revenue growth of 63% year-on-year to $20 billion (Tom's Hardware, April 2026) illustrates what monetised AI infrastructure demand looks like.

Export Controls: A Structural Complication

US export controls have introduced a persistent and increasingly complex overlay on the semiconductor investment landscape. NVIDIA's AI chip market share in China has fallen from over 90% to approximately 50% as of early 2026 (Oplexa / Semiconductors Insight, 2026), reflecting a combination of the 25% Section 232 tariff on qualifying semiconductors, Beijing's domestic procurement mandates, and the January 15, 2026 BIS rule shift from a presumption of denial to case-by-case licensing for advanced AI chip exports to China (Morgan Lewis, January 2026).

In June 2026, the Department of Commerce issued guidance clarifying that licensing requirements apply to all businesses with a headquarters or parent company in China—including subsidiaries operating outside China—closing a loophole that had permitted controlled Blackwell shipments to reach Chinese-headquartered entities abroad (Al Jazeera, June 2026).

Separately, TSMC, Samsung, and SK Hynix moved from automatic exemptions to annual export licence renewals for their Chinese fab operations as of January 1, 2026 (Semiconductors Insight, 2026), introducing recurring policy renewal risk for global semiconductor supply chains.

The policy environment is itself unstable. The AI OVERWATCH Act—which would grant Congress veto power over AI chip export licences—passed the House Foreign Affairs Committee on January 22, 2026 (East Asia Forum, March 2026). Separately, a bipartisan group of eight lawmakers called in February 2026 for a blanket ban on all semiconductor manufacturing equipment exports to China (Semiconductors Insight, 2026). CSIS has noted that export controls reduce revenues critical to sustaining the high R&D levels that have historically underpinned US semiconductor leadership (CSIS analysis, 2026).

The Market Forecast Context

The AI Accelerator Chips market was valued at $45.8 billion in 2025 and is projected to reach $746.2 billion by 2035 at a 32.18% CAGR, with the ASIC sub-segment forecast at approximately 43% CAGR over the same period (SNS Insider, May 2026). The broader AI infrastructure TAM is projected to expand from approximately $160 billion in 2025 to $280 billion by 2027, a roughly 25% base-case CAGR (SparkCo.AI).

One important counterpoint: Allianz Research (March 2026) projects AI capex growth will decelerate sharply—from 51% in 2026 to 13% in 2027 and approximately 5% in 2028. Infrastructure built today may take 18–36 months to generate proportional returns (Futurum Group). Hyperscalers issued $108 billion in new debt in 2025 alone, and Morgan Stanley and JP Morgan project the technology sector may need to issue $1.5 trillion in new debt over the coming years to sustain current investment trajectories (Introl / Quantflow Lab).

Framing the Questions

The AI infrastructure build-out of 2025–2027 is arguably the largest single capital cycle in semiconductor history. The demand signal from hyperscaler commitments and cloud backlog data is credible. The competitive dynamics—ASIC displacement, inference workload growth, export control risk—add texture and nuance to what the headline capex numbers suggest. Understanding the interaction between these forces is the analytical work that remains ongoing.

This post is produced by 1Oak Research for general informational and educational purposes only. Nothing in this post constitutes investment advice, a solicitation, or a recommendation to buy, sell, or hold any security or financial instrument. All investments carry risk, including the risk of total loss of capital. Past growth trends and analyst projections are not guarantees of future performance. Readers should conduct their own due diligence and consult a licensed financial adviser before making any investment decision. 1Oak Research is not a licensed or regulated financial entity.