Source: Visual Capitalist — Charted: The $448B AI Spending Surge by Big Tech

Five companies — Alphabet, Amazon, Meta, Microsoft, and Oracle — spent a combined $448 billion on AI-related capital expenditure in 2025. That's not a projection. It's already in the SEC filings. And it grew at an average annual rate of 72% since mid-2023. If you're an investor, a knowledge worker, or simply someone trying to understand where the economy is headed, this number deserves serious attention.

Visual Capitalist's recent chart, built on Epoch AI data drawn from 10-Q and 10-K filings, makes the scale visceral: combined quarterly capex hit $140.6 billion in Q4 2025 alone. That's a single quarter exceeding what these same five companies spent in all of 2021.

Where the Money Is Going

The spending breaks down into two buckets: physical infrastructure (data centers, power, cooling, land) and compute (GPUs, custom AI chips like Google's TPUs, Microsoft's Maia, and Amazon's Trainium). Both categories are accelerating, but the internal mix is shifting. As of late 2025, a growing share of hardware spend is going to inference chips — optimized for serving AI responses to users — rather than training chips used to build new models. This signals a transition from the R&D phase of AI to the deployment phase.

The three companies that drove the largest absolute increases from Q1 2022 to Q4 2025:

• Microsoft — up $30 billion per quarter, driven by Azure AI capacity and the OpenAI partnership
• Amazon — up $25 billion per quarter, with AWS betting on both its own Trainium chips and third-party model hosting
• Alphabet — up $19 billion per quarter, concentrating on Google Cloud and its Gemini model family

The Inflection Point: Mid-2023

The data shows a clear structural break after mid-2023. Before that point, capex growth was gradual and consistent with historical cloud infrastructure trends. After mid-2023 — which coincides with enterprise AI adoption accelerating following the ChatGPT wave — spending curves visibly steepened across all five companies simultaneously. This kind of synchronized investment surge across competitors is rare in corporate history. It happened with electrification in the 1910s, interstate highways in the 1950s, and internet infrastructure in the late 1990s.

The difference this time: the late-1990s internet buildout led to a bubble and bust because demand didn't materialize fast enough to justify the supply. In 2025, demand for AI compute is already there — large language model inference, enterprise copilots, and AI-assisted search are consuming GPU capacity faster than it can be built. The risk of overcapacity is real but not yet the dominant story.

What This Means for Investors

The $448 billion flowing into AI infrastructure creates identifiable winners beyond the five hyperscalers themselves:

1. Semiconductor companies

NVIDIA captures the lion's share of GPU revenue and has no meaningful competitor for high-performance training chips. TSMC manufactures virtually all the leading-edge silicon that goes into AI accelerators. ASML provides the EUV lithography machines that make TSMC's chips possible. This is a supply chain with very few substitution points — pricing power is concentrated at each node.

2. Power and cooling infrastructure

A single large data center campus can consume 100–500 megawatts of power. The AI capex surge is creating an electricity demand shock in the US, Europe, and Southeast Asia. Utilities serving data center clusters, independent power producers developing new generation capacity, and companies building liquid cooling systems are all seeing order pipelines grow faster than they can hire engineers.

3. Industrial real estate

Data center REITs (Equinix, Digital Realty, Iron Mountain) and raw land near reliable power grids are seeing price appreciation. Virginia's Loudoun County, northern Texas, and Singapore's outer ring are effectively becoming the semiconductor fabs of the AI era — strategic real estate for a strategic industry.

The Skills Dimension

Here's what the capex chart doesn't show: behind every GPU cluster sits a workforce that needs to use it effectively. The demand for people who can work with, build on, and reason about AI systems is growing at least as fast as the hardware investment. This isn't limited to machine learning engineers. Financial analysts who can evaluate AI company fundamentals, marketers who understand AI's capabilities and limits, and anyone who works with knowledge — reading, writing, reasoning — is affected.

Vocabulary and analytical language skills turn out to be unexpectedly relevant here. AI models, especially in professional settings, communicate in precise, technical English. People who have command of nuanced vocabulary — the kind tested on the GRE, used in graduate-level writing, and required in professional environments — interact more effectively with these systems and extract more value from them. The AI infrastructure boom isn't just an investment story. It's a skills story too.

The 2026 Outlook

Announced commitments for 2026 already exceed 2025 actuals. Amazon alone has committed to $200 billion in capital expenditure — nearly 60% higher than 2024, and far above Wall Street estimates at the start of last year. Microsoft's annualized run rate projects to approximately $145 billion. Alphabet has guided $175–185 billion. Meta, $115–135 billion.

The aggregate for 2026 from these four companies alone approaches $700 billion. If Oracle's spending is included and annualized, the five-company total could approach $750 billion — more than 1.5x the 2025 figure.

For retail investors, the question is not whether AI infrastructure spending is real — the SEC filings confirm it is. The question is how to position for the second and third-order beneficiaries without chasing already-priced assets. Semiconductor equipment, specialized power infrastructure, and the human capital layer (education, professional development, AI-adjacent skills training) may offer better risk-adjusted entry points than the hyperscalers themselves, whose AI capex is already largely reflected in valuations.

$448 billion in a single year. The companies building AI infrastructure are not hedging. This is a directional bet of historic scale — and the firms making it have access to data on AI adoption that no outside analyst does.