1. Where does the AI infrastructure boom come from?

Have you ever thought about what the difference is between buying an NVIDIA graphics card and participating in the largest arms race in human history?
Answer: No difference.
The essence of AI infrastructure is a "computing power arms race"—whoever has more computing power first can train stronger models, attract more users, and earn more advertising and cloud revenue. In this race, infrastructure "sellers" like NVIDIA, TSMC, Micron, SK Hynix, and power companies are guaranteed profits regardless.
The main players in this war are four companies: Meta, Microsoft, Google, and Amazon. Their annual capital expenditures (CapEx) directly determine the financial reports of AI infrastructure stocks. When CapEx accelerates, financials improve; when CapEx slows down, stock prices cry.
So, the real question is: Can CapEx continue to accelerate?
📷
To answer this, three things need to be clarified.

2. The first thing: Do they have the money to burn?
Let's look at the numbers to give you a sense of the scale:
2023: The four giants combined CapEx about $155 billion
2024: Soaring to $251 billion, a 62% year-over-year increase
2025: Expected to surpass $380 billion, up about 50%
2026: The latest forecast reaches $725 billion, a 77% increase over 2025
📷
Specifically, for each company’s 2026 CapEx forecast—Amazon about $200 billion, Microsoft about $190 billion, Google about $175-185 billion, Meta about $115-135 billion.
Goldman Sachs even projected a staggering figure: these four companies alone will have a cumulative CapEx exceeding $5.3 trillion from 2025 to 2030.
Where does this money come from?
Free cash flow + credit expansion.
These companies are essentially printing money. The operating cash flow generated annually by Microsoft, Amazon, Google, and Meta exceeds $500 billion.
Moreover, they have the best credit ratings globally.
If the AI four giants (Microsoft, Google, Amazon, Meta) truly adopt a "debt and leverage" mode, they could mobilize incremental funds up to $1.5 trillion to $2 trillion.
Therefore, the most critical feature of this cycle is: these companies are not short of money; what they lack is a "reason not to spend"—and currently, they cannot find one.

3. The second thing: Is token consumption crazy?
The money is in place, but spending is to support AI demand growth. How fierce is that demand?
One increasingly important indicator: token consumption.
Every AI conversation, every code generation, every AI agent’s automatic decision involves countless tokens burning at a rapid pace.
Google Gemini: Q4 2025 daily API token processing volume about 14 trillion tokens
OpenAI: October 2025 daily average about 8.6 trillion tokens
Microsoft: AI annual recurring revenue (ARR) has reached $37 billion, a 123% YoY increase
Amazon Bedrock: Q1 2026 token consumption in a single quarter exceeds the total for all of 2025
📷
Goldman Sachs research predicts: as AI agents are widely deployed, from 2026 to 2030, global monthly token consumption will grow 24 times, soaring from 50 trillion per month to 120 quadrillion (120,000 trillion).
Even more interesting is a "Jevons Paradox" phenomenon: the price per million tokens plummeted from $60 in 2023 to less than $1.50 in 2025, a drop of over 96%.
But total token consumption, tracked by Apollo Global, has doubled since the end of 2025.
The cheaper the price, the more consumption, and the greater the total expenditure—this is the flywheel logic of computing demand.
Tokens become cheaper, AI agents run more aggressively, and total compute consumption actually increases.

4. The third thing: Has ROI improved?
Ultimately, whether CapEx can continue to accelerate depends on whether the investment pays off—after all, boards are not charities.
Good news: the ROI inflection point is approaching.
Meta’s ad revenue in Q1 2026 grew 33% YoY, the fastest growth in recent years, driven by AI-driven precise recommendations.
Google search ads grew 19% YoY, also reaching multi-year highs.
Cloud gross margins are expanding—more inference workloads mean higher billing density.
Goldman Sachs predicts that as the cost per token declines at 60-70% annually, cloud providers will soon see a margin inflection point.
📷
Deeper logic: when giants bundle AI inference as a service in their cloud offerings, tokens become money-printing machines—users pay for tokens, giants convert tokens into revenue, then use revenue to buy more GPUs, creating a cycle that accelerates the flywheel.
Of course, risks are real. Deloitte surveys show that less than 28% of companies worldwide currently see clear, quantifiable returns from AI investments, and nearly half believe ROI will take at least three years to materialize. This "investment-to-return time lag" is the biggest uncertainty in the story.
The overall industry ROI is still relatively low—about 5%-8%, just approaching the cost of capital.
This means:
AI has proven demand. But it has not yet proven it can support trillions of dollars in investment.
So, what Wall Street is truly watching is not model parameters, but ROI.
If in the next few years, the long-term ROI reaches 15%-20%, what does that imply?
It means: today’s AI might just be the beginning.
Because: Google Search, AWS, Meta’s ad system—these super businesses historically had ROI in this range.

5. In summary, these three key points determine how far the bull market can go
To sum up, the ultimate outcome of the AI infrastructure bull market depends on these three variables:
① Giants’ financial strength: ✅ Currently worry-free
Strong free cash flow, ample credit expansion space, the four giants will collectively invest over $700 billion by 2026. Money is not the issue; what matters is what the money is spent on.
② Token demand: ✅ The flywheel has already started
Inference consumption is becoming the core driver of the AI economy, and the explosion of AI agents will push token consumption into exponential growth. The Jevons paradox is playing out perfectly in the compute industry—more affordable, more consumption.
③ ROI improvement: ⚠ The inflection point is near but not fully realized
Signs of cloud gross margin expansion are emerging, ad and cloud revenues are accelerating, but enterprise AI ROI still needs time to materialize. This is the market’s biggest doubt about giants’ CapEx and the main source of short-term volatility.

6. Conclusion: The best people are those selling shovels
In the gold rush, the most profitable are not the miners but the shovel sellers.
In the AI infrastructure bull market, the most stable logic is not betting on which big model wins, but betting on the acceleration of this arms race itself.
As long as giants keep spending, token demand keeps surging, and ROI keeps improving, this arms race will not stop. The sellers of shovels will continue to beat expectations.
The real moment to worry is when a giant suddenly announces, "We are reducing CapEx."
📷
That day might be a turning point or just a buying opportunity.
Because if these three key numbers continue upward, the AI infrastructure bull market may just be entering its second half.
After all, in human history,
the greatest bubbles often grow quietly while being called bubbles.
Data sources: Goldman Sachs, Apollo Global Management, Deloitte, CNBC, company financial reports and investor guidance, data as of Q1 2026.