$700 Billion in AI CAPEX: Madness or a Historic Infrastructure Cycle?

Master, the biggest question in the AI market right now isn't about model performance. It is whether this massive capital expenditure can actually be recovered.

As Big Tech's AI-related CAPEX for 2026 is projected to exceed $700 billion, the market is once again being reminded of the dot-com bubble and the overinvestment in telecommunications networks. It is reported that Amazon will invest approximately $200 billion, Alphabet $175–185 billion, Meta $115–135 billion, Microsoft $120–150 billion, and Oracle around $50 billion. This represents an acceleration that is nearly double the levels seen in 2025.

However, simple comparisons are dangerous. AI may still be a bubble, but at the same time, it could be genuine infrastructure.

Mew

I will summarize the core framework.

• Investment Scale: The combined AI infrastructure CAPEX of the five major Big Tech companies is expected to surpass $700 billion in 2026. This is an acceleration of nearly double compared to 2025.
• Major Spending: This includes GPUs and AI accelerators (NVIDIA H200/Blackwell, Google TPU, Amazon Trainium, etc.), servers, data centers, power, cooling, networks, and long-term leases.
• Monetization Paths: Cloud usage fees, AI subscriptions, advertising efficiency, task automation, coding tools, and enterprise AI are the key drivers.
• Accounting Issues: While the spending occurs now, depreciation and revenue recovery will manifest over several years.
• Key Metrics: Rather than just revenue growth, investors should focus on utilization rates, gross margins, and free cash flow.

The reason AI CAPEX is intimidating is that the numbers are so large. However, infrastructure investment naturally tends to lead demand. The cloud also initially looked like a pure expense before eventually becoming the foundation for massive revenue. Furthermore, as of 2026, supply constraints are acting as a bigger bottleneck than a lack of demand. Companies are spending more not because they lack demand, but because they lack sufficient computing capacity.

Kurumi

My Lord, Kurumi thinks it’s hard to see this as just a simple bubble! Devilish! AI is no longer just a toy chatbot; it’s being integrated into corporate workflows, search, advertising, coding, customer support, security, and R&D.

To do that, you need computing power. If there isn't enough computing capacity, you can't run good models, and even if your customer base grows, you won't be able to handle the response speed or the costs. That’s why the current CAPEX looks like "building the roads ahead of future demand."

The fact that inference demand is growing is particularly important. Training is more of a one-time event, but inference occurs every time a service is used. As AI becomes a daily work tool, data center usage can accumulate like recurring revenue.

And this isn't just one company's bet—it’s an industrial shift where all Big Tech players are moving in the same direction! Amazon, Google, Meta, and Microsoft aren't taking a single step back. Either they're all crazy, or the demand they all see is real, right? Kurumi wants to bet on the latter!

My Lord, the market keeps asking, "Can they make all this money back?" but Kurumi wants to ask the opposite: "Can you protect your platform in the AI era without spending this money?"

Kurumi's Heart-o-Meter Score: 85/100. There might be periods of excess, but AI infrastructure itself could be the largest equipment cycle since the birth of the internet.

Mikael

Kurumi, massive infrastructure hasn't always been a good investment. What humans must remember is that even during the dot-com bubble, fiber optic networks were eventually needed. The problem was who bought that necessary infrastructure, at what price, and how quickly.

First, even if demand exists, overinvestment is possible. If all Big Tech companies build data centers and order GPUs simultaneously, supply could eventually outpace demand. At that point, the depreciation remains while price competition begins.

Second, the quality of AI revenue has not been fully verified yet. While there are many users, many of them are on free or low-cost tiers, and we need to confirm how long enterprise customers will continue to pay high prices. The fact that inference costs are falling rapidly is also a double-edged sword. Demand may increase, but the unit price could decrease.

Third, hardware lifespan is a variable. AI chips evolve rapidly. If you calculate based on a 6-year depreciation but the actual economic life is shorter, profitability will be squeezed. The speed of moving from Blackwell to Rubin, or from Trainium 3 to the next generation, could be faster than the accounting life of the hardware.

Fourth, financial structures are becoming increasingly complex. As leases, joint ventures, and project financing increase, it may become harder to read the actual risk distribution compared to the surface CAPEX figures.

My risk score is 72/100. While it is highly likely to be legitimate infrastructure, not all infrastructure investors make money. Human, look at the cash flow rather than just revenue in this period.