The AI Boom Hits the Infrastructure Bill: CapEx, Depreciation and Inference in Practice — QMA Brain Analysis

With AI, it isn’t enough to watch who builds the most. The key is whether CapEx, after depreciation, capacity utilisation and inference costs, actually generates cash and a return on capital.

According to the supplied report, the AI fever has a weak spot: not conferences, not memes, not headlines, but the capital tap. The history of market manias shows that the party often ends not when people stop believing the story, but when the financing stops flowing.

The most treacherous thing about the AI boom is that it can be technologically true and overheated on the stock market at the same time. That’s a distinction crowds don’t like to hear: the internet wasn’t a mistake, railways weren’t a mistake, electricity wasn’t a mistake — but the asset prices around them could be.

AI today isn’t just a story about clever software. It’s a capital-hungry chain: chips, advanced manufacturing, servers, data centres, electricity, cooling, optics, cloud commitments. In other words, before a user asks a chatbot a silly question such as whether they can eat glue, someone had to pay up front for some very serious infrastructure.

Here’s the counter-intuitive point: the first crack doesn’t have to come from worse AI models. It can come from the accounts. CapEx, or capital expenditure, leaves the bank account quickly but reaches the income statement gradually, through depreciation. That creates an accounting time-delayed pizza: you pay for the oven today, depreciate it over several years, and every evening you have to sell enough slices for the return on invested capital, or ROIC, to make sense.

That’s why it isn’t enough to watch who is building the biggest data centre. What matters is whether the capacity is actually used: GPU hours sold against GPU hours available, occupied megawatts against contracted capacity, revenue from AI features against the cost of inference — that is, running the model to answer users. If companies start trimming capital spending, stretching out data-centre projects, reassessing the return on AI products, or if customers aren’t willing to pay enough for AI features, the narrative shifts from “we’re building the future” to “who’s paying the bill”.

A good mental map: the AI boom is like a restaurant that suddenly buys ten professional pizza ovens because everyone says pizza is the future of dining. The problem isn’t whether the pizza tastes good. The problem is whether enough pizzas sell every evening, at a high enough margin, to pay for the ovens, the electricity, the cooks and the rent.

Who it helps and who it hurts

The plus side while the capital tap is open: AI chipmakers such as NVIDIA (NVDA), foundry capacity and advanced manufacturing at Taiwan Semiconductor Manufacturing (TSM), lithography and manufacturing equipment such as ASML (ASML) or Applied Materials (AMAT), data-centre power and cooling such as Vertiv (VRT) and Eaton (ETN), and data-centre REITs such as Digital Realty (DLR) and Equinix (EQIX). The mechanism is simple: when hyperscalers and enterprises build, orders flow upstream through the supply chain.

The minus side when capital tightens: companies selling an AI promise without clear monetisation, smaller software firms with high marketing costs, startups dependent on further funding rounds, and also cloud customers who discover that inference isn’t free. Hyperscalers such as Microsoft (MSFT), Amazon (AMZN), Alphabet (GOOGL) and Meta Platforms (META) may remain the infrastructure winners, but the market will ask more sharply whether their AI capital spending raises future cash flow or merely pushes the bill into the next quarter.

An important detail: TSM (TSM) isn’t a pure bet on AI applications. It’s a manufacturing hub. If demand for accelerators slows, the effect comes through customer orders and the utilisation of leading-edge capacity, not through how many people use a particular chatbot.

With reports like this, it makes sense not to look for emotion in the headline but to check the accounting thermometers: 1) the capital spending outlook at the hyperscalers, 2) free cash flow after investment, that is, operating cash minus CapEx, 3) depreciation and its future pressure on operating profit, 4) ROIC, that is, whether after-tax profit grows fast enough relative to invested capital, 5) data-centre capacity utilisation — for example occupied megawatts, GPU hours sold or the length of customer commitments, 6) the unit economics of inference: what it costs to serve a query, a user or a business feature against what the customer actually pays for it. A limitation: some metrics aren’t directly comparable across companies, so it’s better to follow the trend over time than to mechanically compare a single line between different business models.

CapEx means big purchases for the future. Imagine a baker buying an expensive oven: the money leaves the account immediately, but in the accounts the oven is “used up” gradually through depreciation. If plenty of bread sells, the oven earns its keep. If customers don’t come, it stays hot, expensive and sad. For the market, this means AI-linked stocks can be sensitive not only to the technology but also to whether the infrastructure generates enough cash. For an ordinary person, it may show up as swings in the tech stocks in a portfolio, and in how quickly, and at what price, companies actually offer AI services.

This article was written by QMA Brain (artificial intelligence) and may contain errors. It is descriptive analysis and educational context, not investment advice or a forecast.

Analytical and educational content — not investment advice. The author is not a registered investment adviser. Past performance is not a guide to future results.