Which companies are actually included in AI stocks?

Simply put, AI stocks are not an official classification but a market term for companies whose business is related to AI^[1]. They can be roughly divided into three layers: the compute layer (selling chips and cloud infrastructure), the large model layer (developing underlying models), and the application layer (turning AI into software that makes money). Each layer has completely different business models, growth logic, and risks, so you can't lump them together.

Think of AI as a gold rush: the compute layer sells the shovels and jeans, the large model layer are the prospectors who claim to know where the gold is, and the application layer are the craftsmen who turn gold into jewelry for consumers. Each layer makes money differently and carries different risks: shovel sellers might earn steadily but face fierce competition; prospectors could strike it rich overnight or lose everything; craftsmen depend on whether consumers buy their products.

For example, Nvidia in the compute layer reported Q2 FY2026 revenue of $46.74 billion, with data center revenue alone at $41.1 billion^[4]; while Palantir in the application layer reported revenue of $1.94 billion for the same period, up 93% year-over-year^[13]. Both are AI stocks, but their size and growth rates are vastly different. This reminds us not to treat all AI stocks as a single entity when investing. You need to first figure out which layer a company belongs to, then assess its investment logic.

Why is Nvidia called the 'pick-and-shovel seller' of the AI era?

The 'pick-and-shovel' strategy dates back to the 19th-century gold rush—prospectors didn't always make money, but those selling shovels and picks did^[2]. In the AI era, Nvidia is that 'pick-and-shovel seller': it doesn't directly make AI applications, but most AI companies need its GPU chips to train and run models^[3].

Why are GPUs so important? Because training AI models requires massive parallel computing, and GPUs are built for that—they have many cores that can handle lots of simple calculations at once, like having a thousand workers moving bricks simultaneously, far more efficient than one strong person. So, most AI companies wanting to train large models rely heavily on Nvidia's GPUs.

The advantage of this strategy is that you don't have to bet on which AI company will win; as long as the whole industry grows, the shovel seller makes money. But note that there's competition among shovel sellers too—AMD's MI300/MI350 series is vying for Nvidia's market^[11]. Also, if the gold rush fades, shovels might not sell—if AI investment slows, Nvidia's high growth could quickly decline.

Retail investors can't buy OpenAI or Anthropic stock—how can they get indirect exposure?

OpenAI, Anthropic, and other top model companies are still private and not listed on US stock exchanges, so retail investors can't buy their shares directly^[3]. But you can get indirect exposure by holding shares in companies that have invested in them.

For example, Microsoft holds about 27% equity in OpenAI, with cumulative investments of $13 billion^[9]; Amazon's investment in Anthropic generated a $16.8 billion pre-tax gain in Q1 2026^[10]. Buying Microsoft or Amazon stock is like indirectly betting on these model companies.

This indirect investment is like taking a detour: you can't buy the star startup's stock directly, but you can buy the stock of its backer and share in some of the growth. However, note that this exposure is indirect, and these big companies have diverse businesses—AI is just one part—so their stock prices won't move entirely in line with the model companies. Also, Microsoft accounts for its OpenAI investment using the equity method, meaning OpenAI's profits and losses are reflected proportionally in Microsoft's income statement, so retail investors can indirectly feel OpenAI's performance swings when reading Microsoft's financials.

How much will the four major cloud companies spend on AI infrastructure in 2026?

Google, Microsoft, Meta, and Amazon—these four cloud giants—have combined capital expenditure plans of approximately $660 billion to $725 billion for 2026, up significantly from about $410 billion in 2025^[5]. Amazon plans about $200 billion, Google $175–185 billion, Meta $115–135 billion, and Microsoft $110–120 billion.

Most of this money goes to AI compute, data centers, and networking. This means demand for the compute layer is very certain in the short term, but it also suggests these giants face considerable cash flow pressure.

Think of it as four giants racing to build highways: they all believe many cars (AI applications) will come, so they're paving roads frantically. The road builders (Nvidia, Broadcom, etc.) naturally thrive, but building roads is expensive, and if traffic doesn't meet expectations, the returns on these investments may not be worth it. So, while these numbers confirm the boom in AI infrastructure, you should also watch whether these giants' capital spending is too aggressive, which could hurt their profits and stock prices.

What is CoWoS, and why is it a bottleneck for AI chips?

CoWoS is TSMC's advanced packaging technology that integrates logic chips with high-bandwidth memory (HBM), and it's standard for flagship AI accelerators from Nvidia, AMD, and others^[6]. It's like building multiple small houses (chips) on a single foundation so they can communicate at high speed.

Why is this technology needed? Because AI chips need to process massive data simultaneously; if the chip and memory are separate, data transfer speed becomes a bottleneck. CoWoS packages them together, like putting a warehouse next to a factory—raw materials are instantly available, boosting efficiency.

But CoWoS capacity is limited, making it one of the bottlenecks in AI chip supply. So TSMC's expansion progress directly affects AI chip shipments. Additionally, SK hynix is a major HBM supplier, and it's expected that all HBM capacity will be sold out in 2026^[12], showing how tight the entire supply chain is.

This is like building a road: you need both a roller (chips) and asphalt (HBM), but asphalt production is limited, so the road can't be built faster. Investors can watch these bottleneck areas because tight capacity often means strong pricing power, and related companies may benefit. But beware: if bottlenecks are resolved, oversupply could lead to price declines.

What's the difference in risk between buying AI chip stocks and AI application software stocks?

AI chip stocks (compute layer) have relatively easy-to-track earnings—like Nvidia's data center revenue or Broadcom's revenue guidance for custom AI chips (which is the company's forecast of future orders and revenue)^[7]. As long as AI demand grows, their orders are fairly certain. But the risk is that if AI investment slows, their high growth could quickly reverse.

AI application software stocks (application layer) are different—they're still validating their business models. For example, Palantir's AIP platform (a software product that helps companies integrate AI models into daily business decisions) is growing fast^[13], but whether it can sustain profitability and whether customers will keep paying long-term are uncertain. So application-layer stocks tend to be more volatile and suit investors who can handle higher risk.

Think of it this way: chip stocks are like selling building materials—as long as the real estate industry is booming, material sales are assured; application software stocks are like selling fully decorated homes—you need customers to appreciate the design, and if the market doesn't, they might not sell. So chip stocks depend more on the industry's overall investment cycle, while application stocks depend more on product competitiveness and customer loyalty. Investors should choose the layer that fits their risk tolerance.

What is 'AI washing,' and how is the SEC regulating it?

'AI washing' refers to companies exaggerating their AI capabilities or just riding the AI wave without real AI business. The SEC, NASAA, and FINRA jointly issued an investor alert in 2024, warning about AI-related fraud^[14].

The SEC has also penalized two investment advisory firms for exaggerating AI capabilities in their marketing, fining them $225,000 and $175,000 respectively^[14]. So, before buying any AI stock, make sure the company actually has real AI business—don't be fooled by the hype.

How to spot 'AI washing'? Look at the company's financials and products: Is it really developing AI technology? Does it have actual AI products or services generating revenue? Or is it just slapping 'AI' on its PowerPoint? Also, watch out for companies that suddenly rename themselves or frequently release AI-related news—they might be chasing trends. The regulatory penalties remind us that even professional investment advisors can exaggerate AI capabilities, so ordinary investors need to be extra vigilant.