Super-large cloud service providers are investing billions of dollars in data centers, and Broadcom ( Broadcom ), which designs and customizes AI chips for them, has always been one of the biggest beneficiaries. According to the company's financial reports, since the end of 2023, Broadcom has achieved approximately $68 billion in AI semiconductor revenue. In the third quarter of fiscal year 2026, which ended on August 2, AI semiconductor revenue more than tripled compared to the same period last year, reaching $16.7 billion.

But on Monday, Broadcom CEO Hock Tan compared many investors attracted by the AI craze to gamblers, and as there were debates on Wall Street about whether the AI bubble would burst, he questioned the economic viability of AI's laboratories.

According to Bloomberg, he stated at the DBS Asia Future Forward Conference held in Singapore: "Many investors have no idea what they are actually investing in." He also said that they "set extremely aggressive expectations and found themselves in a situation almost akin to being in a casino."

Broadcom supplies the hardware required to build systems for cutting-edge AI laboratories. However, although Broadcom can profit from selling chips, the AI laboratories that purchase computing power still must find sufficient paying customers to cover the costs of developing and running models.

The challenge lies in how to transform those demands into sustainable revenue. Tan acknowledges that the $5 billion in annualized revenue disclosed by OpenAI as of the end of September indicates that there is a viable business model for the AI laboratories, but this does not guarantee profitability. Annualized revenue is an extrapolation of the company's current sales pace over a full year and is not equivalent to the actual income collected within 12 months, and the laboratories still have to cover related expenses. OpenAI estimates that training costs alone will exceed $25 billion this year.

Tan said, "Currently, it is still losing money because they have spent a lot on training. So to be honest, this figure doesn't really say much at the moment, except that by providing cutting-edge generative AI models, there is great potential for future revenue growth."

The next bottleneck for AI could be funding.

The development costs of AI have also raised a question: whether investors will continue to buy in. Some analysts expect that the growth rate of AI capital expenditure will slow down next year. Tan indicates that a shortage of funds "may" be the biggest constraint faced by the AI industry, even more so than shortages of memory, semiconductors, or electricity.

There are signs in the market that investors are questioning the scale of funds requested by AI infrastructure companies. Bloomberg reported that Firmus, an Australian data center company supported by NVIDIA, withdrew its IPO plan on October 9 due to weak investor demand for its proposed issue price.

This does not necessarily mean that the computing power demand for AI is declining. Bloomberg quoted Lip - Bu Tan, the CEO of Intel who also speaks on Tan, as saying that the shortage of hardware supply may continue for another two to three years. However, the demand for chips does not mean that companies that purchase them will earn enough money to justify their investment.

The company is “wasting money” on the AI experiment.

The concerns of Tan also extend to enterprises that pay to use AI.

He criticized these companies for “wasting money” on tokens, merely to experiment with AI, without figuring out in which scenarios using AI would actually yield returns. He compared it to children throwing arcade tokens into games only to end up with nothing in return.

He said, "When you were a child, you went to the video game arcade, and your parents gave you 10 dollars in tokens; you used them up in 10 minutes, and you didn't learn anything in return. That's exactly what's happening in companies today."

His concerns resonate with a AI investment return survey report released by McKinsey in August. Although 80% of the respondents stated that AI improved personal productivity, only 37% said that it contributed to their organization's earnings before interest and taxes (EBIT). Approximately one-fifth of the respondents indicated that AI related operating expenses, including token costs, limited their organization's use of this technology.