"Is AI Really Paying Off?" Facing Value Questions... Still, Reasons for Optimism [Weekend Money]

AI Efficiency Up as Token Costs Drop

AI Revenue Outpaces Depreciation for Two Straight Quarters

Hana Securities: "Productivity, Not Scale, Is Now Key"

There is a recurring question every earnings season for big tech companies.

"How much more are you going to spend on artificial intelligence (AI)?"

Up until now, the larger this number was, the more investors welcomed it. Spending a lot of money implied a genuine commitment to the AI business, and it was expected that these funds would directly translate into revenue for semiconductor, data center, and power companies.

However, recently, the prices of AI-related stocks have fluctuated sharply, increasing overall stock market volatility. The familiar formula mentioned earlier is now showing cracks.

Lee Youngjoo, a researcher at Hana Securities, stated on July 25, "When it comes to investing in AI, productivity is just as important to examine as the size of capital expenditures (CAPEX)." He analyzed that, from now on, "it is more important to consider whether the money is being spent effectively rather than simply how much is being spent."

Asking About ‘Value for Money’ Rather Than Speed of Spending

Capital expenditures on AI by hyperscalers (massive companies operating large-scale data centers required for AI) are still on the rise. Global investment bank Morgan Stanley also recently raised its mid- to long-term investment outlook for these companies in a report.

However, as spending on AI infrastructure has already accumulated in the hundreds of billions of dollars, the market's focus is shifting accordingly. Researcher Lee commented, "Previously, the amount by which capital expenditures were increasing was the main issue. Now, the critical concerns are how efficiently the increased investments are being utilized and whether they are economically viable enough to justify further investment."

This change is also highly relevant for investors in semiconductors and power companies, as the AI ecosystem works like a single belt. The growth outlook for semiconductor, data center, power, and network companies ultimately hinges on the assumption that hyperscalers will continue to spend. If hyperscalers close their wallets, the entire industry built upon that foundation will see its growth narrative shaken.

Conversely, if hyperscalers earn enough from their AI businesses, they will reinvest those profits, and that reinvestment will flow back into sales for semiconductor and power companies. For this virtuous cycle to continue, the AI business for hyperscalers must prove to be profitable. Researcher Lee analyzed that several recent data points send encouraging signals regarding this virtuous cycle.

AI Becomes More Efficient... Is It Starting to Deliver 'Value for Money'?

The first sign is "efficiency." According to the LLM Token Spending Index tracked by Bloomberg, after peaking between May and June, the index has recently dropped to levels seen in March and April. This index measures the average cost of processing one million inference tokens.

The decrease in this index does not mean that people are using less AI. Rather, it signals a shift towards selecting cheaper models according to the situation, optimizing the inference process itself, and lowering model prices—creating a structure where the same services can be provided at a lower cost.

Simply put, it's as if a restaurant keeps the same number of customers but improves efficiency in the kitchen workflow and reduces ingredient costs, thereby increasing margins. Even if revenue doesn't grow, net profit can rise.

The second sign is "monetization." According to Exponential View, a technology and AI research institution, AI-related revenue—excluding China—has outpaced the depreciation costs (the accounting concept of spreading equipment purchase costs over usage periods) of currently operating data centers and AI semiconductors for two consecutive quarters.

These figures have yet to reflect costs for future data centers or additional GPUs that may be added. Nevertheless, considering only the existing infrastructure, Researcher Lee assessed that it is significant to see that AI services have started to pay back their costs.

In fact, Microsoft announced that strong demand for Azure and Copilot, both AI services, has been driving cloud growth. Similarly, Alphabet cited demand for AI products and AI infrastructure as key drivers of Google Cloud's growth. While these companies have not yet fully recouped their massive investments, AI has begun to contribute to sales and cash flow improvement.

Meta's recent actions can be seen in the same context. While Meta is expanding its in-house computing capacity through new data centers, it is also considering leasing some of its computing resources to external clouds. This approach aims to maximize utilization of existing infrastructure and create new revenue streams. If infrastructure can also generate external sales, the resources available for further investment increase accordingly.

Researcher Lee stated, "Ultimately, the market's attention will shift from the scale of AI investment itself toward the 'productivity of investment'. If hyperscalers succeed in driving cost efficiency and profitability, the virtuous investment cycle across the AI ecosystem—spanning semiconductor, data center, power, and network companies—has a higher likelihood of becoming even more robust."

This content was produced with the assistance of AI translation services.