Hana Securities researcher Lee Young-joo stated on the 25th that AI investment productivity is now as important as capital expenditure scale, as the market shifts focus from 'how much was spent' to 'whether the spending delivers value.' Bloomberg's LLM token cost index, which measures the average cost to process one million inference tokens, fell from 5-6 month highs to recent 3-4 month levels, signaling efficiency gains through model optimization and price reductions. According to Exponential View, AI-related revenue excluding China exceeded datacenter and AI chip depreciation costs for two consecutive quarters, indicating that existing infrastructure is beginning to generate returns. This productivity shift affects the entire AI ecosystem: hyperscalers' continued investment depends on profitability, which in turn sustains growth for semiconductor, datacenter, power, and network companies interconnected in the AI supply chain.
Market Shifts Focus from CAPEX Scale to Investment Efficiency
Hyperscaler AI investment CAPEX continues to grow, with Morgan Stanley recently raising long-term investment forecasts. Lee stated that the market's focus is transitioning: previously the question was how much CAPEX would increase, but now the priority is how efficiently that investment is being utilized and whether it secures enough economic viability to sustain future investment. This shift directly impacts semiconductor and power investors, as the AI ecosystem operates as an interconnected belt — semiconductor, datacenter, power, and network companies' growth forecasts rest on the premise that hyperscalers will continue spending. If hyperscalers close their wallets, the entire industry's growth story weakens. Conversely, if hyperscalers generate sufficient revenue from AI business, that revenue funds further investment, which flows back to semiconductor and power companies' earnings.
Bloomberg Token Cost Index Falls from 5-6 Month Highs to 3-4 Month Levels
Bloomberg's LLM token cost index, which tracks the average cost to process one million inference tokens, peaked in 5-6 months and recently declined to 3-4 month levels. This decline does not mean people are using AI less; rather, it reflects structural efficiency: users are selecting cheaper models suited to specific tasks, optimizing inference processes, and benefiting from falling model prices, enabling the same service delivery at lower cost. Lee likened this to a restaurant maintaining the same customer volume while improving kitchen workflow and reducing ingredient costs to boost margins — revenue may not increase, but profit does.
AI Revenue Exceeds Depreciation Costs for Two Consecutive Quarters
Exponential View reports that AI-related revenue excluding China surpassed the depreciation costs of currently operational datacenters and AI semiconductors for two consecutive quarters. This figure does not yet account for future datacenter construction or additional GPU procurement costs, but it demonstrates that existing infrastructure is beginning to cover its own expenses. Microsoft cited Azure and Copilot AI service demand as driving cloud growth, while Alphabet identified AI products and AI infrastructure demand as core drivers of Google Cloud growth. Although full investment recovery has not been achieved, AI is now contributing tangibly to revenue and cash flow improvement.
Meta Explores Leasing Computing Resources to External Clouds
Meta is expanding its own computational capacity through new datacenters while also exploring leasing some computing resources to external clouds. This approach maximizes utilization of existing infrastructure and creates a new revenue stream. If infrastructure can generate external revenue, the capacity to sustain future investment increases accordingly. Lee concluded that market attention will shift from AI investment scale itself to 'investment productivity,' and if hyperscalers achieve cost efficiency and profitability, the AI ecosystem's investment cycle across semiconductors, datacenters, power, and networks will become more robust.
FAQ
What did Hana Securities say about AI investment on the 25th?
Hana Securities researcher Lee Young-joo stated on the 25th that AI investment productivity is now as important as CAPEX scale, with the market shifting focus from 'how much was spent' to 'whether the spending delivers value.'
Why did Bloomberg's LLM token cost index decline?
Bloomberg's LLM token cost index fell from 5-6 month highs to recent 3-4 month levels due to users selecting cheaper models for specific tasks, optimizing inference processes, and benefiting from falling model prices, enabling the same service delivery at lower cost.
How does AI revenue compare to infrastructure costs?
According to Exponential View, AI-related revenue excluding China exceeded the depreciation costs of currently operational datacenters and AI semiconductors for two consecutive quarters, indicating that existing infrastructure is beginning to cover its own expenses.
