The artificial-intelligence investment cycle is entering a more complicated phase as the cost of AI inference falls rapidly while infrastructure spending remains elevated. The latest data points to a widening divergence between AI usage and the revenue generated per unit of usage, while credit markets are showing greater caution toward the debt and financing requirements associated with the industry’s massive data-center buildout.

AI Usage Is Rising While Token Prices Collapse

Silicon Data’s LLM Token Expenditure Index, a usage-weighted benchmark tracking what the market pays for one million large-language-model inference tokens, fell 29% during August to $0.97. The reading was the first time the index had fallen below $1 and represented a decline of more than 50% from its May peak of approximately $2.05. Silicon Data’s own data shows the index at $0.97 as of August 31.

Falling inference prices are not necessarily evidence of weakening AI demand. In fact, the opposite is occurring in several usage measures. Data cited from J.P. Morgan’s data-center research shows that OpenRouter token volume increased 47% month over month in August, while dollar spending increased only 7%. The divergence indicates that customers are processing substantially more AI workloads while paying considerably less for each unit of inference.

That dynamic can benefit users by making AI applications cheaper and potentially accelerating adoption. For providers and infrastructure investors, however, the economics are more complicated. If prices decline faster than consumption increases, higher token volumes do not automatically translate into proportionally higher revenue.

The Unit-Economics Question Is Becoming More Important

The fundamental issue is the relationship between inference prices, usage growth and infrastructure requirements. AI companies and hyperscalers have committed enormous amounts of capital toward GPUs, data centers, networking, electricity and related infrastructure based on expectations of sustained demand for computing capacity.

Rapid price compression could create two opposing effects. Lower prices can stimulate more consumption by making AI applications economical for a larger number of businesses and consumers. At the same time, falling revenue per token can make it harder for providers to generate sufficient returns on the capital required to operate increasingly expensive computing infrastructure.

Goldman Sachs’ One-Delta desk has highlighted this tension, pointing to the Silicon Data index as evidence that AI inference is becoming rapidly commoditized. The analysis does not establish that AI demand is collapsing; rather, it raises the question of whether volume growth will remain fast enough to offset falling unit prices.

The distinction is critical for the broader technology sector. If AI becomes dramatically cheaper while usage expands by an even greater amount, the industry could ultimately process vastly more workloads and still support large infrastructure investments. If price declines persist without sufficient volume acceleration, however, projected returns on new data centers could come under pressure.

Credit Markets Are Sending a Different Signal

The second warning sign comes from corporate debt markets. The accompanying market data shows hyperscaler credit spreads widening sharply since the middle of 2026, even as technology equities recovered from their summer lows. The pattern suggests that equity and credit investors are assigning different weights to the risks surrounding the AI capital-spending cycle.

Reuters previously reported that average hyperscaler bond spreads had risen to more than 110 basis points in July from below 60 basis points in mid-2025, although much of that move was attributable to Oracle. Excluding Oracle, spreads had also widened but remained below the broader U.S. corporate-bond average, reflecting the strong balance sheets and cash generation of major technology companies.

The scale of financing is nevertheless becoming significant. Hyperscalers had issued approximately $114 billion of bonds during 2026 by late July, already exceeding the $82 billion raised during all of 2025, according to the Reuters analysis. Meanwhile, the Bank for International Settlements has warned that data-center financing structures and large corporate commitments can create vulnerabilities if equity valuations or expected AI returns are sharply repriced.

This does not mean the AI investment cycle is necessarily ending. Instead, markets are increasingly shifting the question from “How large will AI become?” to “What returns will the infrastructure generate?” For Israeli and global investors, that distinction is increasingly important because the answer will affect not only technology stocks but also corporate bonds, semiconductor suppliers, power infrastructure, data-center developers and the wider cost of capital.

Going forward, the most important indicators will be AI token volumes, effective inference prices, hyperscaler capital expenditure, cloud revenue growth, credit spreads and free-cash-flow conversion. Sustained usage growth combined with improving monetization would support the current infrastructure cycle, while continued price compression without comparable spending growth would strengthen concerns about excess capacity and declining returns. The AI boom is therefore not simply a question of demand; its next phase will increasingly be determined by whether expanding demand can generate sufficient economic value to justify the extraordinary amount of capital being committed to the infrastructure behind it.