AI Token Prices Plunge, Computing Costs Climb, Investment Return Gap Emerges
Nashnova编辑部
AI model prices have been cut three times in a month, yet Nvidia's server systems are set to rise over 15% next year — the scissor between falling prices and rising costs is forcing investors to re-do the math on whether the money spent will ever come back.
Prices down, costs up — how did this scissor form?
AI model pricing is falling fast: OpenAI cut its flagship price three times in roughly a month; a free model called Ox Alpha appeared online with near-frontier performance — its developer unknown.
The Silicon Data LLM Token Expenditure Index — a blended measure of price and usage that proxies what buyers actually pay — keeps sliding.
But building AI is getting more expensive: Nvidia told some large customers its server systems with AI chips will rise over 15% when they ship early next year.
This means → the product is getting cheaper while the raw materials get pricier — margins are squeezed from both ends.
Why are upstream costs rising? What is happening in chips and memory?
Semiconductors are entering structural undersupply — not a brief shortage but capacity itself falling behind demand growth.
Samsung has raised some advanced foundry order prices by up to 15% and locked as much as 70% of its memory output into multi-year contracts with data-center buyers.
SK Group's chairman warned that memory shortages will worsen further by 2027; SK Hynix is exploring joint ventures to fund new fabs.
In plain terms = the chips and memory AI needs are not just more expensive — they are being rationed, and the squeeze will tighten for years.
What keeps the bulls confident?
The core argument is volume: cheaper intelligence means more consumption, so total revenue can still grow.
The data offers support: nearly 60% of U.S. companies now pay for AI; spending across tiers has more than tripled.
The three hyperscalers (AWS, Azure, Google Cloud) posted combined cloud revenue of roughly $106 billion last quarter, up over 40% year-on-year.
Ramp estimates heavy users spend about $7,400 per employee per month on AI, while the median company spends just $12 — this reflects usage concentrated at the top, with most firms still experimenting.
All that spending — where are the returns?
Milos Maricic, founder of AI consultancy Maximand, reviewed 919 earnings calls over three years from the 60 largest U.S. listed financial institutions.
Four out of five calls mentioned AI; over half discussed AI costs — yet only one company, on two occasions, disclosed concrete dollar returns: roughly $19 million combined.
In plain terms = the entire financial sector has talked about AI for three years and spent heavily, yet almost no one can produce an actual return receipt.
How is the credit market reading this?
Broadcom is in talks to raise over $60 billion in debt to fund its AI chip business — part of a broader AI borrowing wave.
The cost of insuring Broadcom's debt has risen about 80 basis points since January, and the premium is still accelerating.
This means → the credit market has begun pricing in the risk that AI investment may not pay for itself.
What is the unresolved question?
Goldman Sachs trading head Rich Privorotsky put it bluntly: if companies ultimately cannot finance every planned build-out, the result is either more equity dilution or less capex — neither supports a higher valuation multiple.
Nvidia's earnings offered some relief, guiding to roughly 70% revenue growth in fiscal 2028; but the company also warned that rising memory costs could compress margins.
The bull case — volume growth outrunning price declines — has not been disproven; AI returns may arrive like e-commerce profits — late, but massive.
In plain terms = whether cheap intelligence can eventually cover all costs, including this mountain of debt, is a question no one can answer yet.
市场有风险，内容仅供研究参考，不构成投资建议。