AI Token Costs Shift Focus From Chips To Returns in 2026
AI spending is getting a cleaner unit of account: the token. On July 7, 2026, the useful signal was not that chips bounced or that one memory producer printed another record quarter. It was that usage, price, and capital intensity are starting to tell different stories.
Tokens Are Becoming The Revenue Meter
For large language models, a token is a small packet of text or data processed by the model. Counting token expenditure is not perfect, but it is closer to demand than another round of vague adoption surveys. It captures both how much customers use models and how much providers can charge for that use.
A token expenditure gauge recently fell from record highs. That matters because the underlying price of a single token has already dropped by more than 90% since 2023. A lower gauge can mean users are choosing cheaper models, providers are cutting prices, or demand is less willing to absorb premium pricing.
This is not the same as saying AI demand has vanished. Only weeks ago, the same gauge was roughly double its level from late last year. The distribution is messy, which is normal when a new compute market moves from novelty to procurement.
A Boston Consulting Group paper looked at token consumption at more than 100 public technology companies with more than $500 million in trailing twelve month revenue. Its useful point was simple. Tokens matter most when companies redesign workflows around AI, not when they merely buy more compute and hope the spreadsheet smiles.
Samsung Shows The Chip Side Still Works
Samsung Electronics gave the chip bulls a hard number. Latest profits rose about 19 times, and the company delivered a third straight quarter of record profit. High memory chip prices helped drive April through June earnings.
Yet Samsung shares still fell. That is the interesting part. Investors can see current profits, but they are also asking whether the next wave of AI investment earns enough return after the capex bill is paid.
SK Hynix is another signal. A US listing for the South Korean memory maker has drawn backing from Situational Awareness, a fund run by a former OpenAI researcher, and Baillie Gifford. Capital still wants exposure to memory. It just wants a cleaner map from chip demand to cash return.
Hyperscalers Are Starting To Look Industrial
Microsoft, Amazon, and Meta are still central to the AI trade because their core businesses throw off cash. The problem is that AI turns software economics into a heavier industrial equation. Chips, data centers, power, and cooling all demand cash before the productivity dividend appears.
Bank of America analysis in market discussions shows a blunt pattern. Much of the free cash flow from the largest hyperscalers is now being absorbed by capital expenditure. In plain English, the cash machine is feeding the chip machine.
That changes valuation logic. Software investors used to pay for high margins and low capital needs. The new model asks them to pay for large future productivity gains while current free cash flow is redirected into physical assets. There is nothing immoral here. It is just less magical.
Infrastructure Winners Face The Same Test
Sterling Infrastructure is a good example below the chip layer. The company reported record first quarter revenue of $825.7 million, up 92% year over year. Adjusted EPS rose 120% year over year, and the stock rallied more than 50% after earnings.
Guidance also moved higher. Management pointed to fiscal 2026 sales of $3.70 billion to $3.80 billion and adjusted EPS of $18.40 to $19.05. Data center construction and backlog visibility support the growth story, but a forward price to earnings ratio near 37 times assumes a lot of execution. A 15% to 20% correction or further earnings revisions would make the math less stretched. Even the picks and shovels need a margin of safety.
The AI supply chain is not only GPUs and cloud instances. Honeywell’s Solstice unit agreed to buy Element for $14.5 billion. The transaction is meant to create an advanced materials company with a combined enterprise value of about $29 billion.
Microsoft’s planned reduction of 4,800 jobs in its Xbox unit adds a colder signal from inside big tech. Capital is being redirected. Weak margins and a hardware downturn in gaming matter less to the AI race than server capacity and cloud returns. Large platforms can fund many experiments, but they still prune.
What To Watch
The first variable is token pricing. If token costs keep falling while usage grows, productivity can still compound, but revenue expectations for model providers may need trimming. Cheaper intelligence is good for users and awkward for some valuations.
The second variable is hyperscaler capex discipline. Investors should watch whether Microsoft, Amazon, and Meta can turn compute spending into durable revenue, not just impressive infrastructure photos.
The third variable is backlog quality. Samsung, SK Hynix, Sterling Infrastructure, Solstice, and Element all point to real demand. The question is not whether AI spending exists. The question is who keeps enough cash after the cycle turns from vertical ascent to cost control.