Binance Research says the AI trade is rotating away from semiconductors and toward software and capital markets as token economics, model selection and capital allocation all shift at the same time.
The report, written by Lim Kim Thye and translated by WuBlockchain, lays out a clear argument. As the cost of capability keeps falling, higher token usage benefits hyperscale cloud infrastructure providers with infrastructure margins of 33% to 38%. At the same time, capex by the five largest U.S. cloud operators has risen from 41% of operating cash flow in 2023 to about 105% in 2026. Markets, in Binance Research’s view, are no longer treating all AI spending the same way: capex that fails to turn into growth is being punished, while spending tied to visible AI revenue is getting rewarded.
The push to maximize token usage ended in Q2
According to the report, the second quarter of 2026 marked the end of the assumption that rising token consumption would automatically translate into revenue for model providers. Companies have realized that token usage and productivity do not move in a linear way, so they are no longer chasing token volume for its own sake. As AI agents spread quickly, attention has shifted to token efficiency per completed task.
Frontier model development has moved as well. The focus is no longer centered on reasoning benchmarks alone. It has shifted toward agent programming, memory architecture and cheaper lightweight inference.
As of early September 2026, weekly token volume routed through OpenRouter had reached 137 trillion, about 20 times the level at the start of the year. Open-weight models accounted for more than half of production inference tokens on the platform, up from one-third in the previous study. The top five models by token usage were all open-weight models.
Binance Research said those models offer about 90% of the capability of closed-source models while costing only about one-sixth as much per call. Stripe, while maintaining 50 million calls per day, cut its GPU cluster to one-third of its previous size and reduced inference cost by 73%.
The implication is that the gains from AI growth are now splitting across different parts of the stack. Enterprises are choosing models task by task, and an AI agent often needs thousands of tokens to finish one task. That makes low-cost open-weight models central to the economics of long-running agents. In an environment where the cost of capability keeps falling, rising token usage helps hyperscale cloud providers more than companies that directly sell model capability, because the latter group has a harder time turning usage growth into revenue.
Capex has moved past operating cash flow
The report says the five largest U.S. cloud operators have guided to combined 2026 capex of $725 billion to $800 billion, depending on whether finance leases and prepayments are included.
For Binance Research, the more important measure is the relationship between capex and operating cash flow. Across the five companies, capex has risen from 41% of operating cash flow in 2023 to about 105% in 2026. In other words, the cash going into infrastructure expansion has overtaken the cash generated by core operations.
That shift is already visible in free cash flow. Alphabet posted negative free cash flow of $5.9 billion for the quarter, the first negative reading since it went public. Meta’s free cash flow dropped from $8.5 billion a year earlier to $784 million. On a trailing 12-month basis, Amazon’s free cash flow was negative $7.6 billion, while Oracle burned $23.7 billion in cash in fiscal 2026.
Combined free cash flow for the five companies moved from positive $246 billion in 2024 to about negative $37 billion in 2026, the first negative turn in this investment cycle.
With internal cash no longer enough to fund expansion, external financing has become a bigger part of the picture. Debt financing as a share of hyperscaler capex rose from 9% in fiscal 2024 to 32% over the trailing 12 months through mid-2026. Total debt across the five companies stood at about $700 billion.
The report also highlighted an observation from the Bank for International Settlements, or BIS. Loan spreads in private AI credit are about 6.2 percentage points, close to the spreads paid by non-AI borrowers. Binance Research said that suggests the credit market has not fully priced the risks that equity investors have already started to recognize.
Demand signals are large, but revenue conversion takes time
The report points to a similar pattern across three companies.
- Google Cloud revenue rose 82% to $24.8 billion, operating margin improved from 20.7% to 35.6%, and remaining performance obligations reached $514 billion. Google Cloud is now processing 22 billion tokens per minute, up from 16 billion in the previous quarter.
- Microsoft’s commercial remaining performance obligations reached $678 billion, up 84% year over year. Azure revenue grew 43%, and paid seats for Microsoft 365 Copilot passed 30 million.
- Oracle’s remaining performance obligations reached $638 billion, up 363% year over year. But the company said only about 12% of that will convert into revenue in the next 12 months, while about 20% will not be recognized until more than five years later.
Among model companies, Anthropic’s annualized revenue was about $47 billion and OpenAI’s was about $25 billion. The report says both companies are still loss-making and are preparing for public listings.
Goldman Sachs expects token consumption to reach 24 times the current level by 2030, yet only 12% of knowledge workers are expected to use agent AI by then.
Those figures are large, but little of that demand converts into short-term revenue. Binance Research says that gap in timing is central to what happens next in markets. Orders are real and capex has already been spent, but hyperscalers and frontier model companies must turn contract demand into recognized revenue fast enough to justify the investment. The underlying assets supporting those businesses also tend to depreciate over four to six years.
The market has changed how it prices AI capex
Investors have already started to price in those factors, according to the report, and the latest earnings season made the shift clear. Index-level averages do not tell the whole story. Individual names do.
Among index constituents, companies that beat earnings expectations rose an average of 0.6%, below the five-year average of 1.0%. Companies that missed fell 2.5% on average, also less than the five-year average of 3.0%. The broad market reaction to both good and bad news has become less intense, but the split inside AI-related stocks has widened.
Looking at the first full trading day after earnings, Alphabet reported 24% revenue growth and 82% cloud growth, yet its stock fell 7.13% because it raised capex guidance. Meta fell 7.95% after missing on earnings per share, lifting capex again and reporting free cash flow of only $784 million.
The other side of the trade was just as visible. Microsoft posted 43% Azure growth and 84% growth in remaining performance obligations, then slightly lowered calendar 2026 capex. Its shares jumped 15.51%, the biggest one-day gain since 2008. Amazon raised capex to $220 billion, but its stock still climbed 15.32% because AWS growth accelerated again to 37%, lifting its market value above $3 trillion. Nvidia rose 8.74%. Palantir, which needs little infrastructure expansion capex, gained 29.45%.
The rule markets are using is now straightforward in the report’s telling: unless a company can show that spending is becoming growth, capex is a burden. That is why Alphabet was punished for raising capex, while Amazon was rewarded for doing the same.
In 2026, leadership is broadening rather than concentrating
Binance Research says market concentration has peaked and has started to ease. The weight of the “Magnificent Seven” in the S&P 500 fell from about 35.3% in October 2025 to 33.2% in September 2026. Through August, the S&P 500 was up 12.28% year to date, the Nasdaq Composite had gained 13.46%, and the Russell 2000 was ahead 19.12%.
Small caps stand out in that comparison. The report says that points to the AI trade spreading into industrials, power and other indirect beneficiaries rather than becoming even more concentrated in mega-cap software names. Equal-weight indexes outperformed market-cap-weighted indexes by about 3.6 percentage points, suggesting that after the easiest gains were captured, capital moved toward more volatile second-order beneficiaries.
There are still limits to how far that rotation can go. The S&P 500 trades at about 19.6 times forward earnings, slightly below the five-year average of 19.9 but above the 10-year average of 19.0. Rates are not offering much help either: the effective federal funds rate is 3.63% and the 10-year U.S. Treasury yield is 4.77%.
The report says that leaves the market in a very different place from last year. In 2026, gains are spreading across more sectors and across companies with different market capitalizations. Passive portfolios are reducing, not increasing, their excess tilt toward AI leaders.
Binance investors are diversifying too, starting from a more concentrated base
Comparing the end of June 2026 with Sept. 4, Binance Research found that changes in S&P 500 sector weights and Binance users’ equity holdings broadly moved in the same direction. Semiconductor weight in the S&P 500 fell from 18.8% to 14.8%, while technology hardware slipped from 6.8% to 6.2%. Binance investors made a bigger adjustment: their technology hardware allocation dropped from 15.92% to 7.94%.
Over the same period, nearly all other sectors in the S&P 500 saw small increases in weight, which the report interprets as a move away from one highly concentrated theme and into multiple sectors. Binance investors continued to cut aerospace and defense, interactive media and services, and industrial stocks, while adding exposure to capital markets, software and broadline retail.
The report highlights three key differences between Binance investors and the benchmark index. First, semiconductor exposure is far higher at 42.08%. Second, capital markets exposure reflects a crypto-linked preference. Third, portfolio concentration is much higher: the top 10 industries account for about 92% of Binance investors’ stock allocation, compared with about 53% for the top 10 industries in the S&P 500.
Monthly fund flows show the rotation in real time
Monthly net flows add another layer to the picture and show how positioning changed over time.
In July, investors were still leaning into the AI theme and remained optimistic ahead of earnings season. Many bought the late-July pullback. Semiconductors absorbed most of the net inflow, and capital markets also drew meaningful capital.
In August, the tone turned more cautious. Investors began taking profits in capital markets stocks while still adding to semiconductors, though at a much slower pace. At the same time, a large amount of capital moved into software, and total monthly net inflow fell to roughly half the July level.
In September so far, semiconductors have posted monthly net outflows for the first time. The report attributes that to higher long-term rates pressuring risk assets. It also notes that the current dataset covers only the first week of September and that the upcoming Federal Reserve rate decision could quickly reverse the pattern.
Across the three months, the portfolio started the quarter heavily concentrated in AI hardware and then gradually spread into software, capital markets and other second-order beneficiaries. Binance Research says that matches the sector rotation seen in the benchmark index, though among a different investor base and at a faster speed.
Trading volume closely matches investor holdings
The report says trading activity lines up closely with asset allocation. The industries that rank highest by trading volume are largely the same ones that dominate Binance investors’ portfolios. That suggests investors are more likely to be building and adjusting existing positions than rotating quickly between unrelated themes.
In September, semiconductors led with 33.84% of trading volume. Capital markets accounted for 16.71% and software for 13.67%. Together, those three groups made up about 64% of trading volume across the top 10 industries. Technology hardware ranked next at 10.42%, consistent with the broader reduction in exposure described earlier.
When trading volume is concentrated in sectors where investors are adding exposure, the report says, those flows are more likely to reflect deliberate allocation decisions rather than short-term turnover.
Pre-IPO perpetuals offer a direct way to price counterparty risk
The final section of the report focuses on Binance Pre-IPO perpetual contracts, which allow investors to take positions on private company valuations before similar products are offered elsewhere.
After Anthropic reported that quarterly revenue had more than doubled and that it had posted a small operating profit, its contract price rose about 35% in August. After OpenAI released its latest frontier model, Astra, its contract price climbed about 23% in early September, showing how quickly the market repriced the news.
Binance Research says those two model companies account for a large share of hyperscaler AI revenue. Wells Fargo estimates that more than 70% of Microsoft’s AI revenue comes from the pair. Barclays estimates they account for about 73% of Amazon’s AI revenue. UBS expects their share of Google Cloud revenue to rise from 28% in 2026 to more than 48% in 2027.
The report also notes that OpenAI’s $300 billion contract with Oracle amounts to about half of Oracle’s $638 billion in remaining performance obligations.
That means actual customer concentration is higher than the aggregate capex numbers suggest. A large share of related revenue is tied to two counterparties that are still private. If either company sees a meaningful slowdown in growth, the effect could flow from hyperscaler AI revenue into Oracle’s order backlog and then into securitized products backed by suppliers financing that infrastructure. For now, Binance Research says Pre-IPO perpetuals are among the few instruments that let investors hedge that specific exposure directly.