The artificial intelligence investment boom is entering a more demanding phase as investors increasingly scrutinise whether record spending on data centres, chips and cloud infrastructure can translate into sustainable revenue growth.
A Binance Research report by Lim Kim Thye indicates that markets are rewarding companies that demonstrate strong AI-driven sales while penalising businesses whose rising capital expenditure is not accompanied by clear commercial returns.
Combined 2026 capital expenditure guidance from the largest US cloud operators now stands at between $725 billion and $800 billion, depending on whether finance leases and prepayments are included.
The report estimates that capital expenditure among five major operators has risen from 41 per cent of operating cash flow in 2023 to approximately 105 per cent in 2026. This means the companies are collectively spending more on infrastructure than their operations generate.
Aggregate free cash flow reportedly declined from a positive $246 billion in 2024 to an estimated negative $37 billion in 2026. Debt has increasingly filled the financing gap, accounting for 32 per cent of hyperscaler capital spending in the 12 months to mid-2026, up from nine per cent in the 2024 financial year.
AI demand remains strong
Despite concerns about spending, demand for artificial intelligence services continues to rise.
Google Cloud revenue increased by 82 per cent to $24.8 billion, while its operating margin climbed to 35.6 per cent from 20.7 per cent. The business also reported a backlog of $514 billion.
Microsoft’s commercial remaining performance obligations reached $678 billion, representing an 84 per cent increase. Azure grew by 43 per cent, while Microsoft 365 Copilot surpassed 30 million paid users.
Oracle reported remaining performance obligations of $638 billion, although only about 12 per cent is expected to convert into revenue within 12 months. Approximately 20 per cent is scheduled beyond five years.
The figures point to substantial demand, but they also highlight the time it may take for signed contracts to generate revenue. Cloud providers must convert that backlog quickly enough to justify infrastructure whose depreciation schedules typically run for four to six years.
Investors reward conversion, not spending alone
Recent stock-market reactions underline the change in investor sentiment.
Alphabet shares fell 7.13 per cent after the company increased its capital-spending guidance, despite reporting revenue growth of 24 per cent and an 82 per cent expansion at Google Cloud.
Meta declined 7.95 per cent after an earnings miss, another increase in capital expenditure and a sharp reduction in quarterly free cash flow.
Microsoft, by comparison, gained 15.51 per cent after reporting strong Azure growth and an 84 per cent rise in contracted future revenue while slightly lowering its 2026 capital-expenditure outlook.
Amazon rose 15.32 per cent and crossed a $3 trillion market valuation after AWS growth accelerated to 37 per cent. The gain came even after the company lifted its capital-spending target to $220 billion.
The contrasting reactions suggest that investors are not rejecting AI spending outright. Instead, they want evidence that the expenditure is producing measurable revenue.
Cheaper models reshape the market
Rapidly falling inference costs are also changing which companies benefit most from the AI boom.
Weekly token traffic routed through OpenRouter reportedly reached 137 trillion in early September 2026, about 20 times its level at the beginning of the year. Open-weight models now account for more than half of production inference tokens on the platform.
The five most-used models were all open weight, offering an estimated 90 per cent capability parity at roughly one-sixth of the cost per call.
This shift could favour large cloud-infrastructure providers, whose margins range between 33 per cent and 38 per cent. Model developers, however, may face greater difficulty capturing revenue as businesses increasingly route different tasks to cheaper models.
Investors diversify beyond AI chips
Market concentration is also beginning to ease.
The Magnificent Seven’s share of the US market declined from approximately 35.3 per cent in October 2025 to 33.2 per cent in September 2026.
Through August, the S&P 500 had gained 12.28 per cent, compared with 13.46 per cent for the Nasdaq Composite and 19.12 per cent for the small-cap Russell 2000.
The Russell 2000’s stronger performance suggests that the AI investment theme is expanding into sectors such as power, industrials, software and capital markets rather than remaining concentrated in large semiconductor and technology companies.
Binance investor activity shows a similar trend. Technology-hardware allocations declined from 15.92 per cent at the end of June to 7.94 per cent in early September, while investors added exposure to software, capital markets and retail businesses.
Semiconductors nevertheless remained the largest allocation at 42.08 per cent, indicating that Binance investors are still considerably more exposed to the sector than the broader market.
Risks remain concentrated
Another concern is the cloud industry’s dependence on a small number of private AI companies.
Analyst estimates cited in the report suggest that OpenAI and Anthropic account for at least 70 per cent of Microsoft’s AI revenue and 73 per cent of Amazon’s. They are also estimated to represent 28 per cent of Google Cloud revenue in 2026.
Oracle’s reported $300 billion OpenAI contract accounts for nearly half of its $638 billion backlog.
A slowdown at either AI developer could therefore affect cloud revenue, infrastructure contracts and the financing arrangements supporting new capacity.
The next phase of the AI boom is likely to be defined less by how much companies spend and more by how effectively they convert that investment into revenue, cash flow and long-term returns.
