Stocks wobble as bonds slump to monthly loss

The 10-year U.S. Treasury yield rose nearly 50bps in September to a 19-year high above 5.27%, while the 2-year yield climbed more than 57bps toward 5% as markets priced in three additional Fed hikes by mid-2027. The bond selloff and Brent crude above $106.60 per barrel pressured Asian equities, while the rate-sensitive Nasdaq fell 0.9% despite Nvidia's $150 billion buyback boost. Chinese technology sentiment remained fragile following proposed U.S. restrictions on Chinese components in data centers, leaving the CSI300 at a one-year low.

Analysis

The key transmission is not a one-day AI sentiment reset but a higher discount-rate regime colliding with an increasingly capital-intensive AI buildout. NVDA can support near-term EPS through repurchases, but that does not solve the market’s emerging question: whether hyperscaler and frontier-model customers can earn adequate returns on infrastructure spending before depreciation, power and financing costs rise. If the reported training interruption proves material or spreads to other model developers, the risk is multiple compression across AI infrastructure well ahead of any reduction in chip orders.

Within semis, the highest-risk exposures are suppliers priced for uninterrupted accelerator-cluster expansion, including NVDA, AVGO and VRT; the second-order vulnerability is utilities and data-center developers whose contracted demand assumptions depend on sustained training intensity. Conversely, oil producers such as XOM and CVX retain a near-term earnings hedge against elevated crude, while rate-sensitive long-duration software and unprofitable AI application names lack such offset. The combination of higher real rates and energy costs also raises the hurdle rate for enterprise AI adoption, potentially shifting spend from speculative model training toward inference, optimization and cost-saving software over the next 6-18 months.

Consensus may be too quick to treat any training halt as an immediate demand collapse for NVDA: installed clusters, backlog conversion and sovereign AI spending can cushion revenue for several quarters. The more actionable issue is whether customer capex guidance, power procurement or depreciation assumptions change during the next earnings cycle. Falsification for the bearish AI-infrastructure view would be unchanged hyperscaler capex plans, sustained NVDA backlog commentary and a meaningful retreat in the 10-year Treasury yield; a break higher in long yields alongside capex deferrals would make the de-rating nonlinear.

Key Decisions for Investors

• Use any relief rally in NVDA to initiate a 1-3 month relative-value hedge: long XOM or XLE versus short NVDA, sized beta-neutral. The trade captures oil-supported cash-flow durability against AI duration risk; exit the short if NVDA customer capex commentary remains intact and the 10-year yield falls materially from current stressed levels.
• Buy 3-6 month put spreads on SMH rather than outright NVDA puts if implied volatility remains contained; this broadens exposure to AI-capex revision risk across NVDA, AVGO, AMD and networking/power suppliers while limiting premium paid. Target a 10-15% sector drawdown; invalidate on reaffirmed hyperscaler capex guidance.
• Avoid adding to VRT and other data-center infrastructure momentum exposure until there is independently verifiable evidence that training demand has resumed and customer project timelines are unchanged. Watch backlog cancellations, power-interconnect delays and hyperscaler depreciation guidance as the decision triggers.
• Maintain or add selective energy exposure through XOM/CVX on a 1-3 month horizon as a portfolio inflation hedge, but take profits if crude retreats below the recent geopolitical risk premium or a credible supply/diplomatic catalyst emerges.