Nvidia Stock Falls 2.8% as Higher Rates Test AI Hardware Valuations
Nvidia and several semiconductor names fell more than 2.8% on Sept. 14 as SoftBank lost 10.7%, South Korea’s Kospi dropped 3.3% and Europe’s technology index fell about 2.1%. The Nasdaq 100 was down only around 0.4% in the latest session reading, leaving a useful cross-section: hardware and other AI-capex beneficiaries took much more damage than the broader growth complex.
Two repricings landed together. Senior AI executives intensified calls for more deliberate frontier-model scaling and stronger safety coordination, weakening confidence in a straight-line extrapolation of training demand. At the same time, the U.S. 10-year yield crossed 5%, increasing the required return on equities whose valuations still assume years of exceptional growth.
The first mechanism changes the expected quantity and timing of future compute demand; the second changes the price investors put on those cash flows. A stock can fall sharply even when current orders remain intact if the market trims the growth rate and raises the discount rate at the same time.
Current hyperscaler disclosures still point to infrastructure expansion. Microsoft said its calendar-2026 capex expectation is about $175 billion after a lease-accounting change and that capacity should remain constrained at least through 2026. Alphabet spent $80.6 billion on capex in the first half of 2026, more than double the year-earlier period, while Google Cloud revenue rose 82% in Q2. Meta expects $130 billion to $145 billion of 2026 capex. Amazon spent $96.3 billion of cash capex in the first half, primarily reflecting technology infrastructure — the majority supporting AWS growth — while AWS sales rose 37% in Q2.
Those figures are not cleanly additive: the companies use different lease and capex definitions, and not every dollar buys Nvidia hardware. They are nevertheless much stronger evidence against an industry-wide cancellation cycle than a single data-centre announcement. The operating question is how much of that spending flows into accelerated compute, at what pace and with what utilization.
Safety-driven restraint could still change the mix. Slower growth in frontier training runs could shift dollars toward inference, networking, storage, applications or existing-cluster utilization. That would matter to Nvidia because its revenue sensitivity depends on accelerator content and cluster expansion, not on an undifferentiated “AI capex” total.
Sept. 14 therefore exposed a valuation asymmetry rather than an operating break. The market is charging AI hardware a higher discount rate while questioning the slope of frontier scaling; the largest customers are still deploying and financing extraordinary amounts of infrastructure. A real capex turn would show up in hyperscaler guidance cuts, order cancellations, lower accelerator deployment or weakening utilization. Monday’s tape establishes none of those operating signals.