The new restrictions are narrower than the headlines suggest, but they hit where it matters most for AI hardware. Nvidia loses its workaround chip for the Chinese market. ASML can't ship the most advanced chip-making machines (lithography systems). Samsung and SK Hynix have to choose between US technology access and their largest customer. The companies most exposed already adjusted their forecasts — but their suppliers, customers, and Chinese chipmakers are only starting to.
AI-related software prices, subscriptions, and cloud computing costs that don't fall as expected. The hardware backbone that powers ChatGPT-style services, AI image tools, and AI deployments at large companies costs more to deploy, and that's being passed through to subscription tiers and enterprise contracts you see at renewal. Companies running their own AI computing systems are reassessing their roadmaps — some projects get delayed, some get scaled back, some get redirected to alternative hardware that takes longer to deploy, with capital-spending plans repricing as the underlying chip cost shifts. If your job touches AI tools (development, marketing, design, sales operations), the tooling roadmap your company laid out is shifting, often quietly. Chinese-market AI services from Nvidia-dependent companies are visibly degrading or being repriced upward as the H20 substitute disappears.
Two parallel chip supply chains are forming — US-aligned and China-aligned — and the commitments being made now (where chip factories direct their production, which design partnerships get signed, which customers get prioritised) won't reverse if restrictions ease. TSMC and Samsung have already begun reallocating capacity away from China-bound advanced chips; that capacity is being committed elsewhere through 2027, which shows up in their forward earnings and in capital-allocation decisions for everyone running on their chip factories. ASML chip-making machines not shipping today mean specific Chinese chip factories lose the ability to make the most advanced chips for the design generation they're in — a 3-5 year window of degraded competitiveness that flows through Chinese tech-stock valuations and to anyone holding exposure to Chinese AI hardware. The computers used to train and run AI for 2026-2027 are being designed now around how long it takes to switch to alternative hardware (3-12 months to alternative platforms, longer for systems that need fine-tuning), with the cost and capability of available compute resetting at a different baseline. For anyone running a software or AI-dependent business, what you pay for compute power and what you can do with it sits at a different floor than it would have without these restrictions — and US AI-infrastructure investment cycles reprice around the new substrate.