A wave of alarm swept through technology markets when Chinese artificial intelligence laboratory DeepSeek unveiled a large language model that appeared to rival leading American systems at a fraction of the development cost. Nvidia shares fell sharply in the immediate aftermath, briefly shedding nearly 17 percent of their value in a single session — one of the largest single-day market capitalisation losses ever recorded for an American company. Yet a growing chorus of Wall Street analysts is now pushing back against the panic, arguing that the structural case for U.S. AI infrastructure investment remains firmly intact. The debate strikes at the heart of a broader question about data center expansion and whether demand for advanced semiconductors could genuinely slow.
According to a Street Journal analysis, the consensus emerging among institutional investors is that DeepSeek’s efficiency gains, rather than undercutting demand for high-performance chips, may paradoxically accelerate it. The reasoning follows what economists call Jevons’ paradox: when a resource becomes cheaper to use, total consumption tends to rise rather than fall, as lower costs unlock new applications and broader adoption.
Why Cheaper Models May Drive More Chip Demand, Not Less
The core of the bull case rests on the assumption that AI model training and inference are not fixed-cost endeavours. As models become more efficient and less expensive to run, enterprises that previously found AI deployment cost-prohibitive are expected to enter the market. Microsoft, Google, Amazon, and Meta have each signalled continued or increased capital expenditure plans for 2025, with aggregate spending across the four hyperscalers projected to exceed 200 billion dollars for the year. Analysts at several major brokerages contend that any marginal efficiency gain from models like DeepSeek simply expands the addressable market rather than cannibalising existing infrastructure spend.
Broadcom, which designs custom AI accelerator chips for several of the largest cloud operators, has seen its shares recover most of the ground lost during the initial sell-off. Analysts note that Broadcom’s business model is structurally tied to long-term silicon roadmaps negotiated directly with hyperscalers, making it relatively insulated from short-term model-level disruptions. Similarly, TSMC, the Taiwanese contract manufacturer that fabricates chips for Nvidia and others, reported that customer order books remained at or near capacity through the first quarter, suggesting no material pullback from major clients.
Geopolitical Risk Adds Complexity to the Investment Calculus
Not all market observers are fully convinced. A segment of the analyst community warns that DeepSeek represents something more significant than a single model release: evidence that Chinese AI development, operating under substantial U.S. export restrictions on advanced chips, has found ways to close the capability gap more rapidly than anticipated. If that trajectory continues, questions about the long-term pricing power of American semiconductor firms could become more pointed. Export controls imposed by the Biden administration, and largely maintained under the current White House, were designed in part to slow exactly this kind of advancement.
The geopolitical dimension adds a layer of uncertainty that pure demand modelling cannot fully capture. Investors are increasingly weighing not just the supply-and-demand dynamics of AI compute, but also the regulatory and diplomatic environments in which these companies operate. Despite those uncertainties, the prevailing institutional view, as reflected in relatively swift market recoveries for Nvidia, Broadcom, and related names, is that American AI infrastructure leaders retain durable competitive advantages rooted in ecosystem depth, software integration, and established enterprise relationships that a single open-source model release is unlikely to dismantle. The episode has, however, reinforced calls for continued U.S. investment in domestic semiconductor capacity and research — a policy conversation that shows no sign of quieting.