Jensen Huang May Be Starting to Feel Anxious

The awkward middle

Jensen Huang may be having an uncomfortable realization: Nvidia (英伟达) sits between two unaligned power structures and neither rewards the GPU vendor as before. On a recent podcast Huang repeated a striking formulation — “you have to turn electrons into tokens” — and described Nvidia as “the architecture that produces the most tokens per watt,” not necessarily the most FLOPS per watt. A company that built its reputation selling GPUs is now trying to recast itself for a world where compute and token economies diverge.

Two different empires of power

On one side of the Pacific, compute is concentrated. It has been reported that EpochAI found Amazon, Google, Meta, Microsoft and Oracle together control roughly 67% of global AI compute (H100-equivalent). These hyperscalers act like landlords: they supply massive datacenter capacity and increasingly build their own silicon, while frontier model labs — OpenAI, Anthropic, xAI — rent that capacity. Nvidia has deployed capital into several of those labs, but Huang himself has warned those investments may be the last. Meanwhile, it has been reported that U.S. export restrictions (often referenced around “H20”-style measures) sharply curtailed Nvidia’s China-facing data‑center pool after 2025, shrinking a previously vital revenue channel.

On the other side, China’s cloud and AI stack is organized differently. Alibaba (阿里巴巴), ByteDance (字节跳动) through Volcano Engine, Tencent (腾讯) and Baidu (百度) are simultaneously owners of compute, developers of large models, and primary consumers of the tokens (词元) those models generate. It has been reported that China’s daily token calls have surged into the hundreds of trillions and that OpenRouter data showed Chinese weekly token volumes overtaking the U.S. The practical result: in China token has become the tradable commodity; compute is the raw input. The Chinese state has even begun to treat tokens as an economic metric, elevating “词元” to an official-like unit in public discourse.

Caught in a squeeze — geopolitics matters

Huang’s “five-layer cake” framework — energy, chips, systems, models, applications — underlines the geopolitical stakes: if the U.S. does not lead in every layer, Huang warned, it risks losing strategic advantage. But the commercial reality is messier. In the U.S., Nvidia’s biggest customers are also building their own chips and services; in China, models and token production are increasingly running on non‑Nvidia silicon such as Huawei (华为) Ascend, after large cloud orders for domestic chips were reportedly placed. So who is left to pay premium prices for Nvidia’s stack? Who even needs to? The answer shapes not just corporate fortunes but the contours of tech rivalry between Washington and Beijing.

Where does Nvidia go from here? It can double down on software and system-level differentiation, chase new upstream partnerships, or accept a diminished role as hyperscalers and domestic Chinese stacks tighten their own supply chains. The question is blunt: can a company that sells the world’s best “shovels” remain indispensable when its biggest customers are quietly digging with their own?