Why China's Top AI Models Still Rely on Nvidia Chips

China's Self-Sufficiency Push Runs Into Nvidia's Grip

Despite years of state-backed investment in homegrown semiconductors, China's leading AI labs continue to train their flagship models on Nvidia hardware. According to reporting on the issue, the practical costs of migrating to domestic chips remain too steep for most Chinese AI developers, who need proven performance and mature software ecosystems to stay competitive 1. That reliance underscores a broader truth about the current AI boom: even amid geopolitical pressure to decouple from American technology, Nvidia's silicon remains the default foundation for serious model training, both in the US and abroad 1.

The Ballooning Price of the Global AI Buildout

Nvidia's centrality is reinforced by just how expensive the broader AI infrastructure race has become. Major US tech companies are burning through cash at an extraordinary rate to keep pace. Amazon, Alphabet and Tesla all posted negative cash flow in their latest quarters, while Meta's cash generation collapsed by 91%, a sign that the capital intensity of building out AI data centers is straining even the best-funded companies 2. Soaring memory-chip prices are compounding the problem, adding another cost pressure on top of GPU spending 2.

That cost pressure is spilling into consumer markets as well. Rising component costs tied to the AI hardware race have pushed gaming PC and console prices up sharply, with average selling prices for gaming gear climbing 16% in the first half of 2026 alone, from roughly $452 to a notably higher figure, as chipmakers prioritize AI-oriented production 4.

Inference, Not Just Training, Becomes the New Battleground

As training costs balloon, attention is shifting toward the expense of running AI models in production, known as inference. Nvidia used its GTC 2026 conference to unveil new chips and software explicitly aimed at making inference cheaper and more efficient, signaling an industry-wide pivot from raw training power toward sustainable, real-world deployment economics 7. Startups are chasing the same problem from different angles: Runware's new Sonic Inference Pods pack 1,200 GPUs into a 20-foot shipping container capable of 1 megawatt of compute, using closed-loop liquid cooling with no water consumption, and the company aims to reach 1 gigawatt of inference capacity by 2027 as traditional data center construction faces multi-year grid connection delays 5.

Enterprises are also trying to control inference costs through software rather than hardware. So-called AI model routers, which automatically direct tasks to the cheapest adequate model rather than defaulting to the most expensive one, have become one of the hottest categories in enterprise AI tooling as agentic AI workloads drive up bills 3.

Consumer AI Products Add to Demand

Even consumer hardware is being pulled into the inference economy. OpenAI is reportedly developing a doughnut-shaped smart speaker, priced as high as $400, that would run ChatGPT's voice mode and use moving parts to signal activity 6. Products like this add incremental but real demand to the inference infrastructure that companies like Nvidia and Runware are racing to build out, reinforcing how thoroughly AI economics now touch training, deployment and everyday consumer devices alike.