Nvidia, China and chip race
Nvidia unveiled the Rubin R100 GPU aimed at agentic AI workloads, emphasising bandwidth-rich memory, while Chinese firms say a 2nm-class AI GPU has entered prototype stages despite unclear production prospects. Analysts and trade pieces also note China has subsidised chipmaking heavily over the last decade, suggesting the competitive landscape is broadening beyond process-node headlines. These developments point to continued premium on memory bandwidth and packaging choices in AI silicon. (markets.financialcontent.com, )
Nvidia is pushing its next artificial intelligence system around memory speed and rack design, as a Chinese start-up says its own advanced artificial intelligence graphics chip has reached prototype testing. (nvidia.com) (scmp.com) Nvidia said on March 16 that its Vera Rubin platform is now in full production with seven chips, including the Rubin graphics processing unit, Vera central processing unit, NVLink 6 switch and ConnectX-9 network card. The company said the system is built for “agentic” artificial intelligence, meaning software that handles multi-step tasks and long chains of prompts with less human intervention. (nvidia.com) Nvidia’s own technical blog said Rubin uses High Bandwidth Memory 4, or High Bandwidth Memory 4, a stacked memory design placed close to the processor to move data faster than standard server memory. Nvidia said the new memory interface is twice as wide as High Bandwidth Memory 3e and that Rubin nearly triples memory bandwidth versus Blackwell. (developer.nvidia.com) That focus reflects how large language models are running into a traffic problem as much as a math problem. Nvidia said reasoning models and long-context systems need to process hundreds of thousands of input tokens while keeping latency, power use and deployment costs under control. (developer.nvidia.com) Nvidia is also selling Rubin as a full rack instead of a single chip. The company said Vera Rubin NVL72 combines graphics processors, central processors, networking, storage and cooling into one pod-scale system, treating the data center rather than one server as the unit of compute. (nvidia.com) (developer.nvidia.com) In China, Shanghai-based Dishan Technology is in “prototype verification” for its first 2-nanometer-class artificial intelligence graphics processing unit, according to a report published April 14 by the South China Morning Post citing local media. The report said the chip was first unveiled in July 2025 and has not yet reached tapeout, the final design handoff before manufacturing. (scmp.com) Dishan said the design uses a hybrid Fin Field-Effect Transistor and Gate-All-Around approach with chiplets, or smaller chip blocks linked together in one package. The company said that design should improve energy efficiency by 40 percent over its prior chip, though commercial production could still be one to two years away and a suitable foundry remains unclear. (scmp.com) (digitimes.com) China’s position in this race is not defined only by whether it can mass-produce a 2-nanometer chip. Tom’s Hardware reported this month that China spent about $142 billion on semiconductor industrial policy from 2014 through 2023, compared with about $39 billion in the United States over the same period. (msn.com) Those spending figures sit alongside export-control limits that still constrain access to the most advanced chipmaking tools. The South China Morning Post said those controls cap China’s practical ability to manufacture beyond 5-nanometer nodes, even as local designers keep moving designs forward. (scmp.com) The immediate contest is no longer just who names the smallest process node first. Nvidia is betting that bandwidth-rich memory, chip packaging and tightly integrated racks will decide how much artificial intelligence work a data center can actually deliver. (nvidia.com) (developer.nvidia.com)
