SAN FRANCISCO, July 9, 2026, 01:59 (PDT)
China is getting ready to allow a few leading AI firms to buy a restricted batch of Nvidia H200 chips, possibly giving the U.S. chip supplier a small but key route back into a market pressed by U.S. export restrictions and China’s drive to use more domestic suppliers.
Nvidia’s H200 GPU is designed for heavy AI workloads. Chinese regulators told Alibaba, ByteDance, and DeepSeek they might get the green light to buy some H200 chips, Reuters said, citing The Information. Beijing could let in fewer than 200,000 chips, which is less than half of what these firms wanted earlier this year.
This is happening now because China’s AI firms want more computing muscle and Nvidia is still working to hold onto its position. The U.S. has let Nvidia sell H200 chips in China and cleared about 10 Chinese buyers, but Chinese authorities hadn’t signed off. Back in October, Nvidia CEO Jensen Huang said the company’s share in China had basically dropped to zero.
Nvidia is also stepping into CPUs now, moving beyond GPUs. CPUs handle key jobs in AI like code and data flow. Perplexity said Tuesday it plans to use Nvidia’s Vera CPU. Nvidia wants to take on Intel and AMD, who have led in this area for years.
Nate Kupp, Perplexity’s VP for computer enterprise and infrastructure, said Vera was a “dead-on fit” for core workloads. He said Vera ran AI agent coding jobs about 1.5 times faster than regular CPUs. AI agents are software systems that take a goal and can use tools to complete a chain of tasks with less human prompting.
Nvidia said it’s looking for $20 billion in revenue from Vera by the close of this fiscal year. The company’s trying to keep the story simple—if agents don’t stop, and CPUs move data to GPUs quicker, data centers can run better.
Nvidia announced a software update Wednesday. The company said its Nemotron 3 Ultra model, tuned with LangChain’s Deep Agents, hit the top accuracy on LangChain’s benchmark for open models. It also ran tasks at 10 times lower inference cost per run than top closed models. Inference is when a trained AI model generates an answer.
LangChain and Nvidia rolled out NemoClaw for LangChain Deep Agents—a blueprint that brings together Nvidia’s Nemotron model, LangChain’s agent tech and Nvidia OpenShell, a secure agent runtime. LangChain CEO Harrison Chase said building stronger agents is about improving “the system around the model.” Nvidia’s Huang added, “Super agents have arrived.” PR Newswire
U.S. regular trading hadn’t started as of the time filed. Nvidia last changed hands at $204.12, up $7.29 from its previous close, putting the company’s market cap at roughly $4.98 trillion, according to market data.
Wall Street is still figuring out what matters most. Bank of America analyst Vivek Arya called Nvidia an “enhanced” buy, according to MarketWatch, as investors keep an eye on rising high-bandwidth memory costs for AI and pressure from custom AI chips like Google’s tensor processing units. MarketWatch
The China opening still isn’t settled. Reuters reported Nvidia, the U.S. Commerce Department, China’s Commerce Ministry, Alibaba, ByteDance and DeepSeek haven’t responded to comment requests, and Beijing hasn’t finalized chip volumes. Export rules could tighten again, China might back local suppliers, or buyers could move spending to internal chips.
Nvidia faces two tests from this update. First, there’s the political part—will H200 orders from China show up? Second, there’s the tech side. Vera and the Nemotron stack need to push agentic AI past just being a data-center pitch and convert it into real sales that can compete with Intel, AMD, and cloud players’ own silicon.
