Nvidia Bets on Open-Source AI Models as Huawei Chip Launch Intensifies China Rivalry

Nvidia is repositioning itself as the dominant force in open-source artificial intelligence, launching its Nemotron 3 Ultra model as a potential successor to Meta's Llama and Google's Gemini in the race to power agentic AI systems. The strategic pivot, articulated by CEO Jensen Huang, reflects a belief that smaller, open models will drive the next wave of AI agents—autonomous systems that perform tasks without human intervention.

The company claims Nemotron 3 Ultra delivers five times greater efficiency and achieves the highest accuracy on its GB200 NVL72 hardware platform. Nvidia has assembled a coalition of global AI labs to advance what it calls "frontier open models," expanding its model families to support agentic, physical, and healthcare AI applications. The move places Nvidia in direct competition with the very open-source community it seeks to lead, a paradox that underscores the blurring lines between collaboration and rivalry in AI infrastructure.

The timing coincides with mounting pressure in China, where Huawei is preparing to ship approximately 750,000 units of its 950PR AI chip this year. Samples were distributed to customers in January, with mass production slated to begin next month and full-scale shipments expected in the second half of the year. ByteDance and Alibaba are among the Chinese tech giants planning orders, according to sources familiar with the matter who were not authorized to speak publicly.

Huawei's 950PR represents a strategic evolution beyond raw computing power. The chip is optimized for inference workloads—running trained AI models to execute tasks—a segment experiencing surging demand in China as the tech sector shifts from model development to real-world deployment. The chip also marks Huawei's departure from its proprietary CANN software system, now offering compatibility with Nvidia's software ecosystem to ease migration for Chinese developers.

(Washington has banned many Nvidia AI chips from sale in China over concerns the technology could enhance military capabilities, creating an opening Huawei is moving aggressively to fill.)

The telecom industry's embrace of agentic AI has accelerated sharply. At Mobile World Congress 2026, industry analysts noted the conversation had shifted from skepticism—one executive previously described AI agents as "drunk interns"—to practical deployment discussions. Nvidia announced partnerships with T-Mobile and global robotics leaders to integrate physical AI applications on AI-RAN-ready infrastructure, while launching an open physical AI data factory blueprint for robotics, vision AI agents, and autonomous vehicles.

Meta's Llama, long the preferred base model for customized AI development, has lost ground to Google's Gemini in recent months. Nvidia's Nemotron 3 Ultra enters a crowded field where model performance, licensing terms, and ecosystem compatibility determine winners. The company's dual role as hardware provider and model developer creates potential friction with customers who also compete in the model space.

The open-versus-closed model debate has become a proxy war for control of AI's economic future. Closed models from OpenAI and Anthropic offer tighter integration and perceived safety controls, while open models promise customization and cost advantages. Huang's bet on open-source reflects a calculation that fragmentation favors Nvidia's hardware business, as diverse models require the computational substrate only advanced GPUs can provide.

Huawei's challenge extends beyond chips to software compatibility, the moat that has protected Nvidia's dominance. By supporting Nvidia's software stack, the 950PR lowers switching costs for Chinese developers, a tactical concession that could prove strategically decisive. Alibaba, China's most aggressive AI spender with over $53 billion pledged across several years, is targeting $100 billion in AI revenue within five years, creating a massive domestic market beyond U.S. export controls.

Nvidia's expansion into automotive AI—with BYD, Geely, Isuzu, Nissan, and Xiaomi adopting its DRIVE Hyperion platform for Level 4 autonomous vehicles—illustrates its push into physical AI. The company is also partnering with Roche to scale AI factories globally for drug discovery and manufacturing, and collaborating with industrial software giants to integrate AI into design and engineering workflows. These vertical integrations aim to lock in Nvidia's architecture across industries before alternatives mature.

The emergence of AI wearables and edge devices adds another dimension. Startups like Button Computer, founded by former Apple employees, are building tiny computers optimized for AI voice commands, while Y Combinator's latest cohort includes AI-powered tools for game development, security camera intelligence, and architectural workflows. These applications favor inference over training, the workload where Huawei's 950PR is designed to excel and where Nvidia's GPU advantage is less pronounced.

The strategic question is whether Nvidia can maintain its hardware monopoly while simultaneously leading open-source model development—a dual role that requires collaborating with competitors and customers alike. Huawei's ability to combine chip performance with software compatibility, backed by China's vast domestic market and state support, presents the most serious challenge to Nvidia's dominance since the AI boom began. The outcome will determine not just market share, but whether AI infrastructure remains concentrated in U.S. hands or fragments along geopolitical lines.