China’s AI Chip Challengers Take Aim at Nvidia’s Software Advantage

Enflame is raising approximately $908 million through its Shanghai IPO after first-quarter revenue increased 1,475% year on year, according to Reuters Breakingviews’ September 7 commentary. Its quarterly net loss nevertheless reached 444 million yuan. Alongside Moore Threads, MetaX and Biren, the company represents a growing domestic challenge to Nvidia. The combination of rapid sales growth and continuing losses raises an important distinction for investors: commercial expansion does not automatically mean a chipmaker has established a profitable business model.

Software helps explain why competing in this market requires more than designing a capable processor. Nvidia’s CUDA platform documentation describes an environment connecting its GPUs with programming languages, accelerated libraries and widely used AI frameworks. For customers, the economic value of such an ecosystem includes the development work and operational knowledge accumulated around it. Moving to another supplier can require code adjustments, testing and staff time. The analytical implication is that a lower purchase price for a competing chip may be insufficient unless the customer can also migrate applications economically and maintain reliable performance.

Enflame’s own vLLM-GCU documentation provides a concrete example of its approach. The company has adapted the vLLM inference framework for its S60 accelerator, with support for model families including Qwen, DeepSeek and Llama. It also supplies benchmarking tools and optimizations for its hardware. This matters because inference—the process of using a trained model to generate an answer—is where an AI application incurs recurring computing costs as usage grows. Supporting familiar models and development frameworks could make domestic hardware easier to evaluate. However, a published compatibility list is evidence of software development, rather than independent proof of superior performance or lower operating costs.

Nvidia is also continuing to develop this part of its business. Its inference software documentation describes tools for coordinating computing resources, managing memory and serving models across multiple GPUs. This points to a competition increasingly decided at the system level. A meaningful customer comparison would consider response speed, workload capacity, power consumption, engineering support and reliability together. A processor that performs well in one benchmark may deliver different economics when handling a changing mix of real customer requests.

For investors assessing China’s AI semiconductor sector, the strongest evidence of progress would therefore extend beyond fundraising and initial shipments. Repeat purchases, adoption across several customers, improving gross margins and reduced implementation costs would indicate that the technology is becoming a durable commercial offering. The opportunity is substantial, but the path requires sustained spending on both hardware and software. China’s challengers can make progress by serving particular workloads effectively; establishing a broadly competitive computing platform requires that success to be repeatable.