Nvidia’s dominance in artificial intelligence chips is facing growing competition as the industry shifts towards a greater focus on inference—the process of AI responding to user queries—rather than just training large models. This transition, accelerated by companies like Chinese start-up DeepSeek, has opened opportunities for challengers to capture a significant share of the AI computing market.
While Nvidia has long held control over AI model training, inference is becoming the primary driver of computing demand. Start-ups like Cerebras and Groq, alongside tech giants such as Google, Amazon, Microsoft, and Meta, are investing heavily in inference-optimised chips. These chips aim to provide faster and more efficient AI responses, a crucial factor as businesses and consumers demand more sophisticated AI applications beyond chatbots like ChatGPT.
According to analysts at Barclays, capital expenditure on inference computing in “frontier AI” will exceed training investments in the next two years, growing from $122.6 billion in 2025 to $208.2 billion in 2026. Barclays also estimates that while Nvidia will maintain near-total market share in AI training, its share of inference computing could drop to 50% by 2028, leaving nearly $200 billion in chip spending for competitors to capture.
Financial Times reports that Nvidia CEO Jensen Huang remains confident, asserting that the company’s latest Blackwell chips are designed to handle inference efficiently. He also points to the widespread adoption of Nvidia’s CUDA software, which creates a significant competitive advantage. However, cloud computing providers and AI developers are increasingly seeking alternatives to Nvidia’s GPUs, favouring specialised inference chips that optimise memory access and response speed.
