The rapid growth of artificial intelligence has created a huge demand for computing power.
While much of the investment in recent years has focused on training AI models, the next major opportunity lies in running them in real time. NVIDIA believes this shift could open a massive revenue opportunity for AI chips.
At its annual developer conference in San Jose, the company said the total market for AI infrastructure could reach at least $1 trillion by 2027, according to Reuters.
The new estimate is significantly higher than the $500 billion opportunity NVIDIA projected earlier through 2026, reflecting the rising demand for AI systems that interact directly with users.
NVIDIA targets the growing inference market
NVIDIA CEO Jensen Huang used the company’s GTC conference to outline a strategy focused on AI inference computing. Inference is the stage at which trained AI models respond to user questions or perform tasks.
During the event, NVIDIA introduced a new central processor and an AI system built using technology from Groq, a chip startup whose technology NVIDIA licensed in a deal worth about $17 billion in December.
Competition intensifies in AI chips
NVIDIA’s graphics processing units (GPUs) have dominated the training of large AI models, powering systems developed by companies such as OpenAI, Anthropic, and Meta.
However, inference workloads are attracting more competitors. Companies like Google are building custom AI processors, while traditional CPUs, dominated by Intel, are also being used to run AI models.
To address this shift, NVIDIA is expanding its own hardware lineup.
Huang introduced a new “Vera” CPU, saying the company is already seeing strong demand for CPUs sold on their own. “We are selling a lot of CPU standalone,” Huang said. “This is already going to be a multi-billion-dollar business for us.”
How NVIDIA plans to run AI faster
NVIDIA also explained how inference workloads will be divided between different chips.
The company’s Vera Rubin chips will handle the first stage, called prefill, in which a user’s question is converted into the “tokens” that AI systems understand.
Groq’s processors will handle the decode stage, which generates the final answer delivered to the user. This approach is designed to speed up responses as millions of people begin using AI systems daily.
NVIDIA also previewed its future chip roadmap, including the Feynman architecture, which is expected to arrive around 2028 after the company’s Rubin Ultra chips.
Beyond chips, NVIDIA also revealed a new platform aimed at autonomous AI agents. The system, called NemoClaw, integrates with the OpenClaw platform and adds privacy and safety controls to software that can perform tasks automatically with minimal human input.
Industry analysts say NVIDIA is increasingly presenting entire AI systems, rather than just individual chips.
NVIDIA shares rose briefly after the announcement before closing the day about 1.2% higher, reflecting cautious optimism among investors.
