Nvidia has signalled a massive escalation in its market expectations, projecting that the revenue opportunity for its AI hardware could reach at least $1 trillion through 2027. This ambitious forecast, delivered by CEO Jensen Huang at the annual GTC developer conference in San Jose, highlights a strategic shift toward “inference computing”, the process of running AI models in real time to answer user queries.
Read: Bandwidth Blog & Smile 90.4FM Tech Tuesday: Apple MacBook Neo!
For years, Nvidia has dominated the “training” phase of AI, where massive models are built using thousands of GPUs. However, as companies like OpenAI, Meta, and Anthropic move from building models to serving hundreds of millions of active users, the industry focus has shifted. “The inference inflection has arrived,” Huang told an audience of over 18,000. “And demand just keeps on going up.”
To secure its leadership in this space, Huang unveiled a new CPU and a specialized AI system built on technology licensed from Groq. Nvidia recently acquired this tech for $17 billion to better compete against traditional CPUs and custom silicon developed by rivals like Google.
Huang outlined a two-step process for future AI interactions that utilizes different hardware for maximum efficiency:
- Prefill Stage: Nvidia’s upcoming Vera Rubin chips will handle the initial phase, translating human prompts into the “tokens” or digital language that AI models process.
- Decode Stage: The newly integrated Groq technology will manage the second stage, generating the final response with high speed and low latency.
Beyond graphics processors, Nvidia is aggressively expanding into the CPU market. Huang introduced the Vera CPU, asserting that standalone CPU sales are already poised to become a multibillion-dollar business for the company. This diversification is critical as competitors increasingly pitch CPUs as a more cost-effective alternative for deploying AI models at scale.
See also
Looking further into the future, Huang teased the Feynman roadmap. While details remain sparse, the Feynman architecture is slated for a 2028 release, following the Rubin Ultra cycle. This platform will integrate AI processors with advanced networking chips to create comprehensive “AI factories” rather than just individual components.
Nvidia is also making a play for the autonomous agent market with NemoClaw. This tool integrates with the popular OpenClaw platform to add essential privacy and safety guardrails. As AI agents begin to execute complex tasks with minimal human oversight, Nvidia aims to provide the foundational infrastructure that ensures these systems operate securely within enterprise environments.
The $1 trillion forecast is a significant jump from the $500 billion opportunity Nvidia cited just months ago. Although Nvidia recently became the first company to reach a $5 trillion valuation, some investors have voiced concerns regarding the sustainability of its growth. Huang’s presentation, which moved from showing off single chips to showcasing massive racks of interconnected systems, seemed to allay those fears. Analysts noted that the roadmap underscores a durable demand for AI infrastructure as the industry transitions from early experimentation to global, large-scale deployment.
