Nvidia sees $1t in AI chip sales through 2027

Nvidia said it expects at least US$1 trillion in sales from its Blackwell and Vera Rubin chips through 2027.

Jensen Huang, CEO, announced the revised outlook at the GTC event, extending a prior projection of US$500 billion by end-2026.

At GTC, Nvidia unveiled a chip using technology it acquired from startup Groq and showed off a computer made up of general-purpose CPUs.

The company said the next flagship AI processor family, Vera Rubin, will appear in systems in H2 2026.

Investors have pressed for evidence the AI market can sustain growth as rivals like AMD and customers building in-house chips increase competition.

Shares were down 3.4% year-to-date heading into GTC and market value was about US$4.4 trillion.

Nvidia used GTC to highlight partnerships and product roadmaps while seeking to keep customers on its platform.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

The competitive pressure comes from many directions, not one rival.
Nvidia is aiming for a second-half 2026 launch for Vera Rubin systems, while AMD plans Helios, its first rack-scale AI system, later in 2026 1.
Big buyers are pushing their own chip programs faster. Google, Amazon, and Microsoft all run annual release schedules for their in-house silicon, custom chips built for their data centers 2.
This shared pace makes a one-year product cycle a requirement for holding share. Mizuho Securities, a Japanese investment bank and research firm, puts Nvidia’s share between 70% and 95% 3.

Power still matters, though buyers increasingly weigh cost per token.
Nvidia says Rubin should deliver 10 times more performance per watt than earlier systems, which can cut inference costs at scale, meaning the cost to run AI models to generate outputs 1.
The server market is moving away from one standard chip for every job.
Custom-designed chips, or application-specific integrated circuits (ASICs), are projected to rise from 20.9% of AI servers in 2025 to 27.8% in 2026 as hyperscalers, the largest cloud companies running massive data centers, build silicon tuned to specific workloads 4.
Nvidia’s roadmap also leans into this split, including Rubin CPX for long-context inference workloads, running models on very long prompts or documents 2.