Nvidia just posted another quarter that would make most companies weep with envy. Revenue hit $57 billion for Q3 FY2026, up 62% year-over-year, with the data center segment alone pulling in $51.2 billion.
The numbers behind the narrative
Nvidia’s data center revenue of $51.2 billion grew 25% quarter-over-quarter and 66% year-over-year. That puts the company’s data center operations above a $200 billion annualized run rate.
The company controls roughly 80% of the AI accelerator market. That dominance isn’t just about hardware. It’s about the CUDA software ecosystem that locks developers into Nvidia’s architecture.
The global AI inference market is projected to grow from $106.15 billion in 2025 to $254.98 billion by 2030, representing a compound annual growth rate of 19.2%. The Blackwell and Blackwell Ultra architectures are specifically designed to improve inference economics, making it cheaper and faster to run AI models in production.
Why inference matters more than training
Training an AI model is a one-time (or periodic) expense. Inference—running that trained model to serve actual users—happens continuously. Every ChatGPT query, every AI-generated search result, every automated customer service interaction is an inference workload. Some industry estimates suggest inference could eventually account for 80-90% of all AI compute demand.
Nvidia’s CEO Jensen Huang has been telegraphing this transition for quarters. The Blackwell architecture is architected to handle the specific computational patterns of inference, where latency and cost-per-query matter as much as raw throughput.
The market’s strange reaction
Despite posting numbers that crushed expectations, Nvidia shares have traded lower. The stock sits at less than 22x forward earnings, a valuation that feels modest for a company growing revenue at 62% annually.
While Nvidia dominates training, inference workloads are more diverse and potentially more accessible to competitors. Custom chips from Google, Amazon, and a growing cohort of startups are all targeting inference specifically. The decentralized GPU network space is also growing, potentially offering inference compute at lower costs by aggregating underutilized hardware.
The sub-22x forward earnings multiple suggests the market is pricing in some version of this risk. Whether that’s an opportunity or a warning depends entirely on whether Nvidia can maintain its dominance as AI shifts from building models to running them at scale.