Nvidia just posted another blockbuster quarter, and the market barely flinched. That tells you something about where expectations sit for the company that essentially became the arms dealer of the AI revolution.

The real conversation coming out of earnings isn’t about the numbers Nvidia just printed. It’s about inference, the less glamorous but potentially far more lucrative half of the AI compute equation, and whether Nvidia can dominate it the way it dominated training.

Training won the war. Inference is the occupation.

Training is the process of building an AI model, feeding it massive datasets until it learns patterns. Inference is when that trained model actually does stuff: answers your question, generates an image, recommends a video, flags a fraudulent transaction.

Training happens once (or periodically). Inference happens millions of times per second across the global economy. Think of training as writing a cookbook and inference as cooking every meal, forever.

The company currently controls approximately 80% of the AI accelerator market, a dominance built largely on the back of training workloads. Every major tech company racing to build foundation models has been writing enormous checks to Nvidia for its H100 and now Blackwell GPUs.

CEO Jensen Huang has been beating this drum for months. He has projected that Nvidia’s AI hardware could drive a potential $1 trillion revenue opportunity by 2027, with inference playing a pivotal role in reaching that figure. During his keynote at GTC, Huang made inference a central theme, positioning Nvidia’s roadmap squarely around the workload that scales with adoption rather than development.

The numbers behind the narrative

Nvidia’s recent quarter delivered approximately $57 billion in revenue, comfortably surpassing analyst expectations. Year-over-year growth was driven almost entirely by AI initiatives.

Hyperscalers like Google, Amazon, and Microsoft are developing custom silicon specifically optimized for inference. Google’s TPUs have been running inference at scale for years. Amazon’s Inferentia chips are designed to undercut Nvidia on price-per-inference. New ASIC vendors are entering the market with chips that sacrifice training flexibility for raw inference efficiency.

AMD presents perhaps the most direct competitive threat. Benchmarks suggest that AMD’s MI300X series may outperform Nvidia’s offerings in certain direct hardware purchase scenarios. But Nvidia remains highly competitive in cloud rental scenarios, which is how most companies actually access AI compute.

Nvidia’s CUDA software ecosystem, built over nearly two decades, acts as a moat that pure hardware comparisons miss entirely.

What this means for investors

The risk is that inference workloads are inherently more price-sensitive than training. When you’re building a frontier model, you’ll pay whatever it costs for the best hardware. When you’re running inference at scale on millions of queries per second, every fraction of a cent per computation matters. That price sensitivity opens the door for cheaper alternatives, whether they come from AMD, custom hyperscaler chips, or startups nobody’s heard of yet.

For crypto-adjacent investors, this dynamic has direct relevance. Decentralized AI compute networks, projects building marketplaces for GPU access, rely heavily on Nvidia’s hardware as their underlying infrastructure. Nvidia’s pricing power and supply allocation decisions ripple through the cost economics of every decentralized GPU project. If inference demand pushes Nvidia GPU prices higher, decentralized compute networks become more attractive as cost-effective alternatives. If competition drives inference costs down, the value proposition shifts.