Quick answer
AI compute demand has continued to outpace chip supply through 2026, driven by both training ever-larger frontier models and, increasingly, the inference demand from agentic AI running many more model calls per task than a simple chatbot response. Nvidia remains the dominant supplier, but AMD, custom silicon from major AI labs, and Groq's specialized inference chips have all captured meaningfully more share than a year ago.
The chip shortage story shifted in 2026 from being purely about training compute (building bigger models) to increasingly being about inference compute (running agents that make many model calls per task) — a demand driver that scales with usage in a way training compute didn't.
What is actually driving demand now
- Training ever-larger frontier models remains a real, ongoing driver, though growth has moderated somewhat
- Agentic AI workflows making many sequential model calls per task, multiplying inference demand per user
- Enterprise self-hosting of open-weight models adding new demand outside the major labs
- Test-time compute — models that "think longer" before answering — directly increasing compute per query
Who is gaining ground
Nvidia remains dominant, but Groq's specialized inference chips have carved out a real niche specifically for fast, cheap inference rather than training, and several major AI labs now run meaningful workloads on their own custom silicon rather than exclusively third-party chips.
The shift from "training compute shortage" to "inference compute shortage" matters because inference demand scales directly with usage — it doesn't taper off the way a one-time training run does.
Related reading
Bottom line
The AI chip shortage in 2026 isn't primarily a training-compute story anymore — it's increasingly about inference demand from agents and reasoning models that scales with usage. That shift changes who benefits from new capacity and how AI pricing evolves going forward.
