The companies that design the processors, accelerators, and compute architectures powering AI training and inference. This is the highest-visibility layer of the AI stack — the chips themselves — where performance leadership, architectural lock-in, and ecosystem gravity determine who captures the majority of value.
Why it matters
- AI training is compute-bound. Larger models, longer context windows, and multi-modal architectures keep pushing total compute demand faster than Moore's Law improves efficiency. The market for AI accelerators is growing faster than any prior semiconductor cycle.
- Inference is the next leg. Training gets the headlines, but inference — running trained models in production — is where the sustained, recurring compute demand lives. The inference TAM is growing as AI agents, search, and enterprise workflows scale.
- Architecture moats are deep. CUDA ecosystem lock-in, custom silicon design partnerships, and IP licensing models create structural advantages that are difficult to replicate even with competitive hardware.
Roster
- NVDA — NVIDIA — GPU-based AI accelerators and the CUDA software ecosystem. Dominant in both training and inference today.
- AMD — Advanced Micro Devices — Instinct MI-series data center GPUs plus EPYC server CPUs. The most credible alternative to NVIDIA in merchant AI silicon.
- ARM — Arm Holdings — CPU architecture IP licensed into virtually every mobile, embedded, and increasingly server and AI edge processor. Royalty-based model with leverage to compute volume.
- MRVL — Marvell Technology — custom silicon (ASIC) design for hyperscaler AI accelerators plus data infrastructure semiconductor IP.
What to watch
- NVIDIA data center revenue growth — the single largest real-time measure of AI compute demand.
- AMD Instinct adoption — any meaningful MI300/MI400 wins at hyperscalers validate the second-source thesis.
- Custom silicon share — Marvell and Broadcom's ASIC backlog signals whether hyperscalers are diversifying away from merchant GPU.
- Arm royalty growth — rising royalty per chip signals the shift toward higher-value compute in server and AI edge applications.
- Supply chain capacity — TSMC advanced-node allocation determines who can actually ship silicon on time.