AI accelerators · GPUs · ASICs
High bottleneckGPUs, TPUs, custom hyperscaler ASICs, and merchant inference silicon designed for training and serving large models.
This layer captures the most visible AI dollars today. It is also the most contested: incumbents face hyperscaler in-house silicon, merchant challengers, and shifting workload mix between training and inference.
Companies
14
11 pub · 3 priv
Public mcap
$18.65T
Public revenue
$1.31T
Wtd YoY
36%
Wtd GM
62%
Wtd OM
42%
Avg fund
70
Avg tech
80
Market structure
Current TAM
$200.00B
5y TAM
$500.00B
5y CAGR
20.0%
Companies
14
Margin structure
Mixed
Concentration
Concentrated
Capex intensity
Medium
Commoditization
Medium
Drivers
What grows the pie
- ↑Training compute scaling
- ↑Inference workload growth
- ↑Sovereign AI buildouts
Constraints
What can break the thesis
- ↓HBM and CoWoS supply
- ↓Hyperscaler in-house competition
- ↓Customer concentration