Silicon Gold Rush
The shape, share, and leverage points of the AI accelerator market — sourced from primary filings and operator notes, not vendor decks.
Market Shape
The AI accelerator market sits at $71.3B in 2024 and is widely projected to push past $91B by year-end 2025 — roughly 30% YoY. The five-year ceiling, depending on inference deployment density, is forecast around $400B.
Who Holds It
70–95% of AI accelerator share. ~78% gross margin. Pricing power so wide it sets the ceiling for everyone else.
Instinct MI300X gaining traction in HBM-heavy inference. ~47% gross margin — visibly thinner runway than NVIDIA.
Gaudi 3 positioning around price/perf parity. ~41% gross margin. Foundry pivot continues.
Google TPUs, AWS Trainium / Inferentia, Microsoft Maia. Vertical integration eats into the merchant silicon TAM.
Where the Leverage Is Shifting
- → GPUs still dominant for training; inference is splitting toward ASICs and accelerators with HBM-rich packages.
- → CoWoS / advanced packaging capacity is the real bottleneck — TSMC's allocation is the real market clearing price.
- → Edge inference is creating a parallel chip market: device-specific NPUs, neuromorphic, ultra-low-power designs.
- → Silicon photonics and chiplet interconnects emerging as the next architectural axis after process-node gains slow.
- → National semiconductor strategies (US CHIPS, EU Chips Act, India's PLI) are reshaping where capacity gets built.