Custom chips gain ground as Nvidia retains revenue lead

Custom AI accelerators are gaining shipment share without displacing Nvidia’s dominance in revenue. TrendForce forecasts cloud providers’ in-house ASIC shipments will grow 44.6% in 2026, compared with 16.1% for GPUs, with ASIC-based systems reaching 27.8% of AI server shipments. Silicon Analysts projects Nvidia’s revenue share at about 75% for 2026. Lower prices for custom chips help explain why shipment gains do not translate directly into comparable revenue gains.

Google is expanding its TPU business beyond internal workloads, offering capacity through Google Cloud and allowing some customers to install processors in their own data centers. Anthropic has agreed to expand its use to as many as 1 million TPUs. Google also open-sourced TorchTPU, enabling PyTorch code and Hugging Face models to run natively without rewrites, potentially reducing switching barriers.

Broadcom’s AI semiconductor revenue reached $16.7 billion, up 221% year over year, reflecting growing custom-chip investment. That business also includes networking chips, limiting direct comparisons with accelerator sales. Its OpenAI partnership adds custom processors alongside, rather than replaces, OpenAI’s GPU capacity.

Purpose-built chips are best suited to stable, high-volume inference workloads where power consumption and cost matter. Nvidia GPUs retain advantages for evolving frontier-model training, and major custom-chip customers continue buying them. Cost and performance claims largely come from chipmakers, while independent testing of Google’s newest systems remains unavailable. Market-share estimates also rely on forecasts and incomplete revenue disclosures.