Hyperscalers Shift to Custom AI ASICs Over Generic GPUs

The AI infrastructure market is transitioning toward custom application-specific integrated circuits (ASICs) as hyperscalers seek alternatives to off-the-shelf GPUs. Broadcom Inc. and Marvell Technology currently control approximately 95% of the custom AI ASIC co-design market, with ASICs projected to reach 27.8% of the AI server compute market by 2026.

Broadcom has secured a long-term agreement with Alphabet Inc. to develop Tensor Processing Units (TPUs) through 2031 and established partnerships with Meta Platforms and OpenAI. Additionally, Anthropic has committed over $10 billion to custom silicon and will access 3.5 gigawatts of TPU-based compute via Broadcom starting in 2027. Marvell Technology, which architects Amazon's Trainium and Microsoft's Maia chips, is in talks with Alphabet for two new custom chips and expects up to $11 billion in AI ASIC revenue for 2026.

While Nvidia remains the largest GPU producer with 85% year-over-year revenue growth, its stock growth has slowed in 2026 compared to memory makers like Micron Technology and Sandisk. In March 2026, Nvidia invested $2 billion in Marvell Technology. Nvidia forecasts that global data center capital expenditures will reach $3 trillion to $4 trillion annually by 2030.