AI infrastructure investing: Are custom AI chips the real bubble?

The main AI infrastructure risk is not collapsing AI demand but the assumption that every proprietary AI-chip architecture will reach sufficient scale and retain attractive economics.

• Custom ASICs emerged as a rational procurement response to scarce and expensive NVIDIA GPUs, offering hyperscalers lower costs, supply diversification and bargaining leverage.
• MediaTek estimates that the AI ASIC market could reach about $70 billion to $80 billion in 2027, but that forecast assumes today’s architectural experimentation becomes a permanent industry structure.
• Hyperscalers are increasingly capturing supplier economics: AMD granted equity warrants linked to large AI purchase commitments from Meta and OpenAI, while Marvell entered a similar arrangement with Google.
• NVIDIA’s advantage extends beyond GPUs to CPUs, networking, rack-scale infrastructure and the CUDA software ecosystem, while tools such as TileRT can improve performance through software without requiring another accelerator.
• SpaceX’s plan to build exclusively around NVIDIA’s Vera Rubin architecture, despite Elon Musk’s experience with proprietary AI silicon, illustrates the hurdle facing custom architectures; NVIDIA’s NVLink Fusion may also let custom accelerators operate inside NVIDIA-based systems.

Takeaway: AI compute spending can continue to grow while far fewer custom architectures achieve economic scale, potentially improving hyperscaler free cash flow as suppliers face lower pricing power and weaker returns.