Per Woofun AI, JPMorgan Chase analysts Harlan Sur and Mayur Ramdhani project that custom AI ASIC unit shipments will overtake general-purpose GPUs by 2027, signaling a structural shift in AI infrastructure procurement. The report estimates total AI accelerator shipments will reach 23.3 million units, with ASICs/XPUs accounting for 53% (12.5 million) compared to GPUs at 47% (10.9 million). This transition is driven by hyperscalers including Google, Amazon, Meta, Microsoft, OpenAI, and SoftBank/Arm seeking to optimize total cost of ownership and power efficiency through tailored silicon solutions.

Broadcom and Marvell are identified as primary beneficiaries of this trend, with Broadcom commanding an estimated 80% to 85% of the high-end ASIC market. JPMorgan forecasts Broadcom's AI revenue could surge from approximately $20 billion in FY2025 to over $600 billion in FY2026, potentially reaching $1.5 trillion by FY2027, supported by projects such as Google TPU and Meta MTIA.

Concurrently, Marvell's data center revenue is expected to grow from $6.1 billion in 2025 to $14.6 billion in 2027, fueled by Amazon Trainium, Microsoft Maia, and advanced optical interconnect solutions. The analysis highlights that while GPUs remain vital for general training, custom chips like Google/Broadcom TPU7x Ironwood offer superior cost-per-watt metrics compared to Nvidia Blackwell architectures, reinforcing a bifurcated investment strategy in AI hardware.