JPMorgan: AI Custom Chip Shipments May Surpass GPUs in 2027, Broadcom and Marvell Seize the Opportunity

BlockBeats News, June 22nd, JPMorgan Chase stated that as large cloud computing companies and tech giants seek to reduce AI ​​computing costs, improve efficiency, and move away from sole reliance on general-purpose GPUs, the custom ASIC chip market is entering a new growth cycle, with Broadcom and Marvell poised to be the biggest beneficiaries of this trend.

In a recent semiconductor industry research report, JPMorgan analysts Harlan Sur and Mayur Ramdhani estimated that the digital AI ASIC market will reach around $600 billion to $700 billion by 2026 and maintain a compound annual growth rate of over 40% to 50% in the coming years. The report stated that Broadcom currently holds about 80% to 85% of the high-end ASIC market share, with Marvell ranking second at about 10% to 12%.

The rapid growth in AI computing demand is reshaping the chip procurement structure. JPMorgan believes that customers such as Google, Amazon, Meta, Microsoft, OpenAI, and SoftBank/Arm are accelerating the development of in-house or custom AI processors to achieve better performance, power consumption, and total cost of ownership. Unlike Nvidia and AMD's general-purpose GPUs, ASICs are typically designed for specific customers, specific software stacks, or specific platforms, making them more suitable for hyperscale cloud providers with large internal workloads.

The report predicts that Broadcom's AI revenue will increase significantly from around $20 billion in FY2025 to over $600 billion in FY2026 and further to over $1.5 trillion in FY2027. Its project pipeline includes Google TPU, Meta MTIA, ByteDance AI video and networking chips, OpenAI XPU, SoftBank/Arm XPU, and Anthropic-related TPU rack-scale solutions.

As for Marvell, JPMorgan expects its data center revenue to grow from around $6.1 billion in 2025 to approximately $9.3 billion in 2026 and reach about $14.6 billion in 2027. Growth drivers include Amazon Trainium 3 and Trainium 4, Microsoft Maia, Google SmartNIC/DPU, CXL controllers, as well as 800G/1.6T optical DSP, coherent lite, and early CPO solutions.

The report also made a key projection: by 2027, the unit shipment of AI ASIC/XPU will surpass that of GPU. JPMorgan estimates that by 2027, the total AI accelerator shipments will reach 23.3 million units, with GPUs at 10.9 million units, representing 47%, and ASICs/XPUs at 12.5 million units, representing 53%. This implies that while GPUs will continue to grow, custom chips may take a larger share in new AI workload deployments.

Using Google/Broadcom TPU7x Ironwood and Nvidia Blackwell as examples, JPMorgan stated that AI ASICs are competitive in terms of cost-effectiveness and power efficiency. The report reveals that TPU7x Ironwood's FP8 performance is close to Nvidia's B200/B300, but with an estimated price of around $13,000, lower than B200's $35,000 and B300's $40,000; its cost per dollar of compute and per watt of compute also outperform the GPU benchmarks.

This projection does not mean a rapid weakening of Nvidia's demand. Instead, it points to a differentiation in AI infrastructure investment: GPUs will continue to serve general training and inference needs, while cloud providers' in-house ASICs will achieve higher penetration in large-scale, stable, and predictable internal workloads.

For investors, JPMorgan's report reinforces the AI hardware chain's shift from GPUs towards ASICs, advanced packaging, HBM interfaces, SerDes, optical interconnects, and CPO diffusion logic. If the predictions in the report materialize, Broadcom and Marvell will not only be AI networking or interconnect chip suppliers but will become core platform companies in the next stage of AI computing architecture migration.