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

June 22 – JPMorgan Chase is highlighting a new growth cycle for the custom application-specific integrated circuit (ASIC) market, as leading cloud computing firms and tech giants move away from exclusive reliance on general-purpose graphics processing units (GPUs) to cut AI computing costs and improve efficiency. Broadcom and Marvell Technology Group stand to be the biggest beneficiaries of this shift, according to the U.S. bank’s analysis. In a recent semiconductor industry research note, JPMorgan analysts Harlan Sur and Mayur Ramdhani project the global digital AI ASIC market will reach $600 billion to $700 billion by 2026, with a compound annual growth rate (CAGR) of 40% to 50% in the coming years. Currently, Broadcom controls 80% to 85% of the high-end AI ASIC market, while Marvell holds a 10% to 12% share, the report states. Booming AI computing demand is reshaping chip procurement models, JPMorgan noted. Customers including Google, Amazon, Meta, Microsoft, OpenAI, SoftBank, and Arm are accelerating development of in-house or custom AI processors to achieve better performance, power efficiency, and total cost of ownership (TCO). Unlike Nvidia and AMD’s general-purpose GPUs, ASICs are tailored for specific clients, software stacks, or platforms—making them a better fit for hyperscale cloud providers handling massive internal workloads. Looking at revenue forecasts: Broadcom’s AI revenue is expected to surge from roughly $20 billion in fiscal 2025 to over $600 billion in fiscal 2026, then climb further to more than $1.5 trillion in fiscal 2027. Its project pipeline includes chips for Google’s TPU, Meta’s MTIA, ByteDance’s AI video and networking chips, OpenAI’s XPU, SoftBank/Arm’s XPU, and Anthropic’s rack-scale TPU solutionsFor Marvell, JPMorgan projects its data center revenue will grow from ~$6.1 billion in 2025 to approximately $9.3 billion in 2026, and reach around $14.6 billion in 2027. Key growth drivers include Amazon’s Trainium 3 and Trainium 4, Microsoft’s Maia, Google’s SmartNIC/DPU, CXL controllers, plus 800G/1.6T optical DSP, coherent lite, and early coherent packaging optical (CPO) solutions. A critical 2027 projection: Unit shipments of AI ASICs and processing units (XPUs) will surpass those of GPUs. Total AI accelerator units are forecast to hit 23.3 million by 2027, with GPUs accounting for 10.9 million (47% of the total) and ASICs/XPUs making up 12.5 million (53%). This means while GPUs will continue growing, custom chips will capture a larger share of new AI workload deployments. Using Google and Broadcom’s TPU7x Ironwood versus Nvidia’s Blackwell as examples, JPMorgan said AI ASICs are competitive on cost-effectiveness and power efficiency. The TPU7x Ironwood has FP8 performance close to Nvidia’s B200 and B300, but is estimated to cost ~$13,000—compared to $35,000 for the B200 and $40,000 for the B300. It also outperforms GPUs on cost per compute dollar and per watt. This projection does not signal a sharp decline in Nvidia demand, though. Instead, it points to a split in AI infrastructure spending: GPUs will remain the go-to for general training and inference needs, while cloud providers’ in-house ASICs will gain stronger adoption for large-scale, stable, predictable internal workloads. For investors, JPMorgan’s report underscores the shift in AI hardware investment away from GPUs, toward ASICs plus adjacent areas like advanced packaging, high-bandwidth memory (HBM) interfaces, SerDes, optical interconnects, and CPO expansion. If the report’s forecasts hold, Broadcom and Marvell won’t just be AI networking or interconnect suppliers—they will become core players in the next phase of AI computing architecture transformation.