JPMorgan: Custom chip shipments will surpass GPU in 2027; Broadcom TPU's "supply chain invisibility" does not indicate questionable orders

AI computing power demand is driving the chip industry from a “GPU-dominated” landscape towards the parallel expansion of both GPU and custom chips. According to JPMorgan’s latest Fall 2026 U.S. Semiconductor & Semiconductor Equipment Industry Update report, by 2027, ASIC/XPU will account for 54% of AI accelerator unit shipments, surpassing GPU; and will further rise to 55% by 2028.

JPMorgan estimates that the custom AI ASIC market will reach $60-70 billion in 2026, with a compound annual growth rate exceeding 40% to 50% in the coming years. Currently, Broadcom and Marvell occupy about 90% of market share, with Broadcom accounting for approximately 80% to 85%, indicating high market concentration.

The report also points out that supply chain information for custom chip projects such as Google TPU is relatively obscure, but this does not imply a lack of certainty around orders. The five-year TPU supply agreement between Broadcom and Google covers 2026 to 2031, involves 3nm, 2nm, and advanced packaging, and includes yearly increasing TPU revenue arrangements, providing strong revenue visibility for related AI businesses.

More broadly, AI capital expenditure remains the most important demand driver for the semiconductor industry. JPMorgan projects that global semiconductor industry revenue will grow 118% year-on-year in 2026, or 32% excluding memory; in 2027, overall growth will reach 35%, or 18% excluding memory. Meanwhile, wafer equipment spending is expected to increase by 31% and 38% in those years, respectively, sustained by demand for AI, memory, and traditional chips—collectively extending the industry cycle.

ASIC penetration accelerates, custom chips to be a key incremental force in AI computing power

Hyperscale cloud providers are aggressively developing ASIC/XPU, not merely to replace GPU, but to optimize performance, power efficiency, and per-token cost for specific workloads, while also reducing dependence on general-purpose GPU supply.

This trend is reshaping the product structure of AI accelerators. JPMorgan forecasts that in 2026, ASIC/XPU will account for about 41% of AI accelerator unit shipments, rising to 54% in 2027 and reaching 55% in 2028. Custom chips will evolve from a supplementary solution to large-scale deployment, becoming an important incremental driver in AI infrastructure expansion.

Currently, the custom AI ASIC market is highly concentrated, with Broadcom and Marvell jointly holding about 90% of the share. As cloud providers such as Google, Amazon, and Microsoft push forward with in-house AI chip development, there remains significant growth potential for ASIC design and related support demands.

Opaque Broadcom TPU information does not imply order uncertainty

One concern the market has about Broadcom’s AI business is the limited supply chain information available for projects like Google’s TPU, making it difficult for outsiders to fully track orders and shipments. However, JPMorgan believes that limited supply chain information does not in itself prove weakened demand; rather, the focus should be on customer agreements, product iterations, and capacity planning.

Taking Google TPU as an example, Broadcom’s supply agreement with Google spans five years, from 2026 to 2031, and covers 3nm, 2nm, and advanced packaging architectures. The agreement also includes annually increasing TPU revenue provisions, ensuring that despite incomplete supply chain transparency, orders and revenue remain highly visible.

Marvell similarly benefits as cloud giants’ in-house chips enter mass production, with its custom chip business supporting Amazon Trainium, Microsoft Maia, and Google XPU projects. As proprietary accelerators move from early deployment to large-scale production, ASIC design, interconnect, and advanced packaging will simultaneously benefit.

Continued AI capital expenditure further strengthens WFE and memory cycle

JPMorgan expects global cloud computing capital expenditures for CY26, CY27, and CY28 to reach $953 billion, $1.41 trillion, and $1.54 trillion, respectively. As AI investment returns gradually materialize, cloud service providers have growing economic incentives to maintain high levels of infrastructure spending.

Equipment spending will also maintain robust growth. In 2026, global WFE spending is expected to grow 31% year-on-year to about $225 billion, and rise another 38% in 2027 to $263 billion. Capacity expansion for advanced process nodes and new DRAM production will jointly boost equipment demand, while cleanroom space constraints are prompting some equipment orders to be moved forward.

Memory is another key support. JPMorgan projects that the blended average price for DRAM and NAND will increase by about 250% in 2026, and by a further 30% and 25% in 2027, respectively. AI servers are fueling DRAM demand, while enterprise SSDs are a major source of NAND demand. Long-term procurement agreements and relatively disciplined capital expenditure help prevent runaway supply increases.

Overall, JPMorgan believes that the current semiconductor upcycle is not driven by a single AI chip demand, but by the concurrent expansion of custom chips, cloud capital spending, memory, and wafer equipment across multiple industry chains. Within this framework, the sustainability of AI infrastructure investment remains the core variable determining whether the semiconductor boom cycle will continue.