JPMorgan: Custom Chip Shipment Share to Surpass GPUs by 2027
nashnova research
JPMorgan forecasts custom chip shipments will hit 54% of AI accelerator units by 2027, overtaking GPUs for the first time; the custom chip market already reaches $60–70 billion in 2026, signaling a structural shift away from GPU dominance.
Why can custom chips overtake GPUs?
Hyperscalers are building custom chips — ASICs/XPUs, processors designed for a specific workload rather than general-purpose computing — not to replace GPUs outright, but to cut power consumption and per-token cost on targeted tasks while reducing reliance on Nvidia's general-purpose GPU supply.
JPMorgan's timeline: custom chips reach ~41% of AI accelerator shipments in 2026, 54% in 2027, and 55% in 2028.
This means → the crossover arrives in 2027 — custom chips take the majority for the first time, and GPUs lose their unchallenged lead.
Who is making the money?
The custom AI chip market is worth roughly $60–70 billion in 2026, growing at a 40–50%+ CAGR over the next several years.
Broadcom and Marvell together hold ~90% of that market; Broadcom alone commands 80–85% — concentration is extremely high.
Marvell's roster spans Amazon Trainium, Microsoft Maia, and Google XPU programs; as these in-house accelerators shift from early deployment to volume production, ASIC design, interconnect, and advanced packaging all scale in parallel.
Broadcom's supply chain is opaque — are orders at risk?
A persistent worry: supply-chain details around Google TPU and similar programs are hard for outsiders to track.
JPMorgan's take: limited visibility ≠ weakening demand — customer agreements, product roadmaps, and capacity plans matter more.
In plain terms = Broadcom and Google have a five-year supply agreement (2026–2031) spanning 3 nm, 2 nm, and advanced packaging, with annually rising TPU revenue commitments — the contract locks in enough to keep order visibility solid.
Not just chips — is the whole supply chain expanding?
JPMorgan projects 2026 global semiconductor revenue growth of 118% year-on-year, or 32% excluding memory; 2027 overall growth of 35%.
Global cloud capex is forecast at $953 billion in 2026, $1.41 trillion in 2027, and $1.54 trillion in 2028 — rising every year with no sign of a slowdown.
Equipment is equally strong: 2026 global wafer-fab equipment spending is expected to rise 31% to roughly $225 billion, then another 38% to ~$263 billion in 2027.
Are memory chips repricing too?
JPMorgan forecasts 2026 blended ASPs for both DRAM and NAND rising roughly 250%, with further increases of ~30% and ~25% respectively in 2027.
This means → memory is not a bystander — AI training and inference demand for high-bandwidth memory is pushing storage prices into a fresh upcycle.
How long can this semiconductor boom last?
JPMorgan's core view: the current upcycle is not driven by a single AI-chip narrative but by custom chips, cloud capex, memory, and wafer equipment expanding simultaneously.
This reflects a shift from a single breakthrough product to a full-chain infrastructure pull.
In plain terms = as long as hyperscalers keep deploying capital, this cycle has room to run — the durability of AI infrastructure spending is the single most important variable deciding how long it lasts.
市场有风险，内容仅供研究参考，不构成投资建议。