💢💢💢

JPMorgan: AI Accelerator Shipments to Surge 62% in 2026

JPMorgan's latest forecast is revealing a significant shift in the AI chip competition:

The AI accelerator market is far from peaking, but market share is rapidly shifting from traditional GPUs to custom ASICs.

JPMorgan projects global AI accelerator shipments will reach ~16.3mn units in 2026, up 62% YoY. Among these, GPUs from $NVIDIA(NVDA.US) and $AMD(AMD.US) will account for ~58%, down from 68% in 2025. (X (formerly Twitter))

This does not imply declining demand for Nvidia and AMD GPUs.

Quite the opposite—with over 60% growth in total accelerator shipments year-over-year, absolute volumes can still surge significantly even as market shares contract.

What's truly changing is:

AI computing is transitioning from an era of "GPU dominance" to one where GPUs and ASICs coexist.

The most notable player here is Google.

$Alphabet - C(GOOG.US) is continuously expanding its deployment of custom AI chips via TPUs, while Amazon's Trainium and self-developed chips from other hyperscalers are driving rising ASIC penetration in AI infrastructure.

This points to a clearer division of labor in the future AI chip market:

GPUs: Versatility, flexibility, training, and complex inference.

ASICs: Specialized workloads, higher energy efficiency, and lower cost per compute unit.

For investors, this shift actually elevates the strategic importance of $Broadcom(AVGO.US) and $Marvell Tech(MRVL.US).

Broadcom is deeply involved in custom AI accelerator projects like Google's TPU, while Marvell benefits from demand in custom compute, networking, DSPs, and high-speed interconnects.

JPMorgan further estimates that by 2027, AI ASIC/XPU shipments could surpass GPUs, reaching ~12.5mn units versus ~10.9mn for GPUs. (KuCoin)

This signals a major structural shift in the AI chip industry:

The past question was:

"How many GPUs can Nvidia sell?"

The future question should be:

"How many AI accelerators does the world need in total?"

If total AI accelerator demand continues to grow rapidly, then even if $NVIDIA(NVDA.US)'s market share drops from 68% to 58%, it doesn't mean the overall AI chip pie is shrinking.

It may well mean:

The pie is getting bigger, with more companies sharing in it.

Thus, the next phase of AI semiconductor investment logic may no longer be about betting on a single GPU leader, but rather finding winners across the entire compute architecture upgrade:

$NVIDIA(NVDA.US) —— General-purpose AI GPU
$AMD(AMD.US) —— Second-source GPU supplier
$Broadcom(AVGO.US) —— Custom ASIC + AI networking
$Marvell Tech(MRVL.US) —— Custom compute + high-speed interconnects
$Alphabet - C(GOOG.US) —— TPU + AI infrastructure
$Amazon(AMZN.US) —— Trainium + Inferentia
The next round of AI chip competition may not be GPU vs. GPU.
It will be:
GPU vs. ASIC.
Regardless of which architecture ultimately captures the larger share, as long as AI compute demand keeps growing, the entire AI accelerator market is likely to continue expanding.
What truly matters isn't who wins 100% of the market, but who can consistently capture revenue from the explosive growth in total AI compute.

$NVIDIA(NVDA.US) $AMD(AMD.US) $Broadcom(AVGO.US) $Marvell Tech(MRVL.US) $Alphabet - C(GOOG.US) $Amazon(AMZN.US)