ASIC Shipments Expected to Surpass GPUs for the First Time by 2027 as AI Reshapes Server Storage Architecture
Nashnova编辑部
DIGITIMES analyst Stella Weng forecasts ASIC shipments will hit 15.3 million units in 2027, surpassing GPUs at 10.7 million for the first time; meanwhile, AI server storage enters an innovation cycle with three competing architecture paths.
What is the "golden crossover," and why 2027?
ASICs — chips custom-built for a specific task, unlike general-purpose GPUs — are projected at 7.2 million units in 2026, up 42.3% year-on-year. GPUs hit 8.5 million, growing 30%. The gap is already closing.
By 2027, ASIC growth accelerates to 110.9%, reaching 15.3 million units. GPU shipments rise to 10.7 million but slow to 25.2% growth. This means → ASICs win not on absolute volume today, but on the slope: growth jumps from 42% to 111% in a single year.
In plain terms = GPUs are still selling well, but custom-chip demand is doubling — whoever's customers are designing their own chips is driving the crossover.
How is Google closing in on Nvidia?
Nvidia's 2027 chip shipments are forecast at 9.1 million units, up 18.2% — still number one. Google is projected at 8 million, up 134%. The gap shrinks from an order of magnitude to just 1.1 million units.
Google's TPU — its in-house AI training chip — draws demand from two sources: internal consumption by Gemini and Google Cloud, and external sales to Anthropic, Meta, and others. This means → Google is becoming a chip supplier, not just a captive buyer.
Supply bottlenecks are the biggest variable: upstream fiberglass cloth, ABF substrates — a key chip-packaging material — and TSMC 3 nm capacity could all constrain output. This reflects a reality where growth forecasts hinge on supply chains, not demand.
Why is CPU demand accelerating too, and what is driving it?
CPU shipments are projected at 36.1 million units in 2026, up 27.1%, rising to 52.6 million in 2027, up 45.7%.
The driver is the spread of agentic AI — AI systems that plan and execute multi-step tasks on their own. These run on general-purpose servers, not on GPU clusters built for model training.
In 2027, general-purpose server CPUs are forecast at 42.2 million units versus 10.4 million for AI-server CPUs. In plain terms = AI is not just a "buy more GPUs" story; general CPU demand runs four times higher than AI-specific CPU demand.
How will storage architecture change in 2027, and which path lands first?
AI server storage is expected to enter an innovation cycle in 2027, advancing along three paths: processing near memory (PNM), 3D memory-logic stacking, and high-bandwidth flash (HBF).
PNM — placing compute logic next to the memory chip rather than inside it — outperforms processing-in-memory (PIM) on power, performance, and area. It is expected to commercialize faster. On the 3D stacking front, Nvidia, AMD, and Google are all active, stacking not just HBM/LPDDR DRAM but also converting planar SRAM to 3D-stacked SRAM.
HBF can exceed HBM capacity by more than ten times, but trails on bandwidth and latency. This means → HBF suits "large capacity, less extreme speed" workloads, yet standardization and the software ecosystem remain immature. Commercial deployment by 2028 is still considered unlikely.
Where is the shared bottleneck across all paths?
Whether TSMC's 3 nm capacity ramps on schedule is the common gate for ASIC volume growth, Google TPU expansion, and new storage architectures alike.
Upstream material elasticity — fiberglass cloth and ABF substrates — matters just as much. Chip designs can move fast, but packaging materials that lag behind stall everything.
In plain terms = the 2027 "golden crossover" and the storage overhaul are drawn on paper. Whether they materialize depends on factories and materials, not slide decks.
Content is for reference only, not financial advice.