TAIPEI, Sept. 25, 2025 /PRNewswire/ — The global market for high-end cloud AI accelerators is approaching a major turning point. DIGITIMES forecasts that beginning in 2026, the supply chain will enter a new “decentralized” phase shaped by three forces: the expansion of custom ASICs, diversification of advanced packaging, and intensifying competition in high-bandwidth memory (HBM) pricing.
Credit: Digitimes
ASICs gain ground on GPUs
Cloud service providers are accelerating in-house ASIC development to gain tighter control over performance and costs. DIGITIMES projects that between 2025 and 2029, the compound annual growth rate for high-end cloud ASICs will reach 21%, far outpacing GPUs at 7%.
Google and Amazon, both leading adopters, are pursuing vertical integration and platform strategies that convert compute resources into cost advantages for customers. Meta and Microsoft, after years of ASIC research, are also scaling deployments. While GPUs remain indispensable for training workloads, custom ASICs are steadily eroding their market share.
Packaging moves beyond TSMC
In advanced packaging, the non-TSMC ecosystem is rapidly gaining traction. By the fourth quarter of 2026, non-TSMC suppliers are projected to account for 31% of global CoWoS-class capacity, up from 26% in late 2024.
ASE, Amkor, and SJSemi are emerging with distinct competitive angles: cost and scale, the US-based manufacturing, and China’s domestic demand, respectively. Intel is also pushing its EMIB technology, embedding silicon bridges in substrates to reduce costs and strengthen its “made in America” position. The result is a clear shift: high-end AI chips will no longer depend overwhelmingly on TSMC’s packaging lines.
Memory costs surge with HBM4
On the memory side, cloud accelerators will begin adopting HBM4 in 2026. DIGITIMES estimates that average selling prices for HBM4 will be roughly 53% higher than HBM3E, pushing memory’s share of total bill-of-materials costs above 50% for top-tier accelerators. Samsung is betting on advanced process technology and aggressive price cuts to win NVIDIA orders, directly challenging SK hynix’s long-standing dominance, while Micron continues steady capacity expansion.
Three emerging opportunities
DIGITIMES highlights three major opportunities in this shifting landscape:
Custom ASIC solutions to give cloud providers stronger cost control.
Regionalized advanced packaging capacity to meet the US manufacturing and China localization demands.
High cost-efficiency HBM offerings to deliver affordable memory in a period of steep price inflation.
Overall, the AI accelerator industry is moving away from a centralized supply model toward a more decentralized, competitive ecosystem. Only those companies that form strategic alliances and integrate technologies effectively will be able to stay ahead in this new era.
