Custom ASIC share climbs
Custom ASICs now account for about 28% of AI server shipments, a figure reported as up roughly 44.6% year‑on‑year, indicating hyperscalers and cloud providers are increasingly optimizing for cost per token rather than peak GPU throughput. The statistic was circulated on social channels as part of broader commentary on compute economics (x.com).
Custom chips built for one customer’s artificial intelligence workload are taking a bigger slice of the server market, with TrendForce projecting application-specific integrated circuit systems at 27.8% of artificial intelligence server shipments in 2026. (trendforce.com) TrendForce said global artificial intelligence server shipments will grow more than 28% year over year in 2026, while total server shipments rise 12.8%. Graphics processing unit systems still lead with 69.7% of artificial intelligence server shipments, but application-specific integrated circuit shipments are growing faster. (trendforce.com) In a separate TrendForce analysis, cloud service providers’ in-house application-specific integrated circuits are projected to grow 44.6% in 2026, compared with 16.1% for graphics processing units. The firm said the contest is shifting beyond raw chip speed toward interconnects and software ecosystems. (trendforce.com) An application-specific integrated circuit is a chip designed for a narrower job than a general-purpose graphics processor, closer to a custom factory tool than a Swiss Army knife. That tradeoff can lower operating cost when a cloud company runs the same model or service at huge scale every day. (aws.amazon.com) The workload mix has also changed. TrendForce said the market in 2024 and 2025 centered on training large language models, but the second half of 2025 brought more focus on inference, the step where a model answers prompts for products such as Copilot and Gemini. (trendforce.com) That shift has pushed buyers beyond dedicated artificial intelligence racks. TrendForce said inference workloads now run on both specialized artificial intelligence servers and general-purpose servers that handle storage plus pre- and post-inference processing. (trendforce.com) The biggest cloud companies are already building their own silicon around that model. Google said its sixth-generation Trillium Tensor Processing Unit is generally available and delivers 67% better energy efficiency than TPU v5e, while Amazon Web Services says Trainium is built for “cost efficiency” across training and inference. (blog.google) (aws.amazon.com) Microsoft has positioned Azure Maia 100 as its first in-house artificial intelligence accelerator for Azure workloads, and Meta said in March 2026 that it plans four new generations of Meta Training and Inference Accelerator chips within two years. Meta’s engineering paper on MTIA 2i said the chip cut total cost of ownership by 44% versus graphics processors in its deployment. (azure.microsoft.com) (about.fb.com) (aisystemcodesign.github.io) NVIDIA still dominates the graphics processor side of this market. TrendForce said in July 2024 that NVIDIA held nearly 90% of the market for graphics processing unit-equipped artificial intelligence servers, even as proprietary chips from Amazon Web Services, Google, Meta, Alibaba, Baidu, and Huawei expanded. (trendforce.com) The near-term result is not a graphics processor retreat so much as a split market: one lane for maximum flexibility, another for lower-cost repetition. TrendForce’s 2026 forecast says both lanes are expanding, but custom silicon is gaining share faster. (trendforce.com)
