DIGITIMES said on 19 October that it expects 25.951 million high-end cloud AI accelerators to ship in 2027, up 64.5% year over year. These are the chips that run AI models in data centers. It expects custom chips, called ASICs, to reach 15.262 million units, up 110.9%. DIGITIMES said that would top GPU shipments for the first time. It credits cloud providers such as Google and AWS, which are building more of their own chips. Nvidia stays the largest vendor at 9.096 million units, up 18.2%, but its share slips to 35.1%. Google is forecast at 8.009 million units, up 134%, with a 30.9% share. That leaves it 4.2 points behind Nvidia.
These numbers are DIGITIMES estimates, not reported sales. The full report sits behind a paywall, so the summary does not explain how the figures were built. DIGITIMES said some parts may still limit supply of top-spec chips. It named special glass fabric, ABF substrates, circuit boards with many layers, and HBM memory. It also said Nvidia is weighing a tiered memory setup for its Rubin Ultra chip to use less memory per chip. The summary does not say what Nvidia decided.
For teams that rent cloud compute, DIGITIMES expects more of it to run on custom chips such as Google's TPU and Amazon's Trainium, not only on Nvidia GPUs. It ties demand to agentic AI, which takes many steps per task and uses more tokens.