Semiconductors fragmenting
- The AI hardware market is shifting from one-size-fits-all GPUs toward more custom chips and diverse foundry choices. - Google split its TPU line into separate training and inference chips, Tesla picked Intel's 14A for Terafab chips, and TSMC updated its process roadmap through 2029. - That combination of custom silicon, Intel’s first external foundry win and China export dynamics is reshaping Nvidia's dominance and supplier risk ( ).
On April 22, 2026, major AI customers and foundries signaled a move away from a one‑GPU‑fits‑all model by choosing different custom chips and process roadmaps. (techcrunch.com) Training and inference solve different problems: training builds or updates model weights using massive parallel math, while inference serves live model responses with low latency and lower energy per query. (cloud.google.com) At Google Cloud Next on April 22, Google unveiled eighth‑generation TPUs split into TPU 8t for large‑scale model training and TPU 8i for inference, saying both will roll out later in 2026. (techcrunch.com) Google claims the new TPUs deliver up to roughly 3x faster training and about 80% better performance‑per‑dollar for inference versus prior generations, and the 8t design scales to thousands of chips in superpod clusters. (techcrunch.com) On the same day Tesla CEO Elon Musk said Tesla will use Intel’s next‑generation 14A process to build chips for its planned Terafab complex in Austin, marking what Reuters called Intel’s first major external 14A customer. (money.usnews.com) TSMC used its April 22 North America Technology Symposium to extend its roadmap through 2029, debuting A13 and A12 (1.3nm/1.2nm class) targeted for 2029, naming an N2U variant, and moving A16 volume production into 2027. (pr.tsmc.com) Those moves matter because Nvidia still posts record AI‑era results—Nvidia reported record fiscal‑2026 revenue of $215.9 billion and a quarterly data‑center haul cited at tens of billions—so custom chips, Intel’s foundry win and TSMC’s bifurcated roadmap could redistribute where hyperscalers buy compute. (investor.nvidia.com) U.S. export policy and China market dynamics are also shaping choices: broad export controls and subsequent licensing shifts since 2024–2025 have constrained how and which high‑end GPUs reach China, prompting customers and countries to consider alternative suppliers and domestic silicon. (gao.gov) This trend continues a multi‑year shift: Amazon Web Services has shipped multiple generations of Trainium and Inferentia chips and disclosed Project Rainier clusters of Trainium chips, while Microsoft and others have long run internal chip programs to cut costs and dependence on third‑party GPUs. (aboutamazon.com) Expect the next checkpoints to be Google’s TPU availability across cloud regions later in 2026, Intel’s upcoming earnings and customer updates in late April, and TSMC’s node‑by‑node volume milestones running through 2027–2029. (techcrunch.com)
