Google/Meta TPU rental chatter
Beyond the $1B rental reports, social commentary frames TPUs as an efficiency win in large‑scale training and inference deployments, with TPU demand pressuring foundry capacity. The chatter amplifies competitive pressure on GPU supply and TSMC wafer allocation. ( )
Meta entered multi‑billion‑dollar talks to rent Google Cloud TPUs in 2026 with an option to buy hardware for on‑prem deployment in 2027, according to reports citing people familiar with the negotiations. (economictimes.indiatimes.com) Google engineering leadership said older TPU generations are running at full capacity — “seven‑ and eight‑year‑old TPUs” at 100% utilization — and the company has turned away customers because demand is oversubscribed. (datacenterdynamics.com) TrendForce reported NVIDIA raised Rubin’s HBM4 per‑pin speed target above 11Gbps in 3Q25, forcing SK hynix, Samsung and Micron to redesign HBM4 samples and pushing HBM4 mass production out to the end of Q1 2026 at the earliest. (trendforce.com.tw) Taiwan’s Commercial Times (via Benzinga) says suppliers are redesigning HBM base‑dies and that the adjustment could delay shipments by roughly one quarter, prompting Nvidia to trim initial Rubin wafer starts and lean on continued Blackwell production. (benzinga.com) Independent analyses project AI accelerators could soak up most N3 logic wafer capacity — SemiAnalysis estimates roughly 86% of TSMC’s N3 could go to AI accelerators by 2027 — while TSMC executives have warned current advanced‑node capacity is “about three times short” of customer plans. (the-decoder.com) Supply‑chain trackers report TSMC’s CoWoS advanced‑packaging lines are effectively full, spurring plans to outsource parts of packaging and to aggressively expand CoWoS output through 2026 to relieve lead‑time pressure. (trendforce.com)
