Google TPU demand is surging
Social signals show Google TPU demand outpacing capacity and a TSMC price target lift tied to AI revenue — summed up as 'Google TPU crushing it' for customers needing metered TPU capacity reported. That surge nudges some buyers toward cloud‑metered compute instead of hardware ownership.
Google’s VP of AI infrastructure Amin Vahdat said older TPU generations — the seven‑ and eight‑year‑old units — are running at 100% utilization and Google has been forced to turn customers away. (datacenterdynamics.com) Anthropic signed a deal for up to one million Google TPUs that Google and Bloomberg describe as bringing “well over” one gigawatt of capacity online and being worth tens of billions of dollars. (anthropic.com) Multiple industry trackers report Google trimmed its 2026 TPU production target from roughly 4 million chips to about 3 million after finding advanced‑packaging (CoWoS) slots at TSMC constrained. (moomoo.com) Fubon Research and other supply‑chain analysts model 2026 TPU shipments in the ~3.1–3.2 million range and warn full CoWoS capacity expansion won’t arrive until 2027, limiting near‑term external sales. (investing.com) Bernstein raised its TSMC price target to NT$2,200 and pointed to AI‑related revenue moving toward ~20% of TSMC’s sales by 2026 as a core justification for higher valuation and the company’s near‑$45B capex push. (finance.yahoo.com) Google already sells TPUs as metered cloud services — on‑demand, spot/preemptible, reservations and per‑chip‑hour billing are documented on Google Cloud’s TPU pricing and provisioning pages. (cloud.google.com) The combination of Anthropic’s megadeal, public performance claims for Ironwood TPUs, and Google’s metered offerings has driven several large labs and startups to shift high‑volume inference workloads to cloud TPUs rather than buying on‑prem DGX stacks, according to CNBC and DatacenterFrontier reporting. (cnbc.com)
