TPU report: Gemini 4 trained on about 200k TPU v7 chips
An ASIC supply chain update says Gemini 4 was trained on roughly 200k TPU v7 chips, while Google still lacks internal compute and allocates much of its 1H26 capacity to external customers such as Anthropic.
ASIC Update: TPU v9 Timeline & v10 Opportunities • We push out TPU v9 Humufish expectations by one quarter Core A slipped ~3 weeks → MTK Core B/I/O mods. Sampling still Nov; expect tape-out in Jan 2027, MP tape-out 2Q27, shipments 1Q28. • Constructive on EMIB-T: root causes of substrate yield (45-50%) identified → potential re-acceleration. • Google still undersupplies its own models. Gemini 4 trained on ~200k TPU v7, but heavy external allocation (esp. Anthropic) in 1H26. More internal in 2H26 helps, yet compute gap widens — Anthropic + OpenAI each +5–10 GW in 2027 vs few GW for Google. • Broadcom unchanged: 460k CoWoS wafers in 2027, shipments late 2026. MTK Zebrafish volumes higher into 2027–28, partially offsetting v9 timing. • TPU v10 (4Q28/1Q29): Icefish (MTK + GUC 3DIC on A14) the only confirmed mainstream SKU. Merope uncertain; Broadcom faces AMD+Marvell competition. EMIB-T momentum rising for v10 • Near-term headwinds for MTK; mid/long-term still like MRVL, MTK (100% v9 + potential new DC customers), and Intel (EMIB-T). #MTK #TPU #Broadcom #INTC #MRVL