Google launches TPU 8tx and 8i Max; SemiAnalysis data shows TPUv7 beating GB300 NVL72 on TCO
AccBalanced · x · 2026-11-11
SemiAnalysis reports Google's announcement of TPU 8tx for confidential inference and TPU 8i Max, building on rapid TPUv7 inference-stack progress (Muse Spark Open, SGLang, RadixArk) that largely transfers to the TPUv8 series. The accompanying InferenceX preview chart shows Muse Spark Open 816B on TPU7x (SGLang) delivering more tokens per $1 TCO than GB200 NVL72 across most interactivity ranges (official preview, may change with validation). Chip ownership cost: $1.86/chip-hr for GB200 vs $1.21 for TPU7x at hyperscaler volume.
Related event: Google Ships Four Eighth-Gen TPU Systems in One Year, Taking Aim at Nvidia(19 posts)→
More from Infra
- First SGLang Summit draws 2,000+ attendees as the project grows from under 10% share to ecosystem core — ying11231 · 2026-11-11
- Neon ships branch-aware object storage rollback, 24 function templates, CJK search — samgoodwin89 · 2026-11-11
- Musk: Grok 4.8 ditches JAX/SGLang, RL and inference stack rewritten from scratch in C/C++ — r0ck3t23 · 2026-11-11
- Anthropic's amended S-1: $170B+ Microsoft/Google compute deals, IPO at up to $1.4T valuation — rohanpaul_ai · 2026-11-11
- Creon: algorithmic gains will eclipse hardware improvements — the reverse bitter lesson again — bradneuberg · 2026-11-11
- Commerce Dept. delays Megaspeed blacklist decision, drawing accusations of export-control failure — GaryMarcus · 2026-11-11