Marvell details Google chip deal worth up to $120B as inference accelerators split from TPUs
demian_ai · x · 2026-10-08
Marvell's investor day spelled out its Google deal: custom chips plugging into the TPU ecosystem, including inference accelerators, worth up to $120B through 2033 if purchase targets are met; a rumored inference-only chip reportedly pairs two-to-one with TPUs (unverified).
Reading a prompt is compute-bound, ideal for TPUs/GPUs, while decoding token-by-token is memory-bound — every token requires pulling model weights and context back from memory. Etched CEO Gavin Uberti noted decode is all a memory game: on Blackwell, one GPU-to-GPU hop takes 4,000ns, so eight GPUs don't make you 8x faster. Hardware is splitting accordingly, putting pressure on memory bandwidth and interconnects — interconnect alone is $37.5B of Marvell's projected growth.
More from Infra
- Will-It-Jev: Rust 5-tier cascade router routes simple prompts away from LLMs with sub-5ms p99 latency — forestmars · 2026-10-08
- Inside Ambarella: the low-power vision chips set to power robots, drones and AI glasses — Scobleizer · 2026-10-08
- Cloud agent runtime plus Tailscale: the "local hands" setup devs prefer — blelbach · 2026-10-08
- NVIDIA tutorial: deploy a standalone offline VLM kiosk on Jetson with Codex — NVIDIA Developer · 2026-10-08
- Liquid Inference launches aggregator routing across six inference providers — markjeffrey · 2026-10-08
- GitHub Copilot will offload tasks to local models like MAI-Code-1.1 Flash via HydraFusion — mariorod1 · 2026-10-08