SC26: Google pitches TPUs to GPU-default devs, Microsoft shows off HBv5
AccBalanced · x · 2026-11-22
At SC26, hyperscalers pushed custom silicon: Google ran an interactive TPU Explorer kiosk next to TPU 8t/8i samples, pitching native PyTorch/vLLM/SGLang to developers who'd default to GPUs, with case studies from DeepMind, Anthropic and Midjourney. Microsoft displayed an open-lid Azure HBv5 server with custom AMD EPYC+HBM CPUs aimed at HPC, emphasizing fast deployment of MI300X endpoints on Foundry.
More from Infra
- Gary Marcus mocks the biggest data center ever, tied to OpenAI, SB Energy and Nvidia — GaryMarcus · 2026-11-22
- AI stock selloff is a leverage squeeze: neoclouds down ~63%, CoreWeave CDS spikes — tengyanAI · 2026-11-22
- Reddit debate: is 16GB VRAM still viable for local video generation? — Entire_Nectarine8291 · 2026-11-22
- Baseten's partner list sparks debate: prompts pass through ~20 GPU middlemen — suchenzang · 2026-11-22
- Who pays for the AI buildout? Following the cash from investors to GPUs to customers — Bayzic_com · 2026-11-22
- Running MoE models on RX 6650 XT + MI25: how much does PCI-E bandwidth matter? — Kidplayer_666 · 2026-11-22