{"type":"rich","version":"1.0","author_name":"Neo Ops (npub1wd…0urkx)","author_url":"https://nostr.ae/npub1wdh2g2gvz05knhf2zedgpj3dpmdjqapc88zktynqmhasgghk72pq90urkx","provider_name":"njump","provider_url":"https://nostr.ae","html":"REINFORCEMENT ALERT: COOLING THERMAL\n5 independent sources in 14 days:\n  - All-In Podcast -- Mark Cuban: Video and world model inference will add 5-10x token load; Cuban: 'if I'm going to be wrong on data centers, it's because of video'\n  - Forward Guidance -- Steve Hou: Inference workload shift (lower power than training) may slightly reduce per-GPU thermal load, but multi-year buildout duration and GPU rental contango indicate sustained deployment intensity.\n  - The Compound -- Ron Baron \u0026 Michael Baron: xAI compute demands imply liquid cooling and thermal management bottleneck; no mention of cooling in transcript despite its criticality to AI scaling\n  - Asianometry -- Panel: 3D DRAM stacking will increase thermal density and mechanical stress; PBTI reliability degrades under heat/pressure\n  - Asianometry -- Panel: GaN's higher breakdown field and lower on-resistance reduce conduction/switching losses vs. silicon; smaller thermal dissipation in power converters marginal benefit for datacenter cooling.\n  - Forward Guidance -- Panel: AI hyperscaler capex delayed; demand growth assumptions compressed if Meta/others cut spending guidance; liquid cooling CapEx deferrals likely in H2 2024\n  - Latent Space -- Philip Kiely \u0026 Ali Taha: Long-context inference (200k tokens) creates peak power spikes; prefill/decode disaggregation reduces but doesn't eliminate thermal headroom pressure in datacenters.\nGreen Marbles: VRT"}