ZettaFLOPS of new AI compute coming online from Oracle late next year (www.theregister.com)

Oracle announced it will bring more than 18 zettaFLOPS of AI compute online by late next year, composed of an 800,000‑GPU Nvidia Blackwell cluster (up to ~16 zettaFLOPS peak, reported in sparse FP4) and an initial 50,000‑GPU AMD Instinct MI450X deployment (~2+ zettaFLOPS ultra‑low precision). Nvidia is supplying GPUs, racks and Spectrum‑X Ethernet networking for OCI’s Zettascale10 offering and will layer cloud AI services on top. AMD’s MI450X will appear in “Helios” rack‑scale systems (72 GPUs per rack) using the open Ultra Accelerator Link (UALink) and the new Open Rack Wide form factor; AMD claims ~2.9 exaFLOPS FP4 / 1.4 exaFLOPS FP8 per Helios rack plus 31 TB HBM4 (1.4 PB/s) bandwidth. Technically this is a major capacity buildout that accelerates the cloud-scale arms race and gives customers unprecedented access to ultra‑low precision throughput, but practical limits remain: FP4 is currently most useful for inference and sparse workloads, and most large training jobs still prefer BF16/FP8 or denser formats. Research (including Nvidia’s NVFP4 work) suggests 4‑bit pretraining may be viable, yet few customers can realistically lock entire clusters to exploit full zettaFLOPS. The deployment also highlights vendor dynamics—Nvidia dominance in scale networking, AMD’s growing footprint via open interconnects and potential ties to OpenAI—and raises questions around power, cost and how software/tooling will evolve to use these exascale resources.