Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-09-05 · source-backed
Announced at IFA 2026 on September 4, the Threadripper Halo Station pairs a 96-core, 192-thread Threadripper PRO 9995WX (Zen 5, 5.4 GHz boost, 384MB L3, 8-channel DDR5, 128 PCIe 5.0 lanes) with two liquid-cooled Instinct MI350P cards at 144GB HBM3E and 4 TB/s each, upgradable to four for 576GB, plus up to 2TB system memory, drawing about 1,550W. AMD's pitch is trillion-parameter models with no cloud connection. No price and no ship date, which is exactly what decides whether it competes with the GB300 DGX Station (TechPowerUp).
Each link below shares sources, entities, or timing with this story.
AMD unveiled its first rack-scale system to directly contest Nvidia at the rack level, with engineering samples in H2 2026 and mass production targeted Q2 2027. Microsoft joins Meta, OpenAI and Oracle as customers; Meta plans 1 gigawatt of Helios racks by year-end against a lo...
Reuters, via Tech Startups, reports capital released against deployment milestones with Anthropic deploying up to two gigawatts of Instinct MI450 starting 2027. Same structure as Nvidia/OpenAI: compute vendor capital flowing to the lab that commits to buy the silicon. A two-gi...
Luu's September 1 post drew 852 points and over 1,000 comments, walking predictions from February 2024 through November 2025 after removing unfalsifiable and tautological ones. Misses include "AI has peaked" (Feb 2024), Meta "dying" (Nov 2024) against revenue going $135B to $2...
June local-inference benchmarks across the 128GB class put NVIDIA's DGX Spark (~$4k), AMD's Strix Halo / Ryzen AI Max+ 395 (~$2 to 3k), and the M5 Max 128GB (~$5k) head to head (Hardware Corner). Prompt processing favors CUDA hard. But token generation lands at a surprisingly...
PR #19378 landed in llama.cpp this week, and I think most people are underselling what it means. Backend-agnostic tensor parallelism via --split-mode tensor makes multi-GPU inference work across AMD, Intel, and Apple Silicon. Not just CUDA. Everything. For context: llama.cpp h...
Designed with Broadcom, built by TSMC, going into manufacturing after a six-week test run found no major issues (Reuters). Iris supplements rather than replaces Meta's Nvidia/AMD GPUs and supports a buildout targeting 7GW by end-2026 and 14GW in 2027 against up to $145B in 202...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.