NVIDIA AI Updates: October 3, 2026

1. DGX Spark Gets a 64GB Configuration Starting at $4,999

NVIDIA DGX Spark. NVIDIA announced a 64GB unified-memory configuration of DGX Spark, sold only through Acer, ASUS, Dell, Gigabyte, HP and MSI starting October 23 at $4,999. It keeps the GB10 Grace Blackwell Superchip, DGX OS, ConnectX-7 networking and the full NVIDIA AI software stack of the 128GB model, and supports models up to 100 billion parameters on device. Two units connected by a QSFP cable pool memory to 128GB for models up to 200 billion parameters, configured through the new NVIDIA Sync Cluster Assistant, and NVIDIA reports up to 1.7x the performance of a single unit on Qwen 3.8 27B. An NVIDIA Sync Model Launcher arriving at the end of October will download and launch Qwen3.8 27B on one unit or a cluster and set up OpenCode to use it. Source

2. OpenAI’s GPT-6 Astra Ultrafast Runs on Blackwell With Up to 8x Faster Generation

NVIDIA. NVIDIA detailed how GPT-6 Astra Ultrafast, now available in the OpenAI API and to eligible ChatGPT Work and Codex users, runs on Blackwell GPUs and delivers up to 8x faster token generation than Astra Standard mode. OpenAI said it used its own internal models to optimize inference software on NVIDIA GPUs, and OpenAI inference lead Philippe Tillet said Astra can write high-performance kernels for Blackwell and Rubin GPUs. NVIDIA positions the speedup for coding agents’ edit-test-debug cycles and tool-call loops. Source

3. NVIDIA Lays Out AI Factory Economics, Citing 30x Throughput per Megawatt for Vera Rubin NVL72

NVIDIA. NVIDIA published a framework for AI factory return on investment built on earning capacity, useful life and demand, noting that each megawatt of AI factory capacity costs roughly $60 million. Citing SemiAnalysis AgentX data, it states Vera Rubin NVL72 delivers over 30x higher throughput per megawatt than GB300 NVL72 and up to 45x lower cost per million tokens on DeepSeek V4 Pro. On hardware lifespan, it cites CoreWeave extending A100 bookings through 2029, Barkr estimates of five to six years of useful life for an eight-GPU H100 system and nine to 10 years for GB300 NVL72, and Silicon Data figures showing a six-year-old A100 still worth a quarter of its original cost. Source