NVIDIA and Intel Unveil Specialized AI Hardware Amid Proposed Data Center Pause
Large Language Models & Efficient Inference
NVIDIA Releases gpt-oss-puzzle-88B: NVIDIA has launched gpt-oss-puzzle-88B, a deployment-optimized model derived from OpenAI's gpt-oss-120b using the "Puzzle" post-training neural architecture search framework. The model is specifically designed to maximize inference efficiency on H100 hardware for reasoning-heavy workloads without sacrificing accuracy.
Liquid AI Demonstrates High-Speed LLM Inference via WebGPU: Liquid AI's LFM2-24B-A2B, a Mixture-of-Experts (MoE) model, has been shown running at 50 tokens per second directly within a web browser using WebGPU. This performance, demonstrated on an M4 Max device, highlights the growing feasibility of running complex AI models locally in web environments using the ONNX format.
AI Hardware
Intel to Launch Affordable 32GB VRAM GPU for AI Workstations: Intel is set to release the Arc Pro B70 and B65 GPUs on March 31, offering 32GB of VRAM for a competitive price of $949. This hardware release aims to support local AI applications and larger models, such as Qwen 3.5, which require significant memory for efficient local performance.
AI Video Generation
The Emergence of Sora Alternatives in 2026: Following the reported discontinuation of OpenAI's Sora, several alternative video generation models like Kling3.0, Seedance2.0, and Veo 3.1 have been identified as the new industry leaders. Kling3.0 is currently highlighted as the premier replacement due to its optimal balance of quality, pricing, and API accessibility.
Policy & Regulation
Proposed Legislation to Halt AI Data Center Expansion: Senator Bernie Sanders has introduced a bill seeking a pause on AI data center construction and a ban on chip exports, citing risks to jobs and societal stability. This legislative move contrasts with recent executive decisions to lift export bans on high-end NVIDIA chips like the H200 to international markets.
Research & Benchmarks
ARC AGI-3 Benchmark Released to Measure Fluid Intelligence: The third iteration of the ARC AGI benchmark has been released, focusing on measuring how efficiently AI can acquire new skills compared to human cognitive abilities. Current top-tier models are scoring only 0.3%, emphasizing the significant progress still required for AI to match human-level fluid intelligence.