OpenAI rewrites AI infrastructure playbook

The closed loop has Nvidia taking non-voting equity while OpenAI commits spend back into 4–5 million GPUs, targeting 10 gigawatts of compute. CEO Jensen Huang insists this won’t squeeze supply for other customers, though the dimensions could redefine how hyperscalers finance AI infrastructure.

Alibaba is on an aggressive run of AI rollouts this week. Qwen3-Next is a faster MoE architecture that expands to 512 experts while activating only 10 plus a shared expert per step. With stability fixes like normalized router initialization and a spotlight gating, it delivers over 10x the throughput of Qwen3-32B on long sequences, handling as much as 256K tokens natively, with experimental paths to 1M. Two 80B variants lead the road: Instruct for assistants and Considering for reasoning, with FP8 releases cutting latency and energy overhead.

Qwen3-Omni pushes into full multimodality with 30B parameters and just 3B lively per inference. Its split Thinker-Talker system enables streaming speech generation with sub-second response, benchmark wins across 32 of 36 audio/video tasks, and support for 119 written and 19 spoken languages. Open-source Instruct, Considering, and Captioner variants extend its reach.

On raw scale, Qwen3-Max pushes past 1T parameters and 36T tokens, with ChunkFlow boosting 1M-token context training. Benchmarks put Instruct in the worldwide top three, beating Claude Opus 4 and DeepSeek V3.1 on coding and agent tasks. The Considering variant, still in training, has already hit perfect scores on AIME 25 and HMMT with test-time scaling and code execution.

Qwen3-LiveTranslate-Flash makes all this tangible by bringing 18-language real-time interpretation at 3s latency, integrating lip-reading, gesture recognition, and semantic unit prediction for near-offline quality. It edges out GPT-4o-Audio-Preview and Gemini-2.5-Flash on speech tasks, while producing expressive dialect-specific voices. Read more.

Google’s upgraded Gemini Live API now runs on a native audio model built for real-time reliability. Function calls, the pipes that allow agents pull live data or execute services, at the moment are as much as 2x more accurate, even in messy multi-function scenarios. Add tighter audio handling and conversations flow like they need to: pauses, side chatter, and interruptions now not break the thread. Next up is “considering mode,” where developers set a reasoning budget, trading speed for depth with transparent traces of the model’s process.

On the buyer side, Google is flexing the identical tech. A brand new Gemini overlay in Google Play interprets on-screen context so gamers can ask for hints without breaking flow. A redesigned “You” tab turns the shop into a customized hub for progress, rewards, and cross-app recommendations.