AI Briefing

Three frontier labs shipped models in the same week and every single one led with price rather than capability, which tells you where the fight has moved. Meanwhile Nvidia wrote a $12.9 billion check for the open-source community's town square, and the people building the data centers that power all of it just put $50 million behind ads because voters have started saying no. Cheap intelligence is here. The bill is landing somewhere else.

01IndustryNVIDIA Blog ↗

Nvidia Just Bought the Place Where Everyone Else's Models Live

Nvidia signed a definitive agreement to acquire Hugging Face for $12,930,300,000, announced by Jensen Huang on September 3. Hugging Face is where 18 million developers share 3 million models and 500,000 datasets, including models from every one of Nvidia's customers and rivals. Huang promised the platform stays open and that Nvidia compute will not be required to build or deploy through it, which is exactly the promise everyone will be watching. The deal is expected to close in the first half of 2027 pending regulatory approval.

02ProductAnthropic ↗

Anthropic's New Model Designed Protein Binders Ten Times Stronger Than the Best Human Entries in a Public Competition

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, the same underlying model shipped with two different safeguard levels. The headline is not the coding benchmarks but the science: given open-source protein design tools, Mythos 5.1 produced binders with affinities ten times higher than the best submissions to Adaptyv Bio's design competitions on three targets, with a hit rate near 50% where 10 to 15% is typical. It also trained a network that remapped a third of Venus at two to three kilometer resolution from 30-year-old Magellan radar data. Fable 5.1 costs about 25% less than its predecessor thanks to a 75% cut on cache reads.

03ProductLatent Space (AINews) ↗

Meta Will Sell You Frontier-Grade Inference at a 90% Discount If You Let It Train on Your Prompts

Meta rolled out Muse Spark 1.3 on September 2, and Artificial Analysis immediately ranked it the third best model in the world, putting Meta Superintelligence Labs back in the frontier conversation after a rough year. The pricing is the real story: the model is more than 90% cheaper if you opt in to letting Meta train on your data, an explicit price tag on the thing every other lab quietly asks for. Zuckerberg also said open weights are coming soon, and community benchmarks showed a 98.1% MRCR score at 512k to 1M context.

04ProductMicrosoft AI ↗

Microsoft Built a Transcription Model That Costs Ten Cents an Hour and Runs Ten Times Faster Than OpenAI's

Microsoft AI launched MAI-Transcribe-2 on September 3, claiming first place on the FLEURS benchmark across 60 languages with a 5.2% average word error rate. Independent evals from Artificial Analysis put it at 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs Scribe v2 and 5x faster than Gemini 3.5 Transcribe. The introductory price of $0.10 per hour of audio through year end is roughly 72% below Microsoft's own previous generation, and the model adds speaker diarization, word-level timestamps and code switching for blended languages like Hinglish and Spanglish.

05PolicyAxios ↗

The Data Center Backlash Got Big Enough That AI Money Is Now Buying Midterm Ads

Build American AI, a 501(c)(4) with about $50 million on hand and ties to the pro-AI super PAC Leading the Future, is pouring millions into Kansas, Ohio and Wisconsin to defend data center construction. Backers include Marc Andreessen, Ben Horowitz and OpenAI co-founder Greg Brockman. The trigger is polling: opposition to new data centers near where people live jumped from 49% to 61% in a matter of months, and Republicans have privately flagged Ohio as a real electoral risk. The group is also launching a new super PAC, Building the Future, to back friendly candidates and oppose the rest.