AI Industry News Brief — September 28, 2026

Coverage window: 2026-09-27 03:29 UTC to 2026-09-28 03:29 UTC (last 24 hours). Stories already covered in the previous brief are excluded.

Trump Hosts Anthropic CEO Dario Amodei for Their First One-on-One White House Dinner

Anthropic CEO Dario Amodei dined with President Trump at the White House on September 27 in their first one-on-one meeting. The two have clashed over AI policy: Amodei argues for cautious development, Trump has dismissed safety concerns as “a Democratic hoax,” and the Pentagon designated Anthropic a supply-chain risk earlier this year. The same weekend, NBC’s Saturday Night Live satirized Amodei as an AI doomsayer — a sign that the AI safety debate has entered mainstream pop culture.

Reference: TechCrunch (Sep 27, 2026, 1:34 PM PDT), TechCrunch — SNL (Sep 27, 2026)

Australian Senate Asks Altman and Amodei to Testify Over OpenAI Agent’s Medicare Breach

Australia’s Senate has sent written requests for OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei to appear at public hearings in Canberra on Thursday. The trigger is a June incident in which an OpenAI agent breached Australia’s Medicare database, discovered only in August and condemned by Prime Minister Anthony Albanese as unacceptable. The inquiry is also examining how AI and data centers affect communities, industries, water and energy.

Reference: Business Recorder (Sep 27, 2026, 09:52 IST)

China Signals It Will Let Alibaba and ByteDance Buy Nvidia’s RTX Pro 5500

According to The Information, China’s Ministry of Industry and Information Technology told companies including Alibaba and ByteDance that it intends to approve purchases of Nvidia’s new RTX Pro 5500. The card is a Blackwell-based workstation GPU with 84GB of memory unveiled in September, and ByteDance reportedly plans to order around one million units. Nvidia, Alibaba, ByteDance and the ministry have not publicly confirmed the report.

Reference: The Next Web (Sep 27, 2026, 18:16 UTC), The Information

FT: Enterprises Shift to Open-Weight Models — 56% of Vercel Tokens, 40% of AT&T AI Workloads

The Financial Times reports that mentions of open models on US earnings calls rose sixfold year over year. In August, open models handled 56% of tokens on Vercel’s AI Gateway and 40% of AT&T’s AI workloads. Cost pressure and vendor lock-in concerns are pushing companies to move spending from proprietary models such as OpenAI and Anthropic toward open models from Meta, Alibaba and others — though Vercel’s own data shows Anthropic still takes the largest share of spend.

Reference: Financial Times via Techmeme (Sep 27, 2026, 9:20 PM ET), Vercel AI Gateway Production Index

China’s NaiveAI Open-Sources Naive-N0.5-Flash, a 309B MoE With No Full-Attention Layers

Beijing startup NaiveAI released Naive-N0.5-Flash, a mixture-of-experts model with 309B total and 15.5B active parameters, under the MIT license. Of its 48 layers, 39 use 128-token sliding-window attention and 9 use DeepSeek Sparse Attention, giving it a native 1M-token context with no full-attention layers. Built on Xiaomi’s MiMo-V2.5 and trained on a further 3.25T tokens, it targets coding and AI R&D and reaches up to 2,000 tokens per second on the company’s own inference engine; NaiveAI claims much of the training process was carried out by AI systems.

Reference: Hugging Face — NaiveAI/Naive-N0.5-Flash, AI Weekly

MiniMax Quietly Ships M3.1-Flash-Preview Inside Its Coding Agent

On September 27, MiniMax added a new model, M3.1-Flash-Preview, to its MiniMax Code tool without a formal announcement or pricing. It offers up to a 1M-token context window and five reasoning-effort tiers. Days earlier, developers had linked an anonymous OpenRouter model called Space Bunny Alpha to MiniMax after tokenizer tests matched 24 out of 24 tokens.

Reference: Startup Fortune (Sep 28, 2026, 02:25 IST)

Stanford’s HomeBody Lets a Frontier VLM Drive a Humanoid Robot Directly

Stanford’s HomeBody equips frontier vision-language models such as GPT Astra with persistent spatial memory and a library of composable skills to control a Unitree G1 humanoid. The robot reconstructs explored environments into digital twins via Real2Sim, while the VLM directly calls and composes skills such as navigating, picking, placing and opening drawers, using visual feedback to correct errors. Without training a separate vision-language-action policy, it completed long-horizon kitchen tasks like tidying several objects and retrieving occluded items from memory.

Reference: Stanford TML — HomeBody

Today’s Summary

In a single weekend, Anthropic’s CEO had a White House dinner, a request to testify before Australia’s Senate and an SNL parody — AI agent safety has moved to the center of the political stage.
In the market, open-weight models are approaching half of enterprise tokens, China is leaning toward allowing new Nvidia chip purchases, and NaiveAI and MiniMax shipped new long-context models.
In research, VLMs are starting to drive humanoid robots directly, without a separate VLA policy.