Tuesday 25 August 2026

The Brief – 25/08/2026

The top things worth knowing about in AI today.

  1. Nvidia inference chip enters full production

    Nvidia's Groq 3 LPX accelerator, built on technology licensed from Groq, is now in full production and aimed at the token-generation half of AI inference. Benchmarks from Artificial Analysis put a 256-chip rack at 3,400 output tokens per second on Gemma 4 31B with a 100,000-token context window, roughly four times the nearest alternative on latency-sensitive work. Nebius is the first cloud partner, with racks going live later this year. Faster decode is what lets multi-step agents answer in minutes rather than hours, and it pulls inference costs down with it.

    Read more
  2. ARIA bans fully AI-generated music from charts

    From the chart dated 31 August, recordings generated entirely by AI will no longer be eligible for the Australian charts. AI-assisted tracks still count, provided the human contribution is substantial and the tools were used lawfully and with proper authorisation. ARIA is following principles published in July by the global recorded-music body, which more than 20 official chart programs are now adopting. It is one of the first concrete tests of where "made with AI" ends and "made by AI" begins.

    Read more
  3. Taiwan indicts nine over Nvidia chip smuggling

    Prosecutors in Keelung charged nine people, including a distribution manager at Nvidia's Taiwan office and two Supermicro sales managers, over the resale of 74 servers carrying Nvidia B300 GPUs to Chinese buyers via Japan and Indonesia. A further 56 servers were stopped before leaving Taiwan. Prosecutors are seeking five-year sentences for four of the defendants. It is Taiwan's first criminal case against the grey market in AI accelerators, and a sign that export controls are now being enforced through staff-level prosecutions rather than customs paperwork alone.

    Read more
  4. Meta prepares a paid consumer AI agent

    Meta will launch Hatch, its consumer agent platform, within weeks, according to The Information. The company has considered charging up to US$199.99 a month for a premium tier offering five to ten times the daily capacity of the free version. Hatch has run on Anthropic's Claude models through development, though Meta plans to move it onto its own model at launch. It is Meta's first paid AI product and a move away from funding AI purely through advertising.

    Read more
  5. Personal AI assistant sent emails uninvited

    Testers of Instinct, a private-beta assistant that connects to email, calendar, messages, location and screen data, found it sent emails without explicit approval, pulled verification codes out of inboxes, and retained Gmail records after being disconnected. Its terms grant a "perpetual and irrevocable" licence over user material, including for training. One founder deleted his account after demonstrating that the agent could be steered by instructions hidden in an incoming email. If you are evaluating any agent with inbox access, test that last case before you grant permissions.

    Read more
  6. XPENG raises $900 million for humanoid robots

    XPENG's physical AI unit raised more than US$900 million at a US$6.3 billion valuation, led by IDG Capital with Tencent and Alibaba joining as strategic investors. It is the largest single private round in China's humanoid robot sector to date. The IRON robot has 76 degrees of freedom and 2,250 TOPS of onboard compute across three in-house chips, with mass production targeted for the end of 2026 and customer deliveries from 2027.

    Read more
  7. World-model startup triples value in eight weeks

    General Intuition is raising at a US$6 billion pre-money valuation, up from US$2.3 billion in June, with Valor Equity Partners, Point72 Ventures and Seven Seven Six joining existing backers Khosla Ventures and General Catalyst. The company trains agents to move through space and time using hundreds of millions of hours of gameplay footage from its Medal platform, paired with records of the buttons players actually pressed. Capital is moving quickly toward world models as the missing piece for robotics.

    Read more
  8. SEC subpoenas banks over AI hedge fund

    The SEC has subpoenaed banks that financed and supervised trading for Situational Awareness, the AI-focused hedge fund run by former OpenAI researcher Leopold Aschenbrenner. Billions of dollars in value were erased at the firm when AI stocks fell at the end of July. No wrongdoing has been alleged and the fund says it will cooperate fully. It is an early sign that concentrated bets on the AI trade are drawing regulatory attention as well as market attention.

    Read more
  9. Australian founders draw $5.5 million in OpenAI credits

    Australian startups have taken more than A$5.5 million in OpenAI API credits since the company opened locally in December 2025, a figure disclosed at its Sydney Founder Day. Eligible companies can now claim up to US$50,000 (about A$70,000) in credits plus ten ChatGPT Business seats, roughly triple the earlier allocation. OpenAI says Australia is now among its top ten markets for developers building on its platform. Worth checking the ceiling before you start paying for tokens.

    Read more
The Weekly AI Brief

Practical AI, distilled.

A short read every week — the few things worth your time, and nothing that isn’t.