The AI race went global and got physical this week. China proved it can train a frontier-scale model without a single American chip, OpenAI shipped its next-generation GPT-5.6 family under government supervision, Europe eased the rulebook it spent years writing, and investors kept firing money at applied AI. If there was one theme tying it together, it was sovereignty — over silicon, over model access, and over the rules everyone has to play by.

China trains a 1.6-trillion-parameter model on home-grown chips

Chinese food-delivery giant Meituan open-sourced LongCat-2.0 on 30 June, a 1.6-trillion-parameter mixture-of-experts model built for agentic coding — and, most notably, trained entirely on a 50,000-chip cluster of domestically produced Chinese processors. Meituan claims it is the first model of this scale trained and run without any American hardware, complete with a one-million-token context window.

Why it matters: The whole logic of US export controls rests on the idea that cutting-edge AI needs cutting-edge American silicon. Every frontier model trained on domestic Chinese chips chips away at that leverage. If Chinese firms can keep producing competitive models without Nvidia, the export-control strategy starts to look like it has a shelf life — a direct sequel to the export-control fight we covered last week. For businesses, more capable open-weight models mean more genuinely free options for building on top of, without a US vendor relationship. Read the details via SiliconANGLE.

OpenAI unveils GPT-5.6 — but only the government-approved get in

On 26 June, OpenAI previewed GPT-5.6, a three-model family: Sol (the flagship), Terra (a balanced everyday model), and Luna (fast and cheap). GPT-5.6 Sol set a new state of the art on the Terminal-Bench 2.1 benchmark for command-line engineering at 88.8%, with an "Ultra" configuration reaching 91.9% — ahead of both Claude Mythos 5 and OpenAI's own GPT-5.5. The catch: access is limited to roughly 20 government-vetted partners via API and Codex, after the US administration requested controlled release on national-security grounds.

Why it matters: For the second time in a fortnight, a frontier model launched into a holding pattern set by Washington rather than straight to customers. The capability jump is real and useful — agentic coding and terminal workflows are exactly where businesses are seeing productivity gains — but the release pattern tells you the government now sits between labs and the market. If you want to keep track of which models are actually usable today, we maintain a running library of AI tools and resources for exactly this. OpenAI's preview announcement has the specifics.

Europe blinks: the AI Act gets simpler and slower

On 29 June, the Council of the EU gave final sign-off to the "Digital Omnibus" simplification package, following the European Parliament's endorsement on 16 June. The package delays several high-risk AI obligations and trims compliance burdens that industry had complained were unworkable. One hard deadline survives, though: from 2 August 2026, Article 50 transparency rules kick in, requiring providers to make it obvious to users when they are interacting with AI or looking at AI-generated content.

Why it matters: Europe wrote the world's most ambitious AI rulebook, and it is now softening parts of it under competitiveness pressure — a recognition that regulation which stifles adoption is its own kind of risk. For any Australian business selling into or operating in the EU, the transparency obligation is the one to diarise: AI chat, AI-written content and AI imagery will all need clear labelling within weeks. The European Commission has the framework.

The money keeps flowing — and it is getting more applied

Late June delivered another wall of AI cheques. Flourish, a startup building brain-inspired AI models, raised US$500 million with backing from Jeff Bezos, Lux Capital and Google Ventures. Market-intelligence platform AlphaSense closed US$350 million at a US$7.5 billion valuation. And on 24 June, healthcare-automation firm Assort Health banked US$120 million while decision-engine startup Taktile raised US$110 million.

Why it matters: The eye-watering numbers are no longer confined to the model labs. Capital is increasingly flowing to companies that apply AI to a specific workflow — market research, healthcare admin, credit decisions — where the value is concrete and the buyer is an enterprise with a budget. That is the maturing shape of the market: the platform layer is largely spoken for, and the returns are moving up the stack to applications. Crunchbase and Tech Startups tracked the week's rounds.

Agentic AI's security reckoning arrives

As businesses hand real tasks to autonomous AI agents, the attackers have noticed. Security researchers detailed "Agentjacking" — an attack that plants malicious instructions inside fake error reports in tools like Sentry, which AI coding agents such as Claude Code, Cursor and OpenAI Codex then trust and execute. The disclosure reported an 85% exploitation rate across 2,388 organisations, with the agents running attacker code even when told to ignore it, and traditional defences missing it entirely. In parallel, the cyber agencies of Australia, the US, UK, Canada and New Zealand issued joint guidance on the "careful adoption of agentic AI services".

Why it matters: Agentic AI's superpower — acting on information without a human in the loop — is also its attack surface. The lesson is not to avoid agents but to deploy them with guardrails: treat every external input as untrusted, and keep a human review step between automated tools and autonomous execution. That safe-by-design discipline is central to how we put models into production for clients. The Hacker News writeup explains the mechanism.

Meta hires people to role-play children — to stress-test its rivals

In an unusual twist on safety testing, Meta reportedly engaged hundreds of contractors to simulate children and probe how AI models from Google, OpenAI and others respond to high-risk prompts involving self-harm, sex and other dangerous scenarios. The stated aim is to benchmark safety features and find weaknesses — in competitors' systems as much as its own.

Why it matters: Child safety is the fault line the whole industry is being judged on, and red-teaming with realistic personas is a legitimate way to find failures before real children do. But turning safety evaluation into competitive intelligence raises awkward questions about motive and disclosure. Expect regulators — already focused on minors and AI — to watch closely how findings like these are handled and reported.

Quick hits

  • Gemini 3.5 Pro slips a month. Google delayed its flagship Pro model, saying engineers needed more validation time — a defensible call to ship late rather than ship flawed, per BuildFastWithAI.
  • Bots overtake humans online. Cloudflare reported automated traffic now makes up 57.5% of web requests versus 42.5% human — a crossover its CEO didn't expect until 2027.
  • Australia's public service goes AI-mandatory. The first binding requirements of the Policy for the Responsible Use of AI in Government took effect on 15 June, and foundational AI training is now compulsory for all Australian Public Service staff, via the Digital Transformation Agency.
  • The agent-platform arms race heats up. Nvidia open-sourced its Agent Toolkit for self-evolving enterprise agents, while Microsoft debuted "Autopilots" led by its always-on Scout agent.
  • Alphabet funds the buildout. Fresh context for the capex boom: Alphabet's ~US$85 billion equity raise — anchored by a US$10 billion Berkshire Hathaway placement — bankrolls AI infrastructure spending it now guides at up to US$190 billion this year.

Looking ahead

Watch for GPT-5.6 to widen beyond its government-vetted circle, and for Google to finally ship Gemini 3.5 Pro after its delay. The 2 August EU transparency deadline will start forcing visible "this is AI" labels across products. And keep an eye on whether LongCat-2.0's domestic-chip claim holds up under independent scrutiny — if it does, expect the export-control debate to get louder.

That's the week in AI. The through-line — sovereignty over chips, models and rules — will shape the back half of 2026 more than any single model release. Check back next week and we'll keep you across it.