China AI Chip Evasion: Nvidia CEO Targeted, DeepSeek Used OpenAI Outputs
Two major China-AI stories break simultaneously: U.S. intel suspected a Singapore-based exec of helping China evade AI chip controls near Nvidia's CEO, and DeepSeek/Moonshot reportedly distilled OpenAI's GPT-6 Astra to train their latest models.
What we know
Per the NYT, a suspected spy helped China evade AI chip export constraints; an account on X IDs U.S. intelligence as focused on Singapore-based exec 'Huang Le'; another account on X adds Nvidia CEO Jensen Huang was a 'potential target' of the suspected Chinese intel op, though Nvidia says the CEO interactions were not unusual. Separately, another account on X cites The Information reporting that both DeepSeek (V4.1-Flash update) and Moonshot AI (Kimi K3.1 Pro) relied heavily on distillation from OpenAI's GPT-6 Astra — raising fresh questions about IP and the U.S.-China AI competition.
How the story developed
Two major China-AI stories break simultaneously: U.S. intel suspected a Singapore-based exec of helping China evade AI chip controls near Nvidia's CEO, and DeepSeek/Moonshot reportedly distilled OpenAI's GPT-6 Astra to train their latest models.
Google released Gemini 4 Carbon, its new frontier model, just two weeks after Argon debuted, with benchmarks reportedly surpassing GPT-6 Astra and Claude Opus 5.5 across most metrics.
Anthropic secretly acquired a one-person biotech company ('Professor Oak') after 6 years of anonymity, while separate data shows GPT-5.5 outscores all tested Chinese AI models — with OpenAI and Anthropic reportedly still holding back their strongest models.
Researchers announced three new discoveries in harmonic analysis, complexity theory, and quantum field theory; OpenAI's Greg Brockman quickly noted P probably still doesn't equal NP.
OpenAI released GPT-6.1 Terra with a focus on computer use, with Greg Brockman declaring it crosses a qualitative threshold where it feels faster than an unassisted human at operating a computer.
OpenAI released GPT-6.1 Terra, offering 3x faster inference at 1/10th the cost with a 1.05M token context window and computer-use capabilities — and immediately classified it as 'critical' in cybersecurity and 'high' for biological and chemical capability under its Preparedness Framework.
Prominent builder Georgios Konstantopoulos says he hasn't intentionally used an OpenAI model since Claude Opus 5.5 launched, declaring Anthropic won the current competitive round.
Revelations that Anthropic's Claude Mythos and OpenAI's Daybreak Red were used to scan PHP for vulnerabilities — yet an outside actor using GPT-5.6 still won — have reignited debate over whether restricting frontier AI access actually improves national security.
OpenAI's GPT-Rosalind surfaced a novel drug discovery lead for a rare neurodegenerative disease, while the broader debate over AI's net societal costs intensifies.
By the Global Eye 24 desk, with AI assistance. Updated as the story develops. Spot an error? Write to corrections@globaleye24.com.