Microsoft Wrote AI Safety Rules With One Non-Negotiable Line: If It Can't Be Shut Down, It Doesn't Ship

Microsoft says artificial intelligence needs a leash, and has written one into its own shipping rules. Samsung says it needs a cheaper chip. A lab in Shanghai, judging by its actions, feels it needs no introduction at all. Meanwhile in New Delhi, a country that imports nine out of every ten chips it uses spent Wednesday trying to change that math.

Microsoft’s rule for AI: if you can’t switch it off, you can’t ship it

On Monday, Microsoft’s AI division published the first draft of a document it calls the Humanist AI Code of Conduct, a 37 page rulebook for how the company’s future models are meant to behave. Mustafa Suleyman, who runs Microsoft AI, called it “a constitution of sorts” for what the company builds next.

The idea behind it fits in five words, which Suleyman used himself: people matter more than AI. From there the document lists a short set of things a model must never do, including helping build weapons capable of mass harm, threatening child safety, or manipulating people at scale. But the line getting the most attention is the plainest one. Every model has to stay interruptible, correctable, and able to be shut down. “If it isn’t, we don’t ship it,” the code states.

Suleyman first floated the idea of “humanist superintelligence” back in November 2025: the notion that the most powerful AI systems should be built with an off switch from day one, rather than have one bolted on after something goes wrong. This draft opens that idea up to public objection. Microsoft is running a six week comment period before it publishes a revised version later this year, one that is meant to actually guide how its models get built through 2027.

Samsung puts $230 million into a chip that isn’t Nvidia’s

While Microsoft was writing rules for how AI should behave, Samsung was placing a very different kind of bet on the same industry. The Korean company co-led a Series A round worth more than 200 million euros, about $230 million, in a Dutch startup called Euclyd, alongside Somerset Capital Partners, EQT’s Scaleup Europe Fund, and Innovation Industries.

Euclyd is not trying to build a faster version of Nvidia’s chips. It is building something structurally different: a chip and memory architecture aimed at inference, the stage where an already trained model answers a question, rather than training, where Nvidia’s grip on the market is closest to total. Inference is the one part of the business where analysts put Nvidia’s share at somewhere between 60 and 75 percent instead of near complete dominance, which is exactly why challengers keep aiming there first.

Founders Bernardo Kastrup and Atul Sinha started the company in 2024 at Eindhoven’s High Tech Campus. Former ASML chief executive Peter Wennink is taking the chairman’s seat, which reads like a signal on its own: a man who spent years running the company that builds the machines that make advanced chips now thinks the next opportunity is skipping the GPU model altogether. Euclyd says it plans to ship physical hardware in 2028 and serve thousands of enterprise customers by 2030, which in chip years counts as close but not around the corner.

A 744 billion parameter model showed up with no announcement at all

Compare that to what Shanghai Artificial Intelligence Laboratory did the previous Friday, which was nothing. On September 11, a repository named Atria Dawn Preview simply appeared on GitHub and Hugging Face, credited to the lab, with no blog post, no press release, and no pricing page attached. It is a 744 billion parameter model built on the GLM-5.2 foundation model, released under an open MIT license, built for agentic work: the kind of task where a model has to use tools, run code, and recover from its own mistakes across many steps rather than answer one question and stop.

Three days later, the lab finally said something, posting a technical report to the preprint server arXiv with more than 140 listed authors. That is the closest thing to a formal announcement the model has gotten. Its benchmark numbers so far come entirely from the lab’s own testing rather than independent evaluation, so how it actually holds up against paid frontier models from OpenAI or Anthropic is still an open question. What is not in question is that a fully open, commercially usable model at this size can now be downloaded by anyone with the hardware to run it, and nobody at the lab felt the need to make a show of that fact.

India tells the world’s chipmakers it is done being only a customer

None of that changes the fact that most of the world’s advanced chips are made somewhere else, which is the exact problem Prime Minister Narendra Modi spent Wednesday trying to chip away at. In New Delhi, he sat down for a roundtable with global semiconductor chief executives, one day ahead of SEMICON India 2026, a three day conference opening Thursday that is expected to draw more than 600 exhibitors, including 300 companies from outside India.

The country still imports close to 90 percent of the chips it uses, at a cost of roughly $30 billion a year. But the numbers behind Wednesday’s roundtable show that gap starting to close. Twelve semiconductor projects have been approved across six states under the government’s Semicon India programme, carrying investment commitments above 1.64 lakh crore rupees, and three of those plants have already begun commercial production. In July, the government approved a second phase of the programme, called Semicon 2.0, with a further outlay of 1.27 lakh crore rupees aimed at building out the rest of the supply chain beyond fabrication alone.

None of that makes India a chip superpower yet. It makes India a country that has moved from talking about building chips to actually shipping some, which for an industry this capital intensive and this slow to build counts as real progress rather than a talking point.