The past day moved AI governance and infrastructure in the same direction: from principles toward operating systems. Three leading labs are discussing a shared safety body, while Microsoft published detailed rules for how its own models should behave. At the same time, Google widened internal access to a rival model and Perplexity described giving GPT-6 Astra more control over production work. New funding for networking and inference chips shows that capital is still chasing the physical bottlenecks behind those capabilities.
Leading labs discuss a shared AI safety body
The Washington Post reported that leaders at Anthropic, OpenAI and Google have discussed creating a new AI safety body while endorsing limits on the pace of frontier development. The conversations are not yet a binding pact, and the companies have different commercial incentives. Still, coordination among direct rivals would mark a shift from company-specific policies toward shared testing and escalation practices.
The hardest questions are practical. A credible body would need independent evaluators, access to unreleased systems, common thresholds for dangerous capability and a process for responding when one member declines to slow down. Without those mechanisms, public commitments may improve transparency but do little to change deployment decisions.
Microsoft publishes conduct rules for MAI models
Microsoft AI opened a six-week public consultation on a draft Code of Conduct. It says MAI models should remain subordinate to people, never resist interruption or shutdown, avoid expanding their own scope and not conceal reasoning from auditors. The draft also includes absolute constraints for weapons of mass harm, child safety and manipulation at scale.
This is more specific than a broad promise to build responsible AI. Microsoft says the code will guide training and evaluation, giving outsiders language they can later compare with model behavior. The open issue is enforceability: rules become meaningful only when evaluations, incident reporting and release gates show whether a system follows them under pressure.
Google gives engineers access to Claude Opus 5
Business Insider reported that Google made Anthropic's Claude Opus 5 available to engineers across the company through Antigravity, its internal development platform. A Google spokesperson said Gemini remains the primary internal model and that third-party access is quota-limited for specialized uses.
The decision shows how coding performance can override ecosystem boundaries. Google is a major Anthropic investor, but it also competes directly with Claude through Gemini. Letting engineers choose a rival model creates an internal comparison at production scale and may pressure each vendor to prove gains in code quality, review burden and time saved rather than benchmark scores alone.
Perplexity puts GPT-6 Astra into end-to-end workflows
An OpenAI customer case study says Perplexity is using GPT-6 Astra to write communications, modify software and monitor production systems. Perplexity cofounder Johnny Ho said the model can build testing programs that simulate external services and check workflows from start to finish, allowing teams to check in less often than with earlier models.
The account comes from the model supplier, so it should be treated as a deployment example rather than independent performance evidence. Even so, it illustrates the next enterprise threshold: systems that do not only generate code, but also test changes and observe live operations. That expands the value of the model and the consequences of weak permissions or incomplete monitoring.
Cornelis raises $205 million for active AI networking
TechCrunch reported that Cornelis Networks raised $205 million and introduced Active Compute Fabric, an open networking architecture designed to let chips process and transmit information at the same time. The company says it can reduce GPU idle time caused by waiting for data and work across accelerators from different vendors.
AI clusters increasingly depend on the network as much as the processor. If Cornelis can improve utilization without locking customers to one chip supplier, buyers gain another way to lower the cost of training and inference. The company is already shipping its current product, but performance and interoperability claims will need validation across large, mixed-hardware deployments.
Euclyd funds a different route to inference
Euclyd announced a Series A of more than €200 million co-led by Samsung, Somerset Capital Partners, EQT's Scaleup Europe Fund and Innovation Industries. The Dutch startup is developing a processor and memory architecture for AI inference rather than another conventional GPU. It plans to sell rack systems for private deployments and license its designs to chipmakers.
The round is unusually large for a young semiconductor company, reflecting both the cost of bringing silicon to market and demand for alternatives to Nvidia. Euclyd is targeting initial physical systems in 2028, so the financing is a bet on execution over several years. Samsung's memory and supply-chain expertise may matter as much as its capital.
MediaTek brings 2nm silicon to on-device AI
MediaTek launched the Dimensity 9600 Pro, its first flagship phone system-on-chip built on a 2nm process. The company says its dual-NPU design improves large-model prefill performance and token generation per watt, supports models up to 30 billion parameters and cuts power use for always-on AI. Phones using the chip are expected this quarter.
Vendor figures still need device-level testing, but the direction is clear. More capable local inference can reduce cloud costs and keep sensitive data on the phone. It can also make agent features more responsive when connectivity is limited. The tradeoff will be how phone makers balance model size, battery life, heat and the permissions required for proactive software.
What matters today
AI development is becoming an institutional and infrastructure problem at the same time. Labs are exploring shared controls because internal policies may not manage competitive pressure. Microsoft is turning values into trainable rules, while Google and Perplexity are testing how much operational responsibility advanced models can carry. Cornelis, Euclyd and MediaTek are attacking bottlenecks in networks, inference and edge devices. The next gains will depend on whether governance, access controls and hardware efficiency advance together.