Europe forces Android open as nations scale physical AI

Europe opens Android to rival AI

Rival AI assistants could eventually operate inside Android with the same kinds of access as Gemini. Europe has now made that a binding requirement, not a voluntary promise.

On 16 July, the European Commission issued specifications under the Digital Markets Act requiring Google to give competing AI services equal access to relevant Android capabilities. The Commission also ordered access to Google Search data for eligible third-party search engines, with privacy protections attached.

The important shift is where competition can happen. An assistant that can interact properly with operating-system features has a much stronger position than one confined to a standalone app. That could let other providers build deeper voice, search and task-completion experiences on Android devices in the European Union.

Nothing changes overnight. Google first has to implement the specifications, and the rules apply to its designated services in the EU rather than Android everywhere. There’s also a difficult privacy problem underneath the policy: useful device and search access can expose extremely sensitive information if the technical boundaries are weak.

For AI developers, the opportunity is potentially substantial, but access on paper won’t be enough. My read is that Europe’s intervention succeeds only if rivals receive capabilities that are genuinely usable while Google can still prevent personal device and search data leaking across services.

Japan builds for physical AI

That regulatory push is about who gets access to existing platforms. Japan, meanwhile, is building an enormous new one for machines that operate in the physical world.

The METI-backed FRONTia project plans a 140-megawatt AI factory operated by Noetra, equipped with 13,750 NVIDIA Vera CPUs and 27,500 Rubin GPUs. Its intended workloads include open multimodal foundation models for robotics, industrial agents and digital twins — virtual representations of physical systems used for simulation and planning. NVIDIA calls it the first national infrastructure dedicated specifically to physical AI, although that description is the company’s claim.

A day earlier, NVIDIA also introduced Cosmos 3 Edge, a four-billion-parameter world-action model intended to perform visual reasoning and generate robot policies locally across Jetson, RTX and DGX systems. That could reduce latency and dependence on constant cloud connectivity near factory equipment. There’s no independent performance or safety evaluation in the briefing, and participation from Japanese manufacturers is still described as planned coalition activity.

The industrial implication is bigger than either announcement alone: Japan is treating robotics-model capacity as national infrastructure. That may accelerate automation and sovereign model development, but it also concentrates a major programme — from central training to edge deployment — around NVIDIA’s hardware and software stack.

Web evidence agents can keep

For enterprise agents, one quieter change could make web research much easier to reproduce.

Parallel Web Search is now a native grounding provider in the Gemini Enterprise Agent Platform. Through the Gemini API or Agent Studio, it can retrieve current public-web information with citations. Developers are allowed to extract, post-process and permanently cache the returned data, including for use with models outside Gemini.

That last part is unusually consequential. Grounding results often disappear into a model response or have to be fetched again later. Persistent retrieval lets a compliance, research or due-diligence system retain the evidence behind an answer, inspect it, and potentially reproduce the workflow without being locked to the model that first consumed it.

There are costs and governance questions. Parallel Web Search is separately metered through Google Cloud. Zero-data retention is available, but the announcement doesn’t describe it as the default. Permanent caching also means organisations need policies for copyright, freshness, deletion and access control; a cited webpage can change or become inappropriate to retain.

For enterprise teams, retrieval is becoming part of the governed data estate rather than a temporary prompt ingredient. Reproducibility should improve, but only if retained evidence has an owner, a lawful purpose and an expiry strategy.

Nokia puts GPUs into the radio network

Now to telecoms, where Nokia wants AI computing much closer to the mobile network itself.

Its newly announced AI-native radio-access platform combines Nokia’s anyRAN software with NVIDIA Aerial accelerated computing. Operators would be able to deploy it through upgrade cards, standalone GPU-powered nodes or cloud-native servers. The goal is to use AI processing to extract more capacity from existing spectrum and radio infrastructure.

Nokia reports spectral-efficiency improvements above 20 per cent and says it is targeting gains above 100 per cent by 2028. Those are company claims, and customers can’t buy the finished platform yet. Pilots are planned for late 2026, with commercial availability targeted for 2027.

The economic case will depend on the full system, not the headline percentage. Spectrum is scarce and expensive, so even a credible capacity improvement could defer new towers or spectrum purchases. But GPUs bring their own capital cost, power draw and operational complexity.

For network operators, this looks worth testing rather than budgeting around just yet. The deciding measure will be usable capacity gained per dollar and per watt under real traffic, not a laboratory efficiency result.

Copilot connects to private models

Developers using JetBrains IDEs now have more control over what sits behind GitHub Copilot.

GitHub has expanded Copilot’s bring-your-own-key support in JetBrains to include OpenAI-compatible custom endpoints. An engineering organisation can therefore keep the familiar Copilot workflow while routing requests to an alternative or privately hosted model under its own controls. The release also adds custom agents, skills and instructions for Claude providers, along with local sandboxing and a debugger skill for agent-assisted diagnosis.

This announcement is specific to the JetBrains integration; it doesn’t establish equivalent support across every Copilot surface. And although a local sandbox can limit what agent-executed code touches, it doesn’t prove the generated code or the execution itself is safe.

For teams constrained by privacy, data residency or internal model policy, private endpoints could remove a meaningful obstacle to adopting agentic coding tools. The trade-off is a transfer of responsibility. GitHub supplies the interface, but customers must secure the endpoint, select a capable model, manage credentials and verify that their configuration behaves as intended.

In practice, this makes Copilot more adaptable for regulated engineering environments — and makes platform teams more accountable for the quality and security behind the button.

An inventory for Kubernetes AI

Before an organisation can govern its AI systems, it has to know where they’re running. Google has released a tool aimed at that basic discovery problem.

The open-source k8s-aibom project is an unprivileged Kubernetes controller that detects deployed inference runtimes, agent frameworks, vector stores and training systems. It produces deterministic machine-learning bills of materials using the CycloneDX 1.6 format. It doesn’t require sidecars, kernel modules or changes to individual workloads, which should make adoption less intrusive across existing clusters.

This is especially relevant to shadow AI: components deployed by teams without a central security or platform group having a complete record. A machine-readable inventory can support audits and expose systems that otherwise blend into a large Kubernetes environment.

Its limit is equally important. Detecting a model server or vector database doesn’t establish whether it is secure, approved or compliant. An inventory is evidence of apparent deployment, not evidence of good governance.

For security and platform teams, AI bills of materials are likely to become the counterpart to software dependency inventories. The tool lowers the cost of discovery; the organisational work begins when each discovered component needs an owner, risk classification and remediation path.

Intel adds European manufacturing capacity

At the hardware level, Intel is putting another €5 billion behind European chip production.

The company plans to upgrade its Leixlip campus in Ireland and install additional advanced manufacturing equipment. The intended capacity will supply Xeon 6 processors and future Xeon chips made with Intel 3 technology for AI and high-performance computing workloads.

For European governments and infrastructure buyers, the appeal is geographic diversification. More data-centre processor capacity within Europe could reduce some exposure to manufacturing concentrated in other regions and give customers another source for critical compute hardware.

But this is an investment commitment, not completed capacity. Intel’s announcement doesn’t provide production milestones in the briefing, and the strategic benefit depends on two things it still has to deliver: competitive processors and a functioning expansion that reaches meaningful volume.

So the near-term consequence is confidence-building rather than extra chips arriving immediately. For buyers planning long-lived infrastructure, Leixlip could eventually improve supply-chain resilience, but it would be premature to treat the announced spending as available supply until Intel publishes and meets concrete production dates.

What Changes for You

One change is already closer to everyday use, particularly for families with teenagers using ChatGPT.

OpenAI says parents with linked teen accounts can now make Study Mode the default for new conversations. The company is also increasing reminders to take breaks after prolonged use and expanding parental notifications when a linked teen account is deactivated for violations involving violent threats or online violence. OpenAI reports interactive learning coverage across more than 250 maths and science topics, plus pronunciation support for over 61 languages.

For a family using linked accounts, ChatGPT can now begin in a more structured learning mode instead of a standard conversation, while parents receive an additional signal in a narrow category of serious enforcement. Schools may also find the default useful when discussing how students use the service for guided learning rather than simple answer generation.

The controls depend on OpenAI identifying the user as under 18 and on the parent and teen accounts being linked. They won’t detect every inappropriate response or risky interaction. The practical improvement is stronger default behaviour and oversight, not a substitute for adult judgment or school safeguards.