AI News Digest: 25 March 2026

25 March 2026

Quick Read: Today's highlights cover the key AI developments from 25 March 2026, including the most important stories for UK businesses and decision-makers.

OpenAI is cleaning house ahead of its IPO, Arm has made a chip that could reshape the data centre industry, and a major supply chain attack just hit one of the most popular AI libraries. Meanwhile, the battle lines between AI companies and government are hardening on both sides of the Atlantic.

OpenAI Kills Sora, Consolidates for IPO

OpenAI has shut down Sora, its AI video generation app, just six months after launch. The standalone app and developer API are both being retired as the company consolidates its product line ahead of a planned IPO. CFO Sarah Friar told CNBC the company needs to be "ready to be a public company."

Under Sam Altman, OpenAI had been running like an incubator, placing bets across video, browsers, hardware, robots, and enterprise coding tools. That era is over. The company is now focusing on ChatGPT as a unified AI assistant and its enterprise Codex product as the two core pillars.

Our take: This is what IPO discipline looks like. Investors want a coherent story, not a dozen side projects. For UK businesses that had been experimenting with Sora for video content, this is a reminder: building workflows on top of startup products carries real platform risk. The underlying video generation technology will likely survive inside ChatGPT, but the dedicated tooling and API are gone. If you are evaluating AI video for commercial work, diversify your tooling now.

Arm Launches First Ever In-House Chip, Meta Is Lead Customer

In a historic move, Cambridge-based Arm has launched the AGI CPU, its first ever in-house chip after 35 years of licensing processor designs to others. The up-to 136-core data centre processor, built on TSMC's 3nm process with Neoverse V3 cores, is purpose-built for AI inference workloads.

Meta is the lead customer, with seven other committed buyers including OpenAI. Arm CEO Rene Haas has insisted the market needs a dedicated AI inference chip and that existing CPU offerings are not optimised for the workloads that will dominate data centres over the next decade.

Our take: This is enormous. Arm has always been the architect behind the scenes. Now it is competing directly with its own customers. For UK businesses, there is a national pride angle here: this is a British company making a play to own the silicon layer of AI infrastructure. But the strategic implications matter more. If Arm's inference chip delivers on efficiency claims, it could meaningfully reduce the cost of running AI models at scale, which flows directly into lower API pricing and more accessible AI for mid-market businesses.

LiteLLM Supply Chain Attack Compromises Major AI Library

LiteLLM, one of the most widely used Python libraries for interacting with AI model APIs (with over 97 million downloads), was hit by a supply chain attack on 24 March. Compromised versions 1.82.7 and 1.82.8 were uploaded directly to PyPI, bypassing the Project's CI/CD pipeline. The malicious code included a credential stealer that executed automatically on every Python process startup.

The threat actor, identified as TeamPCP, had previously compromised the Trivy security scanner and Checkmarx's AST GitHub Actions in the days prior. The compromised versions have since been yanked from PyPI.

Our take: If your development team uses LiteLLM (and if you are building AI integrations, there is a good chance they do), check your installed versions immediately. This is a textbook example of why software supply chain security cannot be an afterthought. Pin your dependency versions, run integrity checks, and audit what your CI/CD pipeline pulls from public registries. The AI tooling ecosystem moves fast, but that speed creates exactly this kind of attack surface.

Billion-Pound AI Data Centre Approved for West London

Pure Data Centres and SEGRO have received planning approval for a 56MW AI-ready data centre at Premier Park in West London. The development, estimated at around one billion pounds in gross capital investment, will be a purpose-built liquid-cooled facility engineered for high-density AI workloads. It will create hundreds of jobs and commit over three million pounds to community benefits.

Our take: This is a significant vote of confidence in UK sovereign AI infrastructure. While the US debates moratoriums on data centre construction (see below), London is pressing ahead. For UK businesses concerned about data residency and the latency costs of running AI workloads on US-hosted infrastructure, more domestic capacity is genuinely good news. The liquid-cooling design also signals that this facility is built for the GPU-dense inference and training workloads that define modern AI, not yesterday's cloud computing.

Senate Democrats Move to Codify Anthropic's AI Red Lines

Senator Adam Schiff is drafting legislation to codify Anthropic's red lines against autonomous weapons and mass surveillance. Senator Elissa Slotkin has separately introduced the AI Guardrails Act, which would limit the Department of Defense's ability to use AI for mass domestic surveillance and autonomous lethal force without human oversight.

This follows the Trump administration's decision to blacklist Anthropic after the company set limits on how the US military could use its Claude model. Anthropic has filed suit, calling the designation unconstitutional.

Our take: The fight between Anthropic and the Pentagon is no longer just a commercial dispute. It is becoming a legislative battleground over the fundamental question of AI governance: who gets to set the red lines? For UK businesses watching from across the Atlantic, this matters because it will shape which AI providers can operate freely in both US and allied markets. If you are building on Claude, the platform risk here is not technical but political.

Trump Appoints Zuckerberg, Huang, Brin to AI Advisory Panel

Mark Zuckerberg (Meta), Jensen Huang (NVIDIA), Sergey Brin (Google), and Larry Ellison (Oracle) are the first four members of Trump's President's Council of Advisors on Science and Technology (PCAST). The panel will advise on AI policy and could grow to 24 members. It will be co-chaired by AI and crypto czar David Sacks.

Our take: Notice who is not on this panel: Anthropic, which is currently suing the administration. The message is clear. If you play ball, you get a seat at the table. If you set limits, you get blacklisted. UK businesses should take note: the regulatory environment in the US is being shaped by the very companies it is supposed to regulate. The UK's own approach, through the AI Safety Institute and proportionate regulation, looks increasingly distinct from the American model.

Quick Hits