Trump Signs AI Executive Order — Frontier Models Get 30-Day Voluntary Review
President Trump signed a scaled-back AI executive order on June 2 titled “Promoting Advanced Artificial Intelligence Innovation and Security,” after postponing it last month. The centrepiece: a voluntary 30-day federal review for frontier AI models before they’re released to the public, with a focus on national security and cybersecurity risks.
Unlike the broader order cancelled two weeks ago, this version gives the NSA a central role in vetting and leaves significant discretion to AI companies about whether to share information. The Cybersecurity and Infrastructure Security Agency (CISA) will work alongside the NSA on assessments.
Critics say the “voluntary” framing means it’s toothless for companies that opt out. Supporters argue the review framework establishes a norm of pre-release safety checks without heavy-handed mandates that could push AI development offshore.
Why it matters: This is the US government’s most concrete attempt at frontier model oversight yet — and the “voluntary with NSA” design says a lot about the administration’s AI philosophy: prioritise security (cyber threats, WMD risk) over broader safety or bias concerns, and lean on existing agencies rather than creating new regulatory bodies. Whether companies actually submit models for review will determine if this is a real safeguard or a press release.
Google Launches Gemma 4 12B — An Encoder-Free Multimodal Lightweight
Google dropped Gemma 4 12B on June 3, a new 12-billion-parameter variant of its open-weight family with a distinctive design choice: no separate vision encoder. Unlike most multimodal models that bolt a vision encoder onto a language model, Gemma 4 12B processes images and text through a unified architecture, treating visual tokens the same as text tokens.
The model is built from Gemini 3 research and targets on-device deployment — think phones, laptops, and edge devices. Google claims the encoder-free approach improves cross-modal reasoning while reducing memory footprint. It supports up to 256K tokens context.
The timing is strategic: with Gemma 4 12B, Google now offers an open model that can run locally while handling images, text, and soon audio natively — all under the Apache 2.0 license. The original Gemma 4 family launched in April with sizes from 2B to MoE 26B.
Why it matters: Encoder-free multimodal is a genuine technical departure. Most lightweight open models still bolt on a visual encoder. If Gemma 4 12B performs as claimed, it changes the economics of on-device AI — you get vision+language in a single 12B package that fits on consumer hardware. The broader shift: Google is increasingly hedging between proprietary (Gemini) and open (Gemma), and the gap between the two narrows with every release.
UK CMA Forces Google to Let Publishers Opt Out of AI Search Results
The UK’s Competition and Markets Authority (CMA) imposed new conduct requirements on Google on June 3, ordering the company to:
- Let UK publishers opt out of AI Overviews — Google’s AI-generated search summaries
- Allow publishers to prevent their content from being used to “fine-tune” Google’s AI models
- Publish clearer data on how AI features affect traffic to publisher sites
The ruling comes under the UK’s Digital Markets, Competition and Consumers Act (DMCCA), giving the CMA direct enforcement power over Big Tech’s gatekeeper platforms. Google must implement the changes within three months.
Why it matters: This is the first binding regulation of AI search features anywhere in the world. The AI Overviews debate has been raging since Google launched them — publishers lose traffic when AI summarises their content instead of sending users to their sites. The CMA just gave publishers an escape hatch. If other regulators follow (the EU, Australia, Canada are watching), it fundamentally changes the economics of AI-driven search.
EU Unveils €320 Billion Tech Sovereignty Plan to Break US Dependence
The European Commission unveiled a sweeping Tech Sovereignty Package on June 3, targeting chips, cloud computing, and AI — and explicitly naming US tech dominance as the threat. The plan calls for:
- €320 billion over 10 years to build European alternatives in cloud, AI, and semiconductors
- A Sovereign Cloud Framework — €180M contract already awarded to four EU providers
- Open-source mandates for publicly funded AI projects
- Reduced reliance on US hyperscalers (AWS, Azure, GCP) for government and sensitive data
The plan’s reception has been mixed. Renew Europe called it “a step when we needed a leap,” saying the ambition is right but the funding mechanisms are vague. The Commission estimates the package needs 4:1 private-to-public co-investment to hit the targets.
Why it matters: The EU has regulated US tech (DMA, DSA, AI Act). Now it’s trying to build alternatives. The €320B figure is eye-watering, but it’s dwarfed by US private AI investment — Meta alone committed $145B. The open-source requirement for publicly funded AI is the most interesting policy lever: if the EU forces European-funded models to be open, it could create real competition but also raise questions about US companies benefiting from EU-funded research.
Meta Backs Down on Worker Tracking — Employees Get 30-Minute Opt-Out Windows
Meta has scaled back its controversial Model Capability Initiative (MCI) — the program that recorded employees’ keyboard and mouse activity for AI training — after weeks of staff pushback and potential GDPR violations flagged by Reuters.
The new compromise: workers can opt out of tracking for 30-minute blocks when they “need to check something personal.” But the opt-out is temporary and requires an active decision each time. The tracking tool could also capture emails and chats from non-US employees, triggering EU privacy law concerns.
Meta CTO Andrew Bosworth reportedly told employees they could be “angry or shocked, but there is no option to opt out permanently.” The 30-minute window is a concession, not a policy reversal.
Why it matters: This captures the tension at every major AI lab right now. Training frontier models requires vast behavioural data — and companies are going after employee data as a gold mine. Meta’s attempt and subsequent partial backdown will be watched closely by Google, Microsoft, and OpenAI, all of whom face similar internal tensions. The 30-minute compromise also demonstrates a fundamental truth about workplace AI surveillance: you can’t really opt out while keeping your job.
DDR5 “RAMageddon” — 32GB Now Costs $375 as AI Demand Crowds Out Consumers
The AI data centre buildout is wreaking havoc on the consumer memory market. 32GB of DDR5 now costs a minimum of $375 — up from $90 in early 2024 — as memory manufacturers prioritise high-bandwidth memory (HBM) for AI GPUs over consumer DRAM.
TrendForce reports DRAM contract prices surged 90-95% quarter-over-quarter in Q1 2026, the largest increase on record. Samsung and SK Hynix continue to raise prices in Q2, as fabs are retooled from DDR5 to HBM production. A threatened Samsung walkout could drive prices higher still.
The impact extends beyond PC builders. AMD’s new Gorgon Halo chips and Nvidia’s RTX Spark face a 63% DRAM cost hike this quarter. Even Apple’s latest devices are affected.
Why it matters: “RAMageddon” — yes, it’s really being called that — is the most tangible way the AI boom affects non-AI consumers. A PC builder who could drop $90 on memory 18 months ago now pays $375, assuming they can find stock. The memory supercycle shows no signs of easing until HBM fab capacity catches up, which analysts project for sometime in 2027. For now, if you’re building a PC, buy RAM the moment you see a deal.
Microsoft’s MAI-Code-1-Flash Beats Claude Haiku — 60% Fewer Tokens
Microsoft released MAI-Code-1-Flash on June 2, a coding model that outperforms Claude Haiku 4.5 on every benchmark while using up to 60% fewer tokens. The 137B-parameter MoE model is built by Microsoft’s Superintelligence team and rolled out to GitHub Copilot users in VS Code.
The release is part of Microsoft’s broader MAI model portfolio, launched at Build 2026, which includes the MAI-Thinking-1 reasoning model and six other variants. The strategic message is clear: Microsoft is building alternatives to OpenAI models for its own stack.
Why it matters: Microsoft is hedging its OpenAI bet. For years, Copilot has depended on OpenAI models. MAI-Code-1-Flash is the first credible in-house alternative — and if it’s genuinely 60% more efficient, it makes financial sense for Microsoft to shift workloads internally. The broader implication: every major platform company is now building its own AI models, reducing dependence on the frontier labs.
Let’s Encrypt Moves to Post-Quantum Certificates
Let’s Encrypt announced on June 3 that it will begin issuing post-quantum hybrid certificates using the ML-KEM (Kyber) key encapsulation mechanism alongside traditional ECDSA certificates. The move makes Let’s Encrypt — which issues certificates for over 300 million websites — one of the largest certificate authorities to support post-quantum cryptography in production.
The hybrid approach means certificates include both a quantum-safe key and a traditional key, ensuring compatibility as the ecosystem transitions. Let’s Encrypt says the transition will be gradual, with full PQC support expected by 2028.
Why it matters: “Harvest now, decrypt later” attacks — where adversaries collect encrypted traffic today to decrypt it when quantum computers arrive — are a real threat. Let’s Encrypt’s move means millions of websites can now deploy quantum-safe encryption without complex configuration. It’s a behind-the-scenes win for internet security that most users will never notice.
🔍 THE BOTTOM LINE
The dominant theme this week is the geopolitical and regulatory pushback against AI concentration: the US tries soft-touch model review, the UK regulates AI search features, and the EU drops a €320B sovereignty bombshell. Meanwhile, the hardware realities of the AI boom hit consumers (RAMageddon) while the software competition keeps accelerating (Gemma 4 12B, MAI-Code-1-Flash). The regulatory landscape is fragmenting fast — and companies will need to navigate at least three very different frameworks (US voluntary, UK binding, EU sovereign) to operate globally.
❓ Frequently Asked Questions
Q: Does Trump’s AI executive order actually require companies to submit models? No — the 30-day review is voluntary. But the NSA and CISA will conduct assessments for companies that participate. The order creates a framework for pre-release reviews without mandatory compliance, betting that companies will participate to avoid liability risk.
Q: Will the UK CMA ruling affect Google Search globally? The ruling specifically covers the UK market, but Google will likely implement changes globally rather than maintain separate search systems for the UK. Similar regulations are under consideration in Australia, Canada, and the EU.
Q: What is the EU’s Tech Sovereignty Package actually funding? Three areas: sovereign cloud infrastructure (€180M contract already awarded), semiconductor manufacturing (reducing reliance on TSMC), and European AI model development (with open-source requirements). Total ambition is €320B over 10 years, but specific funding sources remain unclear.
Q: Why is DDR5 RAM so expensive now? AI data centres consume vast amounts of HBM memory for GPU clusters. Memory manufacturers have retooled fabs from DDR5 production to HBM production, creating a supply crunch for consumer DRAM. Contract prices rose 90%+ in a single quarter. Analysts don’t expect relief before 2027.
SOURCES
- The White House — Executive Order on AI Innovation and Security
- The Verge — Trump signs executive order to review AI models
- The Register — Trump AI executive order sets 30-day frontier model review
- Scientific American — Trump’s new AI executive order
- Google AI Blog — Introducing Gemma 4 12B
- The Verge — UK CMA secures Google publisher AI-search opt-out
- Bloomberg — EU unveils tech sovereignty plan
- Politico EU — EU plots long game against US digital supremacy
- BBC — Meta workers can opt out of tracking for 30 minutes
- Tom’s Hardware — 32GB DDR5 now costs $375 minimum
- PCMag — Meta workers 30-minute breaks from tracking
- Microsoft AI — Introducing MAI-Code-1-Flash
- Let’s Encrypt — A Post-Quantum Future
- Yahoo Tech — What is RAMageddon