Trump Blacklists Anthropic — Then Claude Rockets to #2 on the App Store

The Trump administration made an unusual power move this week, blacklisting Anthropic — the AI safety company behind Claude — from federal government contracts and procurement. The reasoning? Anthropic had pushed back against a Pentagon AI deployment plan, insisting on safety restrictions that the administration deemed too restrictive. Washington doesn't like being told no.

But here's where it gets wild. Within 48 hours of the blacklisting news breaking, Claude shot up to #2 on the Apple App Store. Anthropic's ban became the best free marketing campaign in the history of AI. Millions of curious users apparently decided that the company the government wanted to punish must be doing something right.

Meanwhile, OpenAI quietly finalized its own Pentagon AI deal — and buried in the fine print were safety restrictions almost identical to the ones Anthropic had demanded. So the administration blacklisted one company for safety standards and secretly accepted the same standards from another. The lesson, apparently, isn't about safety. It's about who you know.

New International Report: AI Can Now Write Step-by-Step Bioweapon Instructions

The 2026 International AI Safety Report — compiled by researchers from 30 countries — dropped this week with a finding that should make everyone uncomfortable: current AI models can provide meaningful, actionable assistance to someone trying to create a biological weapon. Not theoretical assistance. Step-by-step synthesis instructions, strain selection guidance, and delivery mechanism suggestions.

The report didn't name specific models, but noted that the capability has emerged across multiple frontier systems in the past year. Safety guardrails exist, but researchers found they could be circumvented through careful prompting strategies that don't require technical expertise. In other words, the attack surface isn't just for experts anymore.

The same report flagged real-time cyberattack assistance as an equally serious concern. AI systems can now help attackers identify vulnerabilities, write exploit code, and adapt attacks in real time — effectively giving every script kiddie the capabilities of an experienced red team. The report called for mandatory capability evaluations before model releases, but stopped short of recommending any enforcement mechanism.

AI Now Eats 4% of All U.S. Electricity — and Data Centers Are Nearly Full

Data centers powering America's AI boom now consume 4% of total U.S. electricity generation — and that number is climbing fast. According to a new industry report, data center occupancy rates are projected to hit 95% by the end of 2026, meaning the physical capacity for AI computation is nearly maxed out. The race to build more isn't just a tech story anymore; it's an energy story.

The power demand is so acute that tech companies are restarting nuclear reactors. Three Mile Island Unit 1 — which had been offline since 2019 — came back online this month exclusively to power a major AI data center campus. Microsoft, Google, and Amazon have all signed long-term nuclear power agreements in the past 90 days. Wind and solar can't ramp fast enough to meet demand that doubles annually.

The bottleneck isn't compute anymore — it's electricity. GPU clusters sit idle waiting for power hookups. New data center construction is running 18-24 months behind demand, and grid interconnection queues are years long. The companies that lock in power deals today will have a structural advantage in 2027-2028 that no amount of model innovation can overcome.

Seven Frontier AI Models Dropped in February. We Are Not Ready.

February 2026 will go down as the most chaotic month in the history of AI releases. Seven frontier models shipped in 28 days: Gemini 3.1 Pro, Claude Sonnet 5, GPT-5.3, Qwen 3.5, GLM 5, DeepSeek v4, and Grok 4.20. That's not counting the dozens of fine-tuned variants, API updates, and multimodal add-ons that came with them. The industry released more capable AI in one month than existed in total three years ago.

Each model claims to be state-of-the-art. Each has a benchmark suite designed to prove it. The reality is murkier — different models excel at different tasks, benchmarks are increasingly gamed, and the actual capability differences between top models are narrowing. What's undeniable is the pace. We are adding more frontier AI capacity every month than most companies will ever need.

The concerning part isn't the number — it's the deployment speed. These models go from training completion to API availability in weeks. The gap between "we built it" and "millions of people are using it" has collapsed to nearly zero. There is no meaningful buffer for discovering what these systems can do before they're already doing it at scale.

Someone posted an AI-generated Stranger Things scene so convincing that half the comments thought it was a leaked Netflix clip. Made with Seedance 2.0, it nails the lighting, the practical effects look, and the character movement of the original show — without a single camera or crew member. Hollywood's most expensive asset used to be production quality. Watch it here → — 10K likes and the comments are genuinely confused about whether it's real.

Perplexity — When the news is moving this fast, you need a source that cites its sources. Perplexity searches the web and shows you exactly where every claim came from, making it the most useful tool for verifying fast-moving AI news without falling for hallucinations or outdated information. Treat it like a search engine that actually reads the results for you.