Welcome back to 4IR. Here's today's lineup:
UAE launches K2 Think—matches GPT-4 at 1/20th the size - Middle East drops fully open-source 32B parameter reasoning model
Mistral AI hits €11.7B valuation as chip giant ASML invests €1.3B - Europe's AI champion gets biggest funding round ever
NVIDIA Blackwell destroys benchmarks—5x faster, 70% cheaper - MLPerf results show massive leap in AI inference speed
OpenAI Fund backs AI security with $55M for Adaptive - Startup building cages for autonomous AI agents
Windows 11 finally brings AI to Intel and AMD chips - September update ends Qualcomm's AI monopoly
🔥 TOP STORY: UAE drops 32B model that matches 600B giants
The story: UAE's Mohamed bin Zayed University and G42 launched K2 Think on September 9, a reasoning model with just 32 billion parameters that matches the performance of OpenAI and DeepSeek systems using 600+ billion parameters. The model scored 90.83% on AIME 2024 mathematics and 71.08% on GPQA-Diamond benchmarks, putting it on par with the world's best reasoning models. Unlike every other major player, the UAE is releasing everything under Apache 2.0 license: model weights, training data, even the deployment code. This marks the Middle East's first serious entry into frontier AI development.
What we know:
32B parameters launched September 9
Matches 600B+ model performance
Fully open-source with Apache 2.0
90.83% on AIME 2024 benchmarks
Built on Alibaba's Qwen 2.5 foundation
Named after K2, world's second-highest mountain
Why it matters: Someone just proved you don't need a trillion parameters to build GPT-4. This breaks the entire "bigger is better" narrative that's driven AI development for three years.
The UAE isn't trying to beat OpenAI. They're making OpenAI irrelevant. Why pay $200/month for ChatGPT when you can run K2 Think for free? The 20x size reduction is insane—it's like building a Ferrari engine that fits in a Toyota. The full open-source release is the power move. No API limits, no usage restrictions, just download and run. This forces everyone's hand. OpenAI has to explain why they need 600B parameters. Anthropic has to justify their pricing. Watch for a wave of K2 Think improvements from the open-source community. The genie is out.
🧠 BREAKTHROUGH: Chip giant ASML bets €1.3B on Europe's AI champion
The story: Mistral AI closed a massive €1.7 billion Series C on September 9, with semiconductor equipment monopolist ASML leading with a €1.3 billion investment for an 11% stake. The French startup's valuation hit €11.7 billion, making it Europe's first AI decacorn and the continent's answer to OpenAI. ASML controls 90% of the extreme ultraviolet lithography market—their machines are required to make every cutting-edge AI chip on earth. The deal includes ASML's CFO Roger Dassen joining Mistral's strategic committee, creating a direct link between AI development and chip manufacturing.
What we know:
€1.7B raised September 9
ASML invests €1.3B for 11% stake
€11.7B valuation ($13.8B)
ASML CFO joins strategic committee
First European AI company above €10B
Partnership targets chip design optimization
Why it matters: The company that enables all AI chips just bought into an AI company. ASML now has skin in the game beyond just selling equipment.
This deal is bigger than the numbers. ASML doesn't do venture investments—they do strategic moves that shape industries. When the company that literally makes chip-making possible drops €1.3B, everyone should pay attention. The board seat matters more than the equity. ASML will shape Mistral's strategy, likely driving exclusive optimization for European chip manufacturing. This isn't a funding round; it's Europe planting its flag in the AI race. Macron gets his champion, ASML gets AI expertise for chip design, and Mistral gets patient capital from a company that thinks in decades, not quarters.
💰 MOONSHOT: NVIDIA's new chips are 5x faster and 70% cheaper
The story: NVIDIA's Blackwell Ultra platform obliterated MLPerf inference benchmarks in results released September 9, delivering 5x higher throughput than the previous Hopper generation and 45% better performance than standard Blackwell chips. The GB300 NVL72 system hit 5,842 tokens per second per GPU on DeepSeek-R1 reasoning workloads, while cutting inference costs by 70%. These are official industry-standard MLPerf results, not NVIDIA marketing claims, and they specifically optimized for reasoning models—the next frontier in AI applications.
What we know:
MLPerf v5.1 results released September 9
5x faster than Hopper H200 chips
45% better than standard Blackwell
5,842 tokens/second/GPU achieved
70% reduction in inference costs
1.5x improvement in FP4 compute
Why it matters: NVIDIA just made AI affordable for everyone. The 70% cost cut transforms the economics—use cases that lost money at $10 per million tokens become profitable at $3.
These aren't incremental improvements. This is NVIDIA breaking the sound barrier. Five times faster means real-time everything: video analysis, instant code generation, actual conversations with AI. The focus on reasoning models tells you where NVIDIA thinks AI is headed—not chatbots, but AI that can actually think. The 70% cost reduction changes the entire market. Every company waiting for prices to drop just got their wish. This is why NVIDIA is worth $3 trillion. They don't just make chips; they make the future cheaper.
📰 BATTLEGROUND: OpenAI Fund drops $55M on AI agent security
The story: The OpenAI Fund announced September 9 it's extending its investment in Adaptive Security, bringing the AI security startup's Series A to $55 million total. Adaptive is building "agent-native" security platforms that monitor and control autonomous AI systems in real-time, addressing what they call "the largest attack surface expansion in computing history." The company was founded by former Anthropic and DeepMind security researchers who've seen firsthand what happens when AI safety measures fail. The investment comes as enterprises rush to deploy AI agents without understanding the security implications.
What we know:
Extended Series A to $55M on September 9
OpenAI Fund leading the extension
Building "agent-native" security platforms
Real-time monitoring and intervention
Ex-Anthropic and DeepMind founders
Targeting enterprise AI deployments
Why it matters: OpenAI is funding the antidote to its own poison. They know autonomous agents are coming—and they know how dangerous unsecured agents could be.
Smart money says this is Sam Altman's insurance policy. Fund the security company before you release the agents that need securing. When regulators ask about safety, OpenAI points to this investment. When enterprises worry about risk, OpenAI offers Adaptive as the solution. The founders matter here—ex-Anthropic and DeepMind means they've seen the failure modes. Traditional security can't handle AI agents that can rewrite their own code or social-engineer humans. If every AI agent needs security, Adaptive becomes critical infrastructure. The $55M is cheap for what could become mandatory.
🚀 VELOCITY: Microsoft brings AI to all PCs in Windows 11 update
The story: Microsoft's September 9 Windows 11 security update (KB5043080) finally brings AI capabilities to Intel and AMD processors, ending Qualcomm Snapdragon's exclusive hold on features like Recall and Windows Studio Effects. The update ships nine major AI enhancements including a completely redesigned Recall app with privacy controls, Copilot Daily that reads personalized news using AI voice synthesis, and semantic search in File Explorer. This marks Microsoft's biggest AI feature expansion since Copilot's initial launch, making AI capabilities standard across all modern Windows PCs regardless of chip architecture.
What we know:
KB5043080 deployed globally September 9
Intel Core Ultra and AMD Ryzen AI now supported
Recall redesigned with local-only privacy mode
Copilot Daily provides AI-voiced news briefings
File Explorer gets semantic search
Windows Studio Effects for all compatible CPUs
Why it matters: Microsoft just flipped the switch on AI for millions of PCs. The Qualcomm exclusive is dead—now Intel and AMD can compete for the AI PC market.
Microsoft's Qualcomm bet didn't work out. Nobody bought ARM laptops when their Intel machines work fine. So Microsoft did what they always do—support everything, everywhere. The Recall redesign is damage control done right: local-only, encrypted, opt-in. Exactly what it should have been from day one. Copilot Daily is the sleeper feature—AI reading you personalized news could kill podcasts and morning news apps. This update makes AI on Windows inevitable. By next year, you won't remember Windows without AI. That's how Microsoft wins—not with flash, but with ubiquity.
Note: Commentary sections are editorial interpretation, not factual claims