Nvidia projects $1 trillion in GPU sales through 2027, and that number is based on actual purchase orders from companies ready to buy. That single stat from Jensen Huang’s GTC keynote might be the most telling signal about where AI is heading right now.
But that was just one story in a week absolutely packed with announcements. The creator behind this breakdown, Matt Wolfe from Future Tools, ran hands-on tests and pulled together every major development worth knowing about. Here’s what stood out.
📌 Two image models dropped, and one clearly won
MidJourney released V8, and the early consensus isn’t great. Wolfe tested it himself with detailed prompts for instruction following, text rendering, and creative scenes. The verdict: it’s extremely fast, but the quality feels like a step backward. Broken arms, mangled fingers, blurry text when the prompt specifically asks for readable words. Where it still shines is weird, surreal, creative stuff. But for anything specific or realistic, it’s struggling.
Meanwhile, Microsoft quietly dropped MAI Image 2, now ranked third on the text-to-image arena behind OpenAI and Nano Banana. Wolfe tested it with the same rigor and found it nailed realism, text rendering, and prompt adherence. A coffee shop menu board came out pixel-perfect. The contrast between these two releases was stark.
📌 Google launched a design-to-code pipeline
Two releases work together here. First, Google Stitch is an AI-native design canvas (think Figma meets AI). You describe what you want, it generates multiple design variations with different color schemes and layouts. You can even talk to it with voice commands.
Second, Google AI Studio now has full-stack vibe coding built in. Wolfe took a design from Stitch, exported it, dropped it into AI Studio, and got a functional website with animations, hover effects, and working filters from a single prompt. The loop between designing and coding is getting remarkably tight.
📌 The LLM updates, rapid fire
- OpenAI released GPT-5.4 Mini and Nano, smaller and cheaper models optimized for agent use cases
- Anthropic made 1 million token context windows available in Opus 4.6 and Sonnet
- Mistral dropped Small 4, an open-weight model performing close to Claude Haiku on reasoning tasks
- Cursor shipped Composer 2, delivering near-GPT-5.4 coding performance at significantly lower cost
- MiniMax released M2.7, a proprietary model that handled 30-50% of its own development workflow during training
📌 Nvidia GTC’s other big moves
Beyond the trillion-dollar projection, Nvidia announced NemoClaw (a security-focused wrapper around OpenClaw that installs with one command), DLSS5 for AI-enhanced game graphics, and early plans for space computing modules, though heat dissipation in vacuum remains unsolved.
📌 Jobs and the data economy
Andre Karpathy released a US job market visualizer showing which roles AI threatens and which are growing. Interesting finding: more jobs are appearing than disappearing. And DoorDash is now piloting a program paying people to collect real-world AI training data, filming tasks and recording speech. Wolfe makes a compelling point that paying people for data could become a significant income stream as AI companies exhaust freely available training material.
This was one of those weeks where every corner of AI moved at once. Check out the full video for the hands-on demos and side-by-side comparisons that really drive the points home.