@NRedl and I built VibeIcons, an #AI#icon generator that creates any icon in the style of Heroicons, Phosphor, Lucide, or Eva. Always consistent, always SVG.
Oracle is set to deploy 50,000 AMD GPUs by Q3 2026 to power its AI “supercluster.” 
This says something loud: if you don’t own compute, you might just be a passenger in the AI value chain.
For founders: models + UX will be table stakes. The competitive moat will increasingly be in infra architecture, cost efficiency, and scaling inference.
I'm back... I just launched https://t.co/lpWau3jgBx!⚡
Real humans write replies that recommend your product and bring customers - while you sleep 😴
Let's go!
@Microsoft just launched its first in-house text-to-image model, MAI-Image-1.
What I like: they’re signaling they want control over the stack, not just layering on third-party models.
What I’m skeptical of: will it meaningfully outperform composable modular systems (e.g. chained diffusion + control nets) in domain tasks?
(Thread 👇)
1/ This shows the trend: big players re-internalizing model + compute.
2/ For a founder, this changes your calculus: can you still “just use OpenAI + fine-tune + wrap”?
3/ But it opens opportunities too: small modular model stacks might beat monoliths in niche verticals.
Ultimately: I’m betting many “AI tools” built on third-party backends will feel brittle 2–3 years out unless you own at least some of the path.