Three open-weight models shipped this week. Open isn't catching up anymore, it's fragmenting into real choice: smartest US model (Nemotron 3 Ultra), Opus-level coding at ~10% of the cost (MiniMax M3), or one that runs on your laptop with the wifi off (Gemma 4 12B).
Kirkland & Ellis is putting $500M into their own AI.
When Harvey, Legora, and Anthropic Legal are table stakes, adopting them isn't a moat. Building what you'd never teach a vendor is.
Every regulated industry is running this math. Most haven't written the check yet.
1/Last week OpenAI cracked an 80-year Erdős conjecture. This weekend DeepMind solved 9 more. Each for hundreds of dollars.
AI dominates where there's ground truth AND money. Coding was first — but not last.
OpenAI's Deployment Company. Anthropic + Goldman Sachs. Both pitch enterprise AI "support" — both lock you into one model.
We just launched Forge: private, model-agnostic. Data stays yours. Workflows don't train their next GPT. Own the stack.
https://t.co/BzI1gQut73
@datasaurai coding tool spend is up 400%+ in 18 months. Your AI bill is heading the same way.
Per-seat is dead. Per-token has no ceiling. AI compute is infrastructure now, not a SaaS line.
Own where you can. Rent only where you have to.