NVIDIA details Rubin GPU roadmap, promises 3.3x training speedup over Blackwell
At GTC, Jensen Huang laid out a year-by-year roadmap through 2027 — Rubin in 2026, Rubin Ultra in 2027 — and announced a new networking fabric called NVLink Switch.
At GTC, Jensen Huang laid out a year-by-year roadmap through 2027 — Rubin in 2026, Rubin Ultra in 2027 — and announced a new networking fabric called NVLink Switch.
Mistral’s newest flagship is trained entirely in EU data centers and ships with a commercial license that strict regulated industries can actually accept.
iOS 18.5 includes a refreshed 3B-parameter model running on-device, with significant improvements to summarization and writing tools.
It doesn’t make TechCrunch headlines often, but Cohere’s Command R+ is winning enterprise RAG deployments on price, latency and on-prem deployability.
We surveyed 800 enterprises on their AI budgets. The headline: spend is up 3.5x year-over-year, but only 14% of pilots are reaching production.
Picking an LLM in 2025 isn’t about benchmarks. It’s about latency, price predictability, data residency and the unglamorous question of who answers your support ticket.