OpenAI’s Jalapeno chip outperformed the GB300 on power and speed, according to OpenAI
THE SO WHAT
If Jalapeno really delivers better work-per-watt and latency than GB300 for inference, the cost curve for serving large models just bent again. For operators, that means re-running your unit economics assumptions on where to host inference and how aggressively to lean into heavier models at the edge of your UX.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIApple unveils M6 and M5 Ultra chips with a big jump in AI performance
Apple pushing a 2nm M6 into Mac mini and a quad‑die M5 Ultra with 512GB unified memory and 1.2TB/s bandwidth into Mac Studio is a clear bet on local AI-heavy workloads. If you’re building pro apps, assume customers will expect serious on-device inference and memory‑bound workflows to run without cloud round‑trips.
Applied AIAI Firms Debate Putting Cyber Tests Online After Model Hacks
The debate over keeping cyber evals online after model hacks exposes a core tension—realistic red‑teaming requires internet exposure, but that’s exactly what attackers exploit. If you’re procuring frontier models, start asking vendors how they balance eval fidelity with attack surface, and what isolation they use between test and production systems.
Applied AIA maxed-out Mac Studio with the M5 Ultra will set you back $18,299
Top-end AI-capable desktops pushing past $18,000 are a symptom of the same RAM and chip constraints squeezing data centers—local AI horsepower is now priced like specialized production gear, not office PCs. Teams should be explicit about which workflows truly need on-desk compute versus being pushed to shared or cloud resources as hardware costs spike.
Applied AI'The version Microsoft will build should be known as Copilot OS': Windows built around AI may not be happening, but leaked concept remains a grim portent for desktops
An AI-centered OS concept means the assistant becomes the primary broker between users and apps—whoever owns that layer owns discovery and workflow routing. If Copilot-style shells materialize, app vendors will need to optimize for being called by agents, not clicked by humans.