2 min read

AI’s full-stack race moves from models to the machinery around them

AI’s full-stack race moves from models to the machinery around them OpenAI says its first custom inference chip, Jalapeño, delivered 1.5–1.9× more AI work per watt and 1.7–3.6× lower end-to-end latency than comparison systems across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. The point isn’t a benchmark trophy: agents make many sequential calls, so latency and power compound across every task. OpenAI is now treating chips, serving software, memory and networking as one product surface. That ra

AI’s full-stack race moves from models to the machinery around them

OpenAI says its first custom inference chip, Jalapeño, delivered 1.5–1.9× more AI work per watt and 1.7–3.6× lower end-to-end latency than comparison systems across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T. The point isn’t a benchmark trophy: agents make many sequential calls, so latency and power compound across every task. OpenAI is now treating chips, serving software, memory and networking as one product surface. That raises the pressure on incumbent accelerators—and makes inference economics a customer-facing feature. Read OpenAI’s Jalapeño results.

What else matters

OpenAI puts workspace administration inside ChatGPT

OpenAI introduced an Admin plugin for ChatGPT Work and Codex, allowing administrators to inspect usage, manage members and permissions, adjust limits, and approve supported actions in one conversation. The practical shift is from AI as a help desk to AI as an authorised control layer; auditability and approval boundaries will matter as much as convenience. Read OpenAI’s Admin plugin announcement.

NVIDIA pushes agentic AI to the edge

NVIDIA announced Jetson Orin Nano 2, an entry-level robotics computer claiming twice the inference performance of its predecessor and 40% lower power at the same performance. With 78 trillion operations per second and 8GB of memory, it targets robots, drones and vision systems that need local, real-time reasoning rather than a round trip to the cloud. Read NVIDIA’s announcement.

GitHub makes Teams a launchpad for coding agents

GitHub’s Copilot cloud agent can now be started from a Microsoft Teams channel, thread or direct message, with participants able to add context and steer the work. Paid-plan users can hand meeting decisions to an asynchronous agent in a cloud sandbox; organisations will need to govern credits, repository write access and sandbox budgets. Read the GitHub changelog.

Continue reading