OpenAI's Jalapeño Chip Outpaces Nvidia in Inference Speed
- ai-news
- openai
- hardware
- enterprise-ai
OpenAI's Jalapeño Chip Beats Nvidia Blackwell in Inference Benchmarks
At this week's Hot Chips conference, OpenAI shared the first detailed benchmark results for its Jalapeño inference chip, developed in close collaboration with Broadcom. Tested against SemiAnalysis' InferenceX benchmark, Jalapeño delivered more tokens per user (i.e., faster responses) and more throughput per kilowatt of power than current state-of-the-art Nvidia Blackwell systems.
"The bottom line is that the results show a very, very significant performance advance over state of the art," said Richard Ho, OpenAI's head of hardware, on a press call. "Jalapeño can serve more AI work per unit of power, while also returning responses more quickly." The chip is expected to deploy in small volumes by the end of 2026, with broader deployment coming in 2027.
Why it matters for businesses: faster response times and lower power costs per AI request ultimately translate into lower service costs and improved response times for businesses relying on AI tools daily — whether for customer service chatbots, data analysis, or office automation agents.
OpenAI Launches Admin Plugin for ChatGPT Work and Codex
Separately, OpenAI introduced a new Admin Plugin for ChatGPT Work and Codex, letting IT admins manage permissions, usage limits, spending requests, and team members — all from a single chat conversation, without switching between tools. The plugin can also automate recurring approval workflows, such as routing requests to Slack or Microsoft Teams. Inside OpenAI's own IT team, similar workflows already resolve about 45% of employee ticket volume automatically.
This is a particularly practical tool for small and medium businesses managing multiple AI accounts and users who want better control over costs and approval processes without dedicated engineering resources.
Claude Cowork Gets Persistent Memory
Anthropic added the same memory system used by Claude's chatbot to Cowork, allowing the tool to pick up tasks without users re-explaining important context each time. Memory is on by default, but the system avoids saving sensitive topics like health or beliefs unless manually enabled. Users can edit or delete any saved memory through the settings menu.
Google Launches Gemini AI for Financial and Legal Services
Google announced a new financial research agent along with a legal AI tool that helps law firms draft briefs, track regulatory changes, and verify legal citations — an increasingly important issue as more lawyers adopt AI tools and encounter hallucinations and errors in their research.
Perplexity and Nvidia Launch a Local AI Computer With Zero Token Costs
Perplexity and Nvidia unveiled "Portable Computer" — a version of Perplexity's agentic platform that runs entirely on local hardware (Nvidia's DGX Spark and Linux machines with RTX GPUs), requiring no cloud connection and no ongoing usage costs. The system keeps sensitive information, like financial documents, on the device itself, only requesting permission to reach a more powerful cloud model when extra capability is needed. The tool is currently available on Linux, with Windows support coming in September.
For businesses handling sensitive data and looking to control long-term AI costs, this points to an interesting direction for the next generation of enterprise AI tools.
Today's trend is clear: leading AI companies aren't just improving the models themselves — they're also upgrading the infrastructure around them: faster chips, business management tools, smarter memory, and flexibility between cloud and local computing. For small and medium businesses, this means more accessible, more affordable, and more precise tools for everyday needs. At RealOffice.AI, we see the same trend: as AI agents become faster and smarter, the potential for real office automation grows — handling inquiries, coordinating tasks, and managing routine work, freeing up valuable time for business owners to focus on what truly matters.