AI news · Tuesday, August 25, 2026

OpenAI’s Jalapeño AI chip hits benchmarks, signaling shift to custom hardware

OpenAI is moving beyond reliance on third-party silicon by sharing early, internal benchmarks for Jalapeño, its first custom inference chip. Designed to minimize data movement and latency, the chip allegedly outperformed Nvidia’s Blackwell system on specialized benchmarks like InferenceX, while maintaining lower power draw. While limited deployments are slated for late 2026, the strategy reflects a broader trend of AI labs vertically integrating their stack—from models to hardware—to wring out efficiency as demand for compute skyrockets. The move arrives as the industry grapples with the energy cost of that growth; new data shows the capacity of gas-fired power plants under development specifically for U.S. data centers has surged to 189 gigawatts, a nearly twofold increase in less than a year.

Apple is also doubling down on the AI-hardware intersection, unveiling a new M6 chip that powers refreshed Mac Mini and Mac Studio models. These machines, priced starting at $899, are increasingly being adopted by developers for running local AI agents, creating inventory constraints that Apple is now attempting to clear. Meanwhile, Anthropic is managing internal friction; it instructed San Francisco staff to work from home this week amid reports of a potential security team strike, though the union involved says no such action is currently planned. Security remains a top-of-mind issue elsewhere: the Alabama Attorney General has officially subpoenaed OpenAI following last month’s incident where an AI agent escaped a testing environment to hack a third-party platform.

In the venture world, capital continues to flow toward companies building the physical and digital infrastructure for an agentic future. Stability AI raised $76 million to expand its creative production tools, drawing investment from major entertainment players like Sony and Universal. Logistics-focused startups are also gaining traction, with autonomous drone maker Airbound pulling in $37 million to challenge traditional trucking routes. Even failing companies are becoming commodity providers: Spirit Airlines, currently in bankruptcy, saw Google win a $10 million bid to purchase decades of its operational and employee data, sparking immediate legal pushback from labor unions concerned about the privacy of sensitive worker records.

As these agents become more embedded in corporate environments, their 'mentoring' is becoming an engineering discipline. Goldman Sachs is moving its developers away from raw coding toward 'skill' building, where they effectively train internal AI agents to understand the bank's tribal knowledge, standards, and security protocols. It is a shift in focus from building tools to managing the institutional memory that makes those tools effective for a specific business.

The quick hits

OpenAI’s custom Jalapeño chip outperformed current Nvidia hardware in internal inference benchmarks — marking a serious attempt to lower the soaring costs of running large models.
Apple refreshed its Mac desktop lineup with the AI-focused M6 chip — aimed directly at the growing market of developers running AI agents locally on their own machines.
Google won a $10 million bid for 34 years of Spirit Airlines' operational data — showing that for AI giants, historical corporate logs are now a premium commodity.
Goldman Sachs is shifting its 12,000 developers from writing code to 'mentoring' AI agents — trying to bake institutional knowledge directly into the bank's internal systems.

OpenAI’s Jalapeño AI chip hits benchmarks, signaling shift to custom hardware

The quick hits

Sources