OpenAI will burn 20% more compute to keep its models in check
Safety monitoring is no longer a side project — it's a line item. OpenAI disclosed this week that changes to its model training pipeline will increase compute overhead by 20 percent of its observed inference workload, and the company says it won't pass that cost on to customers. The figure lands in the same announcement where OpenAI confirmed it paused reinforcement learning on its latest deployment models for two weeks, after internal testing showed Astra — its next major model — may have crossed the company's "Critical" cybersecurity threshold.
The20 percent commitment is specifically for chain-of-thought monitoring: running the model's reasoning traces through safety classifiers in real time, catching misalignment signals before they compound. Ethan Mollick, the Wharton professor and AI commentator, flagged the number on social media as a signal that alignment problems are becoming "a pretty serious concern" — not a theoretical risk, but something that demands dedicated silicon. OpenAI's blog post frames the overhead as the price of doing business at the frontier: as models become more capable, the risks of developing and testing them internally also grow, so the company is choosing to absorb the compute cost rather than slow capability gains or, worse, ship without adequate guardrails.
The financial implications are real. OpenAI's inference costs were already the dominant expense in its Q2 results, where revenue hit $6.7 billion but margins continued to sink. Adding a 20 percent compute tax on top of that — for a monitoring layer that produces no direct revenue — widens the gap between top-line growth and profitability. But it also signals something about the industry's trajectory: if alignment monitoring becomes standard practice, every frontier lab will face the same overhead. The question is whether this becomes a competitive moat (safety as a differentiator) or a tax that slows everyone equally.
Alipay's "Abao" AI agent rolls out to 16 automakers in China
Alipay is bringing its agentic AI assistant to the smart cockpit. At the first Intelligent Agent Business Ecosystem Partner Conference in Hangzhou on August 17, Alipay announced that its "Abao" cross-platform service now covers 16 major automakers — including BYD, Geely, Li Auto, NIO, and Dongfeng — reaching over 16 million connected vehicles. The company also signed 60-plus new designated partnerships and released the AHA protocol, a multi-agent cross-device interoperability framework designed to let different AI agents from different manufacturers communicate and coordinate.
The AHA system is built on three core modules: an intelligent interaction protocol, an agent interconnection protocol, and a device perception-and-execution protocol. Together they enable what Alipay calls "trusted handshakes" between agents — cross-device service delivery where a user can, for example, reserve a parking spot, join a restaurant queue, and order food entirely by voice from inside the car, without touching a phone. The protocol uses domain-based authorization and data isolation to keep user data separate across agents and manufacturers.
Alipay is positioning this as the opening move in an "agent business era" for automotive. The company says over 10,000 lifestyle services have already been adapted for its AI layer, and that merchant-side agent integrations are growing fast enough that development slots are booked through next year. The play is straightforward: embed Alipay's transaction infrastructure deep enough in the vehicle that every in-car purchase flows through its ecosystem, turning the smart cockpit into a commerce surface.
Sources: The Register · OpenAI · Techmeme · Leiphone · BusinessWire
