Apple alleges OpenAI used its circuit schematic
OpenAI is reportedly testing outcome-based deals as Google pushes Gemini Live from voice chat into workflows.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:003:40
Read transcript
Apple’s trade-secrets case against OpenAI has sharpened, with a filing that puts an engineer’s alleged use of confidential hardware work—and an AI agent’s role in it—at the center. Apple says former engineer Chang Liu accessed a company MacBook after leaving, downloaded a circuit schematic in March, and used it in OpenAI work on power conversion. The filing says Liu ran the design through LTspice, an electrical-engineering simulator, then trained an AI agent to run simulations, inspect results, and tune a compensation parameter. Apple also alleges forensic evidence shows efforts to conceal the access. After learning about Apple’s internal investigation, Liu allegedly instructed an OpenAI colleague to destroy evidence; Apple says that colleague agreed. Those are Apple’s allegations, not established findings. OpenAI has asked the court to dismiss the suit, calling the claims meritless, and says it does not have or want Apple’s trade secrets. The immediate question is procedural: Apple wants faster access to a Mac mini and other devices still in Liu’s possession, arguing the Mac mini synced with the recovered laptop and that the material may remain in active use. Apple is also seeking an injunction barring OpenAI from using its secrets, and has sought to halt hardware-development work, including a rumored screenless device. Judge Edward J. Davila is scheduled to hear arguments in October. For builders, the important signal is how quickly agentic tooling turns a potential IP violation into an evidence, access, and governance problem. That governance question is also appearing in contracts. OpenAI has reportedly offered some large customers a model in which payment is triggered when the AI completes an agreed task, such as resolving a support request, rather than simply charging for access or usage. OpenAI declined to comment, and the arrangement has not been publicly disclosed. The appeal is obvious: customers pay for work, not tokens. The hard part is defining success and proving causation. A conversion or cost saving might also reflect marketing, seasonality, or a product change. Sierra, Fin, Cognition, Salesforce’s Agentforce, and Adobe are testing related outcome- or value-based structures, in different forms. The commercial frontier is therefore not just whether agents can act, but whether vendors and buyers can write a defensible definition of a completed result. That shift from chat toward action has a medical counterpart: researchers led through Imperial College London presented an AI system that reads a routine ECG in under two seconds. In a US trial of 67,000 patients, it identified up to 81 percent of heart-failure cases and up to 90 percent of valve-disease cases. Those are case-finding results, not diagnoses or rule-out tests. An echocardiogram is still required. The likely role is triage: flag higher-risk patients for faster ultrasound scans, potentially shortening waits that can last months. The next test is whether this fast routing layer improves care in real clinical workflows. And the same action-oriented pattern is visible in OpenClaw 2.0, a self-hosted agent platform. It adds guided model setup, a rebuilt Control UI, SQLite-backed sessions, and shared cloud sessions that let another person join or take over ongoing work. It can reuse verified Codex, ChatGPT, or Claude sign-ins, API keys, provider logins, and local models such as Ollama or LM Studio. But the release is explicit: shared-session permissions are not tenant isolation or a security boundary. Sandboxing is off by default, and host execution remains possible if the operator enables it. Across these stories, the practical watchpoint is consistent: as AI systems move from answering to acting, value depends on measurable outcomes—but trust depends on provenance, attribution, and tightly scoped access.
