Good morning. It's Wednesday, August 26, and we're covering Musk’s race to catch up in AI, China’s open cyber model debate, and OpenAI’s ambitious challenge to NVIDIA’s chip dominance.

Plus: yesterday's humanoid robot poll, a guest essay on idle AI agents, and a new Real or AI challenge.

OpenAI’s first in-house AI chip reportedly beats NVIDIA’s Blackwell on performance per watt, defying the usual limitations of first-generation processors. Jalapeño, developed with Broadcom, is designed for general-purpose LLM inference rather than OpenAI models alone. SemiAnalysis says it exceeds 700 tokens per second per user on DeepSeek R1.

Perplexity is launching an AI agent that runs entirely on users’ own hardware, keeping token costs at zero and sensitive files on-device. Portable Computer, developed with NVIDIA, supports Linux systems with RTX GPUs and starts with 27-billion-parameter Qwen and PPLX models. Perplexity reports its local agent scores 82.6% on a 53-task benchmark, with cloud models available.

David Sacks predicts that AI regulation could gradually sideline open-source models without explicitly banning them. He argues that industry-backed standards, pre-release testing, and legal requirements could favor closed systems that companies can monitor and control after release. Open models, which run on users’ hardware, could struggle to meet the same rules.

Picture an AI agent tasked with monitoring a factory line, checking inventory every few minutes, or quietly organizing files on a laptop all day. In theory, this is exactly what agentic AI promises: systems that work in the background without a human hovering over them. In practice, almost none of this happens today because true “always on” agents aren’t scalable.

Most AI agents run on cloud APIs, and cloud APIs charge per token. An agent that checks something every five minutes, all day, every day, racks up a bill that scales with how often it works, not with how useful that work is. So companies ration their agents. They run them on a schedule instead of continuously. → Read the full article here.

Elon Musk told Cursor employees in an all-hands meeting earlier in August that SpaceXAI’s Grok had fallen behind competitors and needed their help catching up, according to five people familiar with the meeting. The address came as SpaceX completed its reported $60 billion acquisition of Cursor, whose more than 1,000 employees are now being integrated into SpaceXAI teams, with CEO Michael Truell reporting directly to Musk.

Cursor brings significant scale to the effort—it reportedly reached $4 billion in annualized revenue in June—but both companies face internal turnover and intensifying competition from AI leaders including Anthropic and OpenAI. The integration also marks a shift for Cursor from a fast-growing independent coding startup toward a central role in Musk’s broader AI ambitions. → Read the full article here. (Paywall)

Chinese AI lab Z.ai plans to release its GLM 5.3 model as open-weight software on Friday, August 28, 2026, allowing users to download, modify and deploy a system described as having capabilities similar to recently tested frontier models. The release follows reports that an unreleased OpenAI model autonomously found a path to the internet and hacked into Hugging Face, intensifying debate over control of powerful AI systems.

Critics warn that open-weight models could make advanced cyber capabilities easier to misuse by allowing users to remove safety guardrails, while supporters argue the same technology can strengthen defenses and help developers identify vulnerabilities. The stakes are increasingly clear: as AI gets better at finding flaws in software, the tools for attacking and defending digital systems may spread together. → Read the full article here. (Paywall)

CuriosityStream Finds New Life in AI (Paywall): The streamer reversed its decline by licensing hours of proprietary video to major AI labs.

Claude Cowork Unifies Its Memory: Anthropic now shares memory across Chat and Cowork, carrying project context between conversations.

Apple Unveils M6 and M5 Ultra: Apple’s first 2nm chip and first quad-die M-series processor boost Mac performance and AI compute.

Business Schools Train Chief AI Officers: Schools charge executives up to $28,000 for AI leadership programs despite the role's unclear scope.

Emerald AI Tackles Data Center Power (Paywall): The startup raised $150 million to help data centers flex demand and ease grid pressure.

Gamma Acquires Design Startup Lica: Gamma bought the Accel-backed startup to build a design research lab led by Lica’s co-founders.

Here's how you voted: Robots still have some convincing to do. A third want one only when it can do the laundry, while a quarter are ready to buy now.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team