The Guardrails
Beijing is unwinding the last restraints of the AI industry's most consequential forced deal reversal, while fresh capital flows into the businesses that test models and wire the data centers running them. China plans to lift its travel ban on Manus' co-founders, and CEO Xiao Hong intends to return to Singapore as the company nears the end of its separation from Meta, the Financial Times reports.
The Frontier
Frontier models can hold their own on math and code, but Moonshot AI's new benchmark suggests most of them still can't reliably see — and what gets blamed as a reasoning failure may actually start at the eyeball. Moonshot AI, the lab behind Kimi, released PerceptionBench, a benchmark that isolates visual perception from reasoning — and no frontier model clears 60 percent accuracy on it.
The Arena
The observability market just got its biggest AI-native consolidation play yet: Dynatrace is buying Arize, the startup that effectively created AI observability as a category, for $915 million. Dynatrace has agreed to acquire Arize, the AI observability and AI development lifecycle platform, for $915 million — roughly $815 million of it in cash — with the deal expected to close in the third quarter.
The Frontier
A Chinese university team just proved that an open-weight model plus strong agent scaffolding can out-hack the labs that built the models. Fudan University's Whitzard agent jumped to No. 2 globally — and No. 1 among universities — on CyberGym, the AI security benchmark, cracking 91.2% of the real-world vulnerabilities it faced. The new leaderboard, released this week, puts the team led by Professor Yang Min's lab in the world's first tier of AI attack-and-defense research.
The Frontier
Two research stories this morning, one about what open models are trained on, one about what self-improving agents remember. Denmark's Foundation Models project says a 1-billion-parameter model built only on legally clean data goes toe-to-toe with models four times its size, and a new study quantifies how easily agents bake unsafe behavior into their persistent skills. Danish researchers say their 1B model, trained only on permissible data, matches rivals four times its size.
The Guardrails
Big Tech is about to spend a trillion dollars this year on AI data centers. The problem: money is the only input that isn't scarce. Goldman Sachs now expects global AI data-center spending to hit $1 trillion in 2026 — yet the build-out is colliding with bottlenecks that cash can't buy its way past, from power grids to skilled electricians. JPMorgan forecasts $697 billion in US spending alone, and Bank of America models a path to roughly $1.2 trillion by 2027.
The Guardrails
Washington is drawing a hard line under the global AI cold war: partner with the US-led Pax Silica initiative, or with China's rival framework — not both. The US plans to tell 35 partner countries they will be excluded from the US-led Pax Silica initiative if they join China's rival AI framework, according to a Reuters exclusive based on a letter and a source familiar with the matter.
The Arena
The AI trade just produced its first big named casualty. Jane Street, one of the world's most formidable quant firms, lost roughly $15 billion in July — a rare and unusually specific disclosure from a house that almost never talks about its numbers. Jane Street posted a roughly $15 billion loss in July after getting wrongfooted by the AI-stock reversal that also sank Leopold Aschenbrenner's hedge fund Situational Awareness.
The Arena
A money day for the frontier: Anthropic finally turns an adjusted profit as it preps what investors expect to be a record IPO, and Nvidia's latest filing reveals a $50-billion-plus war chest of stakes in Musk's empire. Anthropic told prospective investors it generated more than $11.5 billion in revenue in Q2 with positive adjusted operating income — its first profitable quarter.
The Arena
Two very different AI frontiers moved today: California finally let autonomous trucks onto its public highways, and a ChatGPT core contributor pulled in one of the biggest Pre-A rounds in Chinese embodied AI. California's DMV has granted testing permits to Aurora Innovation and Kodiak AI, clearing autonomous heavy trucks to operate on public roads in the state for the first time — and Kodiak has already started, running a handful of test trucks around its Mountain View office.
Quick Hits
An evening catch-up: Anthropic nudges its misalignment risk up a notch, Thrive's Joshua Kushner tells Silicon Valley to cool it on AI, and Z.ai opens its vulnerability books. Anthropic raised its misalignment risk estimate from "very low" to "low" and says it has no plans to release its stronger internal model, "Model 2." In its August risk report, the company cited recent cybersecurity incidents for the downgrade and flagged "signs of acceleration" in models' ability to conduct automated R&D…
The Arena
OpenAI's business just crossed a milestone it didn't expect to reach until year-end: enterprise now generates more revenue than the ChatGPT-led consumer business, CFO Sarah Friar told investors on Friday. Friar said the two revenue lines "have now crossed," with enterprise accounting for the majority of OpenAI's revenue — months ahead of the company's own forecast.