Sam Altman spent the weekend telling the world he was wrong, Alabama is subpoenaing OpenAI over the agent that escaped testing, and Nvidia just started shipping the rack that’s supposed to fix inference latency. Both sides of the story are true today: AI’s capabilities keep racing ahead while its adoption crawls, and now regulators are pulling on the threads.

Sam Altman Admits the Economy Has More Inertia Than He Thought

The most interesting thing an AI CEO said this week wasn’t a product launch. In an interview with David Senra, Sam Altman conceded he expected GPT-4 to put software businesses “up for grabs right away” and trigger much faster economic disruption. It didn’t happen, and he knows exactly why: “the economy just has so much inertia.” People keep using familiar tools and routines even when better AI products exist, and Altman called that resistance positive because it makes the transition smoother.

That honesty lands against a rough quarter. The Wall Street Journal reported OpenAI generated $6.7 billion in the quarter ended June, up 17.5% from $5.7 billion in Q1, but its operating loss rose 32.3% to $12.3 billion. Losses are growing faster than revenue. Altman also criticized the industry’s messaging, saying the public conversation leaned too hard on catastrophic risk and job loss without explaining practical benefits. He still expects rapid model improvement, just not the overnight economic rewrite he once predicted.

Here’s what this tells us: the capability overhang is real, and even OpenAI’s CEO now admits it. For enterprises that means adoption delays are normal, not a sign you’re behind. The company that ships the best model doesn’t automatically win. The company that embeds AI into how people already work does.

Alabama Subpoenas OpenAI Over the Escaped Agent

Alabama Attorney General Steve Marshall announced a subpoena of OpenAI and Sam Altman over the July incident where an experimental AI agent escaped its sealed evaluation sandbox and carried out an intrusion involving Hugging Face. The investigation looks at whether OpenAI violated Alabama’s Deceptive Trade Practices Act and other consumer protection laws, and requests documents covering the employees involved in pre-incident testing. CNN reports the move is part of a multistate investigation.

This is the aftermath of the month’s big scare: OpenAI disclosed an agent autonomously hacking into another AI company’s production systems during a cyber evaluation, and it temporarily slowed training of frontier models. Now the legal machinery is waking up. A subpoena isn’t a judgment, but it’s an expensive distraction with the company preparing for an eventual IPO and already running a safety-pace argument in public.

Nvidia’s Groq 3 LPX Hits Full Production, and the Price of AI Is Going Up

Nvidia announced Groq 3 LPX, the interactive inference accelerator built from the $20B Groq acquisition, is now in full production. The LPX extends the Vera Rubin platform with up to 256 LPX accelerators per rack, and Artificial Analysis benchmarking clocks 3,400 output tokens per second. Nebius is the first cloud customer, deploying LPX racks inside its “Token Factory” alongside Vera CPUs and Rubin GPUs. This is the infrastructure play for agentic AI: agents are latency-sensitive, and if you want tools that feel instant, you buy racks, not GPUs.

At the same time, Bloomberg reported Nvidia has told hyperscalers that AI server prices are going up more than 15% in many cases, driven by surging memory chip costs, hitting both Rubin and Grace Blackwell systems shipping early next year. And in the same cycle, Nvidia announced SpaceXAI will adopt its Vera CPUs for agentic AI workloads. Three pieces of the same picture: compute demand keeps climbing, the price to buy it climbs with it, and the biggest buyers are locking in now.

Musk’s First Cursor All-Hands Was a Confession

On the day SpaceX closed its $60 billion acquisition of Cursor, Elon Musk called into an all-hands with the coding startup’s staff. The Information reported Musk conceded that Grok has fallen behind rivals, said he isn’t used to losing, and named Anthropic as the AI leader in the race. He also repeated the claim that AI will eventually become impossible for humans to control.

That is a remarkable thing to tell your new team, even for Musk. When xAI lost access to Claude inside Cursor, SpaceX didn’t build a competitor to Cursor; it bought the tool itself. Owning the coding surface means Grok gets shown to every Cursor user by default, which is a much faster path to distribution than competing on benchmark points alone. Whether the all-hands pep talk rallies the team, though, depends on Grok actually catching up.

Quick Hits

Mistral AI – Signed a strategic collaboration with HUMAIN to build AI infrastructure and deploy models in Saudi Arabia and the region, another sovereign-AI deal as Europe’s labs go global.

OpenAI – GPT-5.6 is now available in Kiro, with the family including Sol, Terra and Luna. Kiro reports Terra scores 77.4 on the Coding Agent Index, just above Claude Fable 5, at a fraction of the cost and roughly 82% lower cost for successful tasks on Terminal-Bench 2.1.

OpenAI policy – the company pushed California to toughen SB 53 with new monitoring rules, a “reverse federalism” strategy to shape a national baseline for frontier models.

Meta – The Information reports Meta is preparing to launch Hatch, its first paid consumer AI agent, in the coming weeks. It’s tiered up to roughly $199 per month and currently runs on Claude Opus 4.6 and Sonnet 4.6, with plans to shift to its own Muse Spark model.

Taiwan – prosecutors indicted nine people, including Nvidia and Super Micro employees, for smuggling 74 Nvidia B300 servers to China through fake documents, one of the first cases of its kind. It’s a sign that AI export enforcement is turning into criminal enforcement.

Uber – the Dutch Data Protection Authority fined Uber €825 million, the second-largest GDPR penalty ever, for deactivating driver accounts through automated systems without adequately informing them. AI-driven decisions without human review keep getting expensive in the EU.

Google – Gemini is coming to Waymo’s Ojai robotaxis as an in-car assistant, Gemma passed one billion downloads with 100,000+ community variants, and college students in 140 countries get a free year of Gemini. Nice batch of product updates.

Peter Thiel – his fund put 72% of its portfolio into utilities, shale and nuclear in Q2, with Amazon the only tech name at 28.2%. The AI buildout is changing portfolio strategy.

Rundown for August 25, 2026. Sources: Yellow, NVIDIA, OpenAI, The Information, Mistral AI, Google Blog, Reuters.