SPEAKER_00 0:00

Welcome back to Sumo Growth AI Insights. I'm walking you through this week's AI news the way an operator actually uses it, not as a pile of headlines. Lead Story: a lawsuit claims OpenAI, Anthropic, Google, and SpaceX AI made an illegal AI slowdown deal. A proposed class action filed Friday in the U.S. District Court for the Northern District of California alleges those labs violated antitrust laws by coordinating to slow AI development, cutting the value paid subscribers get from ChatGPT, Claude, Grok, and Gemini. Plaintiffs point to anthropic CEO Dario Amode's September 12th essay urging labs to pace the frontier. Plus, same-day public agreement from Sam Altman, Elon Musk, and Demis Hassabis, and a July 2026 statement from lab employees about pressure not to slow alone. The suit says rivals may each choose caution but may not agree among themselves that competition is too dangerous. Defendants did not immediately comment. Plus story NVIDIA CEO Jensen Huang says there is a 0% chance AI destroys humanity by 2030. Huang rejected extinction by AI forecasts and called such warnings unsupported doomsday narratives. His comments land against recent caution from Anthropic's Dario Amode, OpenAI's Sam Altman, and Elon Musk. Huang argued governments should apply existing laws on unauthorized computer access, product liability, contracts, and damages before inventing new AI regimes. The stance also matters commercially. Nvidia supplies the GPUs behind much frontier training, so slower development or heavier new rules could hit data center demand. Three more headlines. First, Anthropic is weighing a new model as GPT-6 Astra takes about 13% of tracked enterprise AI spend. Reuters, citing three sources, says Anthropic is considering a new release to counter OpenAI's GPT-6 Astra, launched September 3rd, even after CEO Dario Amode's slowdown essay. Ramp data put Astra at about 13% of tracked enterprise AI spending versus about 8% for Claude Fable, and open router users spent more on OpenAI than anthropic models in the Cited Week for the first time in over two and a half years. Anthropic declined to comment and is evaluating safety and profitability. Second, Oaklo and X-Energy joined a $200 million federal push to power AI data centers with nuclear. The Trump administration initiative aims to speed nuclear plants dedicated to AI data centers. About $60 million over three years is earmarked for national labs and universities, including UT Austin, per Department of Energy documents cited in market coverage. Third, India says it sees no reason to pause AI research while tightening incident reporting. India's IT ministry said the country will keep building applications and models suited to domestic needs while planning tighter incident reporting rules. That is a clear counterpoint to Western pace the frontier talk. Keep building but force clearer disclosure when systems fail. Tools of the day SupersetMobile at superset.sh lets you orchestrate coding agents from your iPhone, review diffs, and keep the same workspaces when you get back to your desk. Arcjet at Arcjet.com is runtime security inside your AI code that detects prompt injection, authorizes tool calls, and blocks abuse before an action fires.work is an autonomous computer fleet that clicks and types through real apps so screen work gets done without waiting on APIs. Business AI tip of the day enforce agent actions at runtime, not in the prompt. This week's lawsuit says Frontier Labs cannot privately agree to slow competition, while Nvidia's Jensen Huang says apply existing liability law instead of freezing progress. Anthropic may still ship a countermodel to GPT-6 Astra. For agencies, the practical lesson is not pick a side, it is to ship agents that stay useful even when the policy weather changes. For every go high level or client agent you ship this week, put CRM rights, calendar invites, payments, and deletes behind server-side off. The model cannot rewrite. Authorize each tool call at runtime with a guard like Arcjet, not a paragraph of Neverdo X in the system prompt. Log the allowed action list on the opportunity so support can prove what the bot was permitted to do. Keep model routing portable across OpenAI, Claude, and backups so a slowdown, outage, or share shift does not strand clients. And use a screen driving helper like Psy only with human approval gates on irreversible steps. Agencies that treat prompts as policy will inherit the next incident. Agencies that enforce at the action boundary keep shipping while the lawyers argue. Pro tip add a webhook allow list for agent actions and reject anything not on it. Pair that with mobile review of coding agent diffs via superset mobile so humans stay in the loop when the bot asks for a merge. That is your Sumo Growth AI Insights briefing. Enforce at the Action Boundary this week, and I will catch you on the next issue.