Episode Details

Anthropic's "Pace the Frontier" essay set off a week in which nearly every major AI leader, and two heads of state, weighed in on whether the frontier should slow down. Patrick Moorhead and Daniel Newman read the incident as something larger than a safety debate. Whoever sets the pace of the frontier also sets the pace for open-weight models, enterprise custom models, and the trillion-dollar valuations riding on top of them. That contest, not the essay itself, is the through-line of Episode 320 of The Six Five Pod.

1. Anthropic's AI Slowdown: Safety or a Market Power Play? Anthropic asks for a slowdown, regulation, and antitrust waivers so labs can coordinate; Jensen Huang, Mark Zuckerberg, David Sacks, and the U.S. and Chinese governments push back, reframing the issue as ordinary product safety. Daniel argues the slowdown request reads as regulatory capture by two private labs that want to lock in a duopoly before open models and cheaper compute erode model-layer pricing power. Patrick lands on the product-safety frame: test before you ship, keep the testing procedures contained, and avoid an FDA-style regime that would stall deployment. (The Decode)

2. Trilateral safety coordination without the waiver. OpenAI, Anthropic, and Google DeepMind confirm they are coordinating on safety in the same week they disagree on whether they need antitrust protection to do it. Patrick treats the coordination as plan B after the regulatory ask failed to land, and flags the competitive risk to smaller labs and open-source developers if three frontier companies set standards together. Daniel argues self-regulation plus regulators "along for the ride" beats leading with regulation, pointing to Europe as the cautionary case for growth. (The Decode)

3. The frontier sets the pace for open models. Both hosts agree on one point last week's debate largely missed. Open-weight models from China and the UAE match frontier output at a fraction of the cost, but many are distilled from frontier models, so a frontier slowdown slows open source too. Patrick adds that these model makers are also shipping real architectural gains (an 85% KV cache reduction, on-device DeepSeek V4.1 Flash running comparably to a frontier model on Spark hardware), which drives inference cost down regardless of who leads on capability. (The Decode)

4. Dreamforce and the end of the "SaaSpocalypse" narrative. Sam Altman, Dario Amodei, and Jensen Huang all appear at Salesforce's Dreamforce, which Daniel reads as confirmation that systems of record remain the control point for enterprise AI workflows. Salesforce's Agentforce and Data Cloud lines are growing at triple-digit rates. The company is pairing a Claude-based Clawed Force with a multi-model harness, and training its own custom model, Koa, on NVIDIA Nemotron and 30 years of CRM data. Patrick sees Koa as the week's underappreciated signal: enterprise software vendors building on open weights instead of renting frontier intelligence. (The Decode)

5. OpenAI pulls even in developer traffic. OpenRouter data cited by Gavin Baker shows OpenAI climbing from roughly 20% to 50% of API share versus Anthropic since June, and Signal65's Pinnacle benchmark now places Gemini and Grok above Anthropic's latest model on several enterprise personas. Patrick credits OpenAI's early compute lock-in for the swing. Daniel's takeaway is that leadership rotates every few months, which means the model is a shrinking moat and compute, energy, and distribution are the durable ones. (The Decode)

6. AI Infra Summit: CPUs re-enter the conversation. NVIDIA presents Signal65 data on its main stage showing Vera running 1.64x faster per core than the leading shipping x86 processor, even as NVIDIA hedges