Good morning. It's Wednesday, September 16, and we're covering NVIDIA and Salesforce’s new enterprise reasoning model, AI safety cooperation between rivals, and Meta’s expanding AI subscription push.

Plus: a new poll on AI development slowdown.

Meta launches Meta One, bundling expanded access to AI image, video, and editing tools across Facebook, Instagram, and WhatsApp. The plans build on Meta’s push to monetize its Muse AI models following a $14.3 billion Scale AI investment. Instagram Plus revenue now averages $1.2 million daily, up 475% week over week, while Facebook reaches $528,000, up 143%.

TotalEnergies and Mistral are committing more than €100 million to build frontier AI models for reservoir exploration and engineering. The three-year program combines nearly a century of geoscience expertise with almost 10 petaflops of subsurface data. A joint scientific lab will develop agentic AI tools to analyze complex data and generate exploration scenarios.

TeRAM emerges from stealth with $37 million to develop 3D SRAM for AI data centers, targeting a bottleneck that increasingly affects inference rather than training. The company says models above two trillion parameters require new approaches to memory capacity, bandwidth, and power efficiency. Its technology integrates 3D SRAM directly onto AI compute chips.

Salesforce and NVIDIA introduced Koa on September 15, 2026, a reasoning model built on NVIDIA’s open-weight Nemotron and post-trained for sales, marketing, and customer-support tasks. Salesforce says Koa uses synthetic data rather than customer information, while its token-efficient design is intended to reduce the cost of handling multi-step tasks in Agentforce.

The model gives Salesforce an enterprise-focused alternative to frontier models such as Anthropic’s Claude and OpenAI’s ChatGPT, with routing through Salesforce’s AI gateway and the company’s existing data and security controls. Salesforce is not dropping those providers: its new Claudeforce partnership lets customers use Claude while keeping enterprise data within Salesforce’s infrastructure. → Read the full article here.

OpenAI is working with Anthropic and Google DeepMind on artificial intelligence safety measures, according to global policy chief Chris Lehane. Lehane said on September 15, 2026, that discussions among the three companies had been underway for several weeks as concerns grow about AI’s potential economic and security risks.

He said OpenAI does not believe the companies need an antitrust waiver to coordinate on safety issues. The effort brings three leading AI developers into direct cooperation on safety while they continue competing in the broader AI market. → Read the full article here. (Paywall)

As models become more capable, AI leaders are arguing that safety measures need to keep pace with the technology.

MediaTek Unveils 2nm Chip: Its Dimensity 9600 Pro uses TSMC’s 2nm process for premium phones with more on-device AI.

Apple May Let Users Replace Siri: A reported Extensions Framework could let iPhone and Mac users choose Claude or ChatGPT as Siri’s AI backend.

AI Agents Get Misbehavior Hotlines: Two new services let AI agents report peers’ cheating, sandbox escapes, and unauthorized cyber operations.

SPS Commerce Expands AI Tools: New agentic AI, visibility and analytics products use network data to support faster supply chain decisions.

Blykalla, Microsoft Apply AI to Nuclear Licensing: Azure AI tools aim to streamline nuclear reactor permitting in Sweden and the U.S.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team