Thursday, June 25 June 25, 2026

Good morning. It's Thursday, June 25th, 2026. I'm your MorningAI brief. Today the story is silicon, sovereignty, and who gets to own the future of AI infrastructure. Let's get into it.

The biggest news of the last 24 hours: OpenAI and Broadcom have unveiled Jalapeño — OpenAI's very first custom AI chip. This is a reticle-sized ASIC built specifically for large language model inference. Not training, inference — the part that actually runs ChatGPT for hundreds of millions of users every single day. What makes this remarkable is the speed: nine months from concept to silicon. That is an extraordinarily fast development cycle for custom hardware. OpenAI calls it part of their plan to "build the full stack," meaning they no longer want to depend entirely on Nvidia for the compute that powers their products. Broadcom brings the silicon design expertise; OpenAI brings the model-specific workload knowledge. The chip is optimized for the attention layers and token generation patterns that define modern LLMs — and it's built with future agentic workloads in mind. This is a direct shot across Nvidia's bow, and it signals that the AI infrastructure race is no longer just a software game.

Speaking of OpenAI's ambitions — a second major story dropped in the last 24 hours that you need to hear. CNBC is reporting that Sam Altman and the White House are in active talks about a potential US government equity stake in OpenAI. Yes — the federal government taking an ownership position in the world's most prominent AI company. No deal is done, and details are sparse, but the fact that these conversations are happening at all tells you something about how Washington now views AI leadership. This is not just industrial policy — this is the government treating AI capability as a national strategic asset. If it goes through, it would be one of the most unusual public-private arrangements in modern tech history.

On the model front: Claude Fable 5 officially exited its free window on June 23rd. After 13 days of open access, Anthropic's most capable model is now behind a paywall, available only to Pro, Max, Team, and Enterprise subscribers — and even then, it costs usage credits at ten dollars per million tokens. Meanwhile, ByteDance is rolling out paid subscriptions for Doubao, China's largest AI chatbot. The global AI monetization playbook is converging fast: free tier to hook users, paid tier to fund compute.

Over at MIT, researchers published a paper today on a system called Murakkab — an automated framework that streamlines the design of agentic AI workloads and optimizes deployment for real-world applications. Agentic AI — where multiple models chain together to complete complex multi-step tasks — is notoriously inefficient and expensive. Murakkab aims to cut both latency and energy consumption. As AI agents move from demos into enterprise software, efficiency breakthroughs like this become critical infrastructure.

On the policy side, Google published a comprehensive white paper pushing a middle-ground approach to AI regulation. Rather than backing either a strict liability framework or a self-regulation-only stance, Google is arguing for a targeted, risk-tiered model — regulate the highest-risk deployments, leave lower-risk applications relatively open. This is clearly a lobbying move, but it's also an attempt to shape the narrative before Congress acts. Meanwhile, the FedScoop is reporting that federal agencies are being warned: just having a human in the loop is not sufficient AI governance. Oversight needs to be meaningful, documented, and auditable — not just technically present.

And in the background, a theme that doesn't go away: AI and the 2026 midterm elections. WUSF is reporting that AI-generated video and deepfakes have gotten good enough that even experts are struggling to distinguish them from real footage. At least several campaigns are already using AI-generated content. The disinformation risk is real, and the guardrails are still thin.

Today's business idea: an AI inference cost benchmarking service for enterprise buyers. With OpenAI launching custom silicon, Nvidia dominant, and AMD and Google TPUs all vying for enterprise contracts, most companies have no reliable way to compare actual inference cost-per-token across providers for their specific workloads. A SaaS tool that runs standardized benchmark suites against each provider's API, tracks pricing changes weekly, and delivers cost optimization recommendations could command serious B2B revenue — and the Jalapeño announcement just made the conversation a lot more urgent.

That's your MorningAI brief for June 25th. Stay sharp out there.