Welcome, humans.

The AI industry spent the weekend discovering that intelligence may be abundant someday, but the boring stuff around it is still painfully scarce: memory chips, clean delegation rules, safety leadership, robotics trust, and managers who know when not to ban the calculator with a law degree.

SK Hynix warned the memory shortage could get worse in 2027 and stretch into 2030, right as AI labs, cloud companies, and device makers are trying to put models everywhere. Turns out the cloud still needs very physical little rectangles.

Here’s what happened in AI today:

• 😼 SK Hynix warned the AI memory shortage could peak in 2027 and last into 2030, keeping the hardware crunch at the center of AI economics.

• 📰 OpenAI's head of safety is leaving as the company further integrates safety and research teams.

• 📰 Microsoft is both making GPT-5.6 the preferred model in Microsoft 365 Copilot and routing some Excel/Outlook prompts to internal models to cut inference costs.

• 🍪 Mistral launched Robostral Navigate, its first robotics model for factories, warehouses, and physical automation.

• 🎓 DeepMind's delegation framework offered a practical way to decide when humans and AI agents should hand work to each other.

...and a whole lot more below.

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us! 

Love robots? We just launched a robotics newsletter! Sign up for it here.

😼 AI's Next Bottleneck Is Not the Model

Every AI demo wants you to look at the magic. Monday's news was about the plumbing.

SK Hynix said the memory shortage could be worst in 2027 and last until 2030. That matters because the AI boom runs on high-bandwidth memory, the specialized chips that feed data to GPUs fast enough for giant models to work.

Here's what happened:

• Bloomberg framed SK Hynix's huge U.S. debut as a bet that AI demand can break the old boom-and-bust memory cycle.

• Tom's Hardware reported that SK Hynix expects memory shortages to tighten next year and possibly persist through 2030.

• Rohan's Bytes reported Microsoft is routing some Excel and Outlook prompts to internal models to reduce reliance on OpenAI and Anthropic.

• OpenAI said GPT-5.6 became the preferred model in Microsoft 365 Copilot, showing the opposite pressure: customers still want the strongest model when quality matters.

Why this matters: The next phase of AI competition is about matching the job to the right resource. Sometimes that means a flagship model. Sometimes it means a cheaper internal model. Sometimes it means waiting for memory supply, power, packaging, and cloud capacity to catch up.

That is also why Benedict Evans keeps asking what happens to token pricing after the supply crunch eases. If model access becomes cheap and abundant, labs may start looking less like magical software companies and more like very expensive infrastructure providers.

Our take: The useful AI winners may not be the teams that always pick the smartest model. They will be the teams that know when to spend, when to route, when to delegate, and when the real blocker is a chip factory, not a prompt.

The future still looks automated. It also looks very dependent on procurement.

FROM OUR PARTNERS

Agent Accountability for Enterprise AI

AI agents don't just need connectivity. They need accountability. Gravitee gives every agent a verified identity, enforces what it can access, and records what it actually did. As agents move into production, "it's connected" isn't enough; enterprises need proof of what happened and why. Gravitee is the platform built to hold AI agents accountable at scale.

Explore the Platform

🎓 AI Skill of the Day: Delegate With a Fallback Plan

When you hand work to an AI agent, don't just assign the task. Assign the handoff rules.

A new DeepMind delegation framework argues that capable agents need more than task lists. They need clear authority, monitoring, validation, fallback paths, and accountability when the work changes midstream.

Try this before delegating a messy workflow:

Act as an AI delegation planner. Break this project into tasks, then decide which tasks should be handled by me, by an AI agent, or by another specialist. For each task, define: authority level, success criteria, monitoring checks, failure signals, fallback plan, and who is accountable for final approval.

Favorite insight: delegation is not a productivity hack unless someone owns the failure mode.

Want more tips like this? Check out our AI Skill of the Day Digest for July.

Have a specific skill you want to learn? Request it here. 

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. Robostral Navigate is Mistral's first robotics model, built to help robots follow plain-language navigation instructions across factories, warehouses, and industrial sites —pricing not public.

2. GitHub Spec Kit gives you a spec-first workflow for coding agents, forcing requirements, clarification, planning, and task breakdown before implementation —free/open-source.

3. Railway Agent now works from Slack and Discord, with CLI usage controls for agent and workspace access —pricing not public.

4. Microsoft Research Flint helps agents turn compact chart specs into polished visualizations —free/open-source.

5. Scarlett works inside Slack to run recurring reports, update business tools, send messages, and complete multi-step tasks across 3,000+ integrations —paid only rn ($50/mo).

6. Toyo triages Gmail and Slack, prepares you for meetings, remembers follow-ups, and can call you with updates —pricing not public.

7. PlugThis turns a plain-English idea into a working Chrome extension with source code, backend, and store-listing assets —paid only rn ($9.99/mo).

8. Agent Draw draws and updates a TLDraw canvas while you describe what you want out loud —free to try.

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube 

📰 Around the Horn

• OpenAI's head of safety is leaving as the company further integrates safety and research teams.

• Hyundai put Boston Dynamics' Atlas into a live World Cup environment, using the humanoid for halftime goal celebrations and ceremonial match-ball delivery.

• UC San Diego researchers completed preclinical surgeries with teleoperated humanoid robots, including a gallbladder removal by a surgeon-robot team.

• Indeed Hiring Lab found agentic AI may be flipping some AI-exposed jobs from destruction narratives toward new job-posting growth.

• Steven Sinofsky compared UChicago Law's AI/device ban with Harvard Law's 1982 computer ban, arguing institutions keep relearning the same technology panic.

• The Substrate forecast that China could produce a Mythos-like frontier cyber model around February 2027, while HOLA memory research proposed a cache method for improving long-context recall in efficient models.

FROM OUR PARTNERS

Agentic Capabilities in Slack

AI agents are only as powerful as the context they can access. This IDC white paper explores why conversational context is the defining variable for AI agent success, and how Slack provides the coordination layer production agents need.

Download the IDC report

😹 Monday Meme: The Anti-AI Backlash Got Organized

The AI backlash is moving from comment-section grumbling into organized activism. The Wall Street Journal reported on hard-line anti-AI activists who are ramping up for a broader fight over extinction risk, labor displacement, and the power of AI labs.

That belongs next to the law-school bans, the safety departures, and the model-access fights. AI adoption is no longer just a product rollout. It is a trust negotiation, and a lot of people are entering that negotiation with a brick in one hand and a policy memo in the other.

A Cat’s Commentary

P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you?

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.