Saturday, May 30 May 30, 2026

Good morning. It's Saturday, May 30, 2026, and if yesterday's news had one overriding theme, it was this: artificial intelligence has become a dependency — for developers, for executives, and for investors — and now the bill is coming due. We've got coders who literally can't work without AI tools anymore, companies blowing through AI budgets with nothing to show for it, a chip startup raising hundreds of millions on a contrarian memory bet, and Microsoft quietly assembling an AI super app. Let's get into it.

---

The headline story today comes from the research lab METR, and it's a little startling. When the AI safety and evaluation organization set out to replicate their landmark 2025 study on developer productivity — the one that found AI actually slowed coders down when you measured objectively — they ran into a brick wall. Developers refused to participate in any task without their AI tools. Full stop. METR couldn't even run the controlled experiment because the control group no longer existed.

Instead, METR pivoted to a self-reported survey, published in May. The results? Developers perceive themselves as twice as productive with AI. But the objective picture is murkier. TechCrunch reports that Amazon shut down an internal token-tracking leaderboard called Kirorank after employees started gaming it — running up AI usage costs without generating real output. Uber burned through its entire 2026 AI budget in four months, and the company's COO recently admitted on a podcast that the spending hadn't translated into measurable productivity gains. Meanwhile a viral analysis from code-review startup CodeRabbit found that AI-generated code produces one-point-seven times more problems than human-written code — and Singapore Management University published research in April warning that AI code introduces long-term maintenance costs into real software projects. The punchline: companies are now spending roughly forty-four percent of their AI tokens just fixing bugs that AI created in the first place.

---

That dependency culture extends straight to the C-suite. TechCrunch ran a piece this week asking whether CEOs have AI psychosis — and Box founder Aaron Levie thinks most of them do. The concern is that top executives are making sweeping strategic bets on AI without adequate grounding in what it actually does, and aren't held accountable when those bets don't land.

Cognition CEO Scott Wu pushed back on the replacement framing in an interview with TechCrunch. His company — maker of the Devin AI coding agent — just raised one billion dollars at a twenty-six billion dollar valuation. Yet Wu told TechCrunch that Devin was never designed to replace humans. "We've never thought about it as replacing humans," he said. He sees agents handling the long-tail maintenance work developers don't enjoy anyway — the platform migrations, the legacy updates — freeing engineers for the creative side. He's honest about the limits though: Devin operates "somewhere between a junior and a mid-level engineer," depending on the task. And yet Cognition claims eighty-nine percent of code committed at the company itself was committed by Devin.

---

In chip infrastructure, two significant funding shots fired in the last twenty-four hours.

First: Groq is reportedly raising $650 million from existing investors. This comes after Nvidia's $20 billion not-an-acquisition deal with Groq last December — a deal that paid out investors and pulled senior executives to Nvidia, but left Groq intact as a business. Now under interim CEO and CFO leadership, Groq is doubling down on inference: the processing layer that fires every time an AI prompt is sent. Inference is where the money is in 2026, and Groq believes its homegrown LPU chip architecture gives it the edge. Existing backers Disruptive and Infinitium have agreed to backstop the round if other investors pass on their pro-rata.

Second: A startup called XCENA raised one hundred thirty-five million dollars on the thesis that the real AI bottleneck isn't compute — it's memory. While the industry conversation has focused obsessively on GPU scarcity, XCENA argues that memory bandwidth is now the binding constraint on inference speed and model efficiency.

---

On the product front, OpenAI expanded Codex's computer-use feature to Windows. The capability — which lets Codex see your screen and operate your device autonomously — launched on Mac first, and now Windows users are in. You can also manage and review Codex tasks remotely through the ChatGPT mobile app. Separately, OpenAI is quietly sunsetting ChatGPT Canvas: the side-by-side editing interface is being dropped from GPT-5.5 Instant and GPT-5.5 Thinking. Subscribers can still access Canvas through legacy models for a limited time.

Microsoft, meanwhile, is reportedly building an AI super app that rolls up GitHub Copilot, the Copilot chatbot, Copilot Cowork, and a new agentic workflow layer internally codenamed "Autopilot." Fortune broke the story. The timing is notable: Microsoft Build is next week, and this sounds like an announcement waiting to happen.

Also from Microsoft: Copilot Health is now in public preview for Microsoft 365 subscribers. The feature can analyze medical records, connect to wearables and Apple Health data, and surface insights on personal health trends — putting Microsoft squarely in the AI health assistant fight alongside OpenAI and Anthropic.

Finally, a quieter but notable trend: startups are now paying everyday people to film themselves doing household chores — loading dishwashers, folding laundry, cooking — to generate embodied training data for robotics models. The Verge reported on this yesterday. What looks like a gig-economy oddity is actually a data gold rush. Physical AI is the next frontier, and real-world demonstration data is the scarce input nobody has enough of yet.

---

That last item actually points to this morning's one business idea.

Here's the gem: build a curated marketplace for human demonstration data — specifically targeting the physical AI, robotics, and embodied intelligence space.

What it is: a platform where households, tradespeople, and physical laborers are matched with robotics companies and embodied AI labs that need high-quality video and sensor data of real-world skill execution. Not just chores — plumbing repairs, electrical work, warehouse picking, kitchen prep, medical procedures. Think Getty Images for physical motion, with structured metadata, consent management, and quality filtering baked in.

Who pays: the buyers are robotics companies, automotive OEMs building autonomous systems, surgical robot manufacturers, and defense contractors. This is enterprise B2B with meaningful contract sizes. The data sellers are gig workers, tradespeople, and eventually institutions like hospitals and culinary schools.

Why now: the window is open because physical AI is transitioning from research labs to products, demand for embodied training data is spiking, but no dominant marketplace has emerged yet. Most companies are doing one-off contractor arrangements or proprietary internal collection — expensive, slow, and non-scalable.

What makes it defensible: data quality standards are the moat. A structured taxonomy, rigorous labeling, chain-of-custody provenance, and consent documentation create switching costs for buyers and a trust premium that raw scraping cannot replicate. Once you own the quality benchmark and the supply network, newcomers have to build both from scratch.

This is a picks-and-shovels bet on the physical AI wave — one that today's Verge story confirms is already happening at the ground level, just without infrastructure.

---

That's your AI Morning Briefing for Saturday, May 30, 2026. The industry is at an inflection: dependency is total, accountability is lagging, and the infrastructure layer is still flush with capital. The next twelve months will sort the companies that measured ROI from those that tokenmaxxed into the void. Stay sharp, stay curious, and we'll see you Monday morning.