I read a number of newsletters and Substacks about AI to try to keep up with what’s happening. One of the top questions I get asked is, “What do you read to stay current on what’s happening in AI?”
I decided to try my hand at using Claude to synthesize and track what’s happening. This is an experiment in action.
I created an agent that reviews 12 sources I’ve identified for AI news on a daily basis. It collects the full text of whatever was published that day (or most recently), then analyzes it to find patterns emerging that would otherwise be missed from reading each newsletter individually. It then applies a market research industry lens to those patterns.
I ‘m also having it track the patterns over time to see what’s emerging, what’s continuing, and what’s fading. I’ve had this running since mid-April.
I decided today to have it create a weekly synthesis arc, reviewing everything it had gathered from last Thursday and create a newsletter showing its findings, including applicability to the market research industry.
Below is the first of the weekly syntheses. If there’s ongoing interest, I’ll continue sharing these every week.
The Big Story This Week
Big AI companies stopped fighting over computer chips and started renting them from each other. Computer chips that run AI (called GPUs, made mostly by Nvidia, the largest AI chip maker) used to be a private weapon — whoever owned the most won. This week, Anthropic (the company that makes the Claude AI assistant) signed a deal to lease an entire chip farm from SpaceX (Elon Musk's rocket company), even though Anthropic's CEO and Musk publicly attacked each other a few months ago.
The story started small on Tuesday. A startup called Panthalassa raised $140 million to build floating data centers powered by ocean waves, sending results back through SpaceX's Starlink satellites. (The Rundown Tech, 2026-05-05)
By Thursday it had exploded. Anthropic agreed to lease all of a Memphis chip farm called Colossus 1 — 220,000 chips, 300 megawatts of power (about as much as a small city). (The Rundown Tech, 2026-05-07)
The same Anthropic also promised to spend $200 billion over five years on Google Cloud while keeping a separate $38 billion deal with Amazon. One AI company, three of the biggest cloud providers. (The Rundown Tech, 2026-05-07)
Musk explained the deal as renting compute "to AI companies taking the right steps" — meaning who he leases to depends on values, not just price. (The Rundown Tech, 2026-05-07)
Smaller players are escaping the giant data centers entirely. A startup called Span teamed up with Nvidia this week to mount tiny chip boxes on the outside walls of houses — 8,000 units installed, six times faster and one-fifth the cost of a normal big data center. (The Rundown Tech, 2026-05-06)
What Built Momentum
Stories that got stronger as the week went on
Big AI labs are now selling directly to businesses through investment firms
This story showed up on Tuesday as a funding announcement. By Thursday it had become a complete sales channel with real products. Anthropic and OpenAI (the makers of ChatGPT) are both building joint ventures with private equity firms — companies that own pieces of many businesses — to push AI directly into the businesses those firms own.
Tuesday: Anthropic's deal with Blackstone and Hellman & Friedman is worth $1.5 billion. OpenAI raised $4 billion from 19 investors. (Neatprompts, 2026-05-05)
Tuesday: A company called Sierra raised $950 million at a $15 billion value selling AI customer service to large companies — it already serves more than 40% of the 50 biggest U.S. companies. (The Rundown Tech, 2026-05-05)
Thursday: Anthropic shipped 10 ready-made AI helpers built for finance and insurance work — pitchbooks, screening new clients, earnings reviews, valuations — plus a Microsoft 365 plug-in and connections to financial data providers like Dun & Bradstreet, Verisk, and IBISWorld. (The Rundown Tech, 2026-05-06)
Builders agreed: the trick is the system around the AI, not the AI itself
For two years, AI work meant writing better instructions (called "prompts"). This week, multiple AI builders all said the same thing: that's not enough. The real work is building a system around the AI — memory of past mistakes, automatic checks, multiple AI helpers reviewing each other, and folders of context the AI can pull from. The idea got a name on Thursday.
Monday: The chief marketing officer at UiPath (a workflow automation company) said 70-80% of company AI experiments fail because they treat AI as a standalone tool instead of part of a larger workflow. (The Rundown Tech, 2026-05-04)
Tuesday: Five separate AI builders converged on the same conclusion — context and orchestration matter more than the AI model. (Simple AI / Dharmesh, 2026-05-05)
Thursday: A practitioner who writes the AI Maker newsletter named the idea "agent harness" — like the rigging on a horse, it's the gear around the AI that decides what gets remembered and reviewed. (Wyndo / AI Maker, 2026-05-07)
What Peaked and Faded
Stories that were loud early in the week but quieted down
AI hiring discrimination lawsuit (Mobley v. Workday) — strong Monday, gone by Thursday. A U.S. court certified a class action against Workday (a company that sells hiring software) for AI screening that allegedly rejected candidates by race, age, and disability all at once. The story came in hot Monday but no source picked it up the rest of the week. (The Slow AI, 2026-05-04)
AI products lose money when you do the math — strong Tuesday, gone by Thursday. Two newsletters independently showed that the standard "free version with a paid upgrade" software model breaks for AI, because every user action costs the company computer time. One example showed a stated profit margin of 81% turning into 45% once subsidies were stripped out. The argument got sharp on Tuesday and then no one returned to it. (Lenny's Newsletter, 2026-05-05)
What's Been Around All Month
Signals showing up week after week — the slow-building trends
The "demo to desk" gap — 7 weeks running
AI demos look amazing. Real-world reliability is patchy and sometimes random. New research from Anthropic shows that on hard tasks, bigger AI models fail more randomly, not less — meaning the failures cannot be caught by reviewing the output, because there's no pattern to catch.
This week's clearest example: Aurora (a self-driving truck company) ran more than 280,000 miles of fully driverless deliveries between Dallas and Houston with 100% on-time delivery, while Tesla passed 10 billion miles of self-driving but cars still need a human watching the wheel. Same week, two AI products, two different reliability stories. (The Rundown Robotics, 2026-05-07)
Robots and AI getting built into the physical world — 3+ weeks running
Last month robot companies were demoing single units. This week multiple companies have factories, customers, and shipping schedules. The pattern: every robot company now controls the AI brain, the physical hardware, and the training simulation all at once.
A French startup called Genesis AI showed off a robot hand that can crack eggs, play piano, and solve a Rubik's Cube. A company called 1X opened a factory in California aiming to make 100,000 humanoid robots a year by 2027. Japan Airlines started using Chinese-made Unitree humanoid robots to handle bags at Tokyo's Haneda Airport. (The Rundown Robotics, 2026-05-07)
Orchestration over prompting — 4+ weeks running
Mentioned every week since late April. This week it crossed from "good idea" to "settled doctrine," with a name (agent harness) and a working example (Anthropic's 10 finance agents). See "What Built Momentum" above.
What This Means for Research
Why any of this matters if your job involves understanding what people think or want
Market research — the work of figuring out what customers, voters, or audiences actually think — got squeezed from two sides this week. On one side, the big AI companies are now bundling research-style work (earnings reviews, client screening, valuations) directly into deals with private equity-owned businesses. That means an AI vendor's salesperson can deliver findings before a research firm even gets a chance to bid. On the other side, the consulting firm KPMG just rolled out an internal dashboard that tracks how often each employee uses AI tools — but employees already say one prompt a day counts as "active." The people doing serious system-building work and the people gaming the count look identical on the dashboard.
The deeper signal from this week's arc: AI is reshaping both how research gets sold (faster, bundled with software) and how it gets measured (by usage counts, not insight quality). Teams that build real systems — memory, checks, source-material folders — will pull ahead of teams that just buy more AI seats. The quality gap will take time to show up in any number.
Also Worth Watching
Connecticut passed a sweeping AI law covering employment and chatbots the same week Colorado moved to weaken its own landmark AI law — U.S. state rules are now moving in opposite directions at the same time. (Future Perfect, 2026-05-07)
A former OpenAI executive named Mira Murati testified that CEO Sam Altman told her the legal team had cleared an AI model to skip a safety review, which she later confirmed was false. (The Rundown Tech, 2026-05-07)
South Korea's largest Buddhist group ordained a Chinese-made humanoid robot as a monk, with custom rules combining Buddhist teachings and AI safety guardrails — a religious institution writing AI governance text. (The Rundown Robotics, 2026-05-07)
Anthropic is in talks to buy a British chip startup called Fractile to make its AI run more cheaply — silicon ownership is now reaching the AI labs themselves, not just the cloud giants. (The Rundown Tech, 2026-05-05)
Google's DeepMind (its AI research lab) bought a stake in a video game studio so it can use an offline copy of the online game EVE Online as a training ground for AI to plan over long time spans — proof that long, durable tasks remain the biggest weakness in AI evaluation today. (The Rundown Tech, 2026-05-07)
This newsletter covers May 4 – May 7, 2026. Sources: The Rundown Tech, The Rundown Robotics, The Slow AI, Lenny's Newsletter, Future Perfect, Neatprompts, Wyndo / AI Maker, Simple AI / Dharmesh. Note: no daily digest existed for May 6, and three RSS feeds (Luiza Jarovsky, Slow AI, Lenny's) were blocked late in the week — direct continuity on those sources is degraded.