Happy Friday, here is your weekly AI rundown 👋

The story this week is the supplier problem. Nvidia's biggest customer built a chip that beats it. Amazon shut down the human marketplace that trained the models now replacing it. And in Beijing, a robot ran the 100 metres faster than any human ever has. Everyone's climbing the stack — and eating whatever's above them.

At the Hot Chips conference, OpenAI published the first benchmarks for Jalapeño, its custom inference chip — and it posted 1.5 to 1.9× more AI work per watt than Nvidia's Blackwell systems, with far lower latency. Built with Broadcom on TSMC and taped out in just 16 months, it's OpenAI's answer to a $1B+/month Nvidia bill. Two honest caveats: Jalapeño uses newer HBM4 memory, so the truly fair fight is Nvidia's Rubin (also shipping now), and it only reaches real volume in 2027. But the signal is unmistakable — Nvidia's biggest customers (OpenAI, Google, Amazon, Meta, Anthropic) are all now building the silicon they used to buy.

Amazon is shutting down Mechanical Turk on September 30, ending the 21-year-old marketplace Bezos once called "artificial artificial intelligence." MTurk paid armies of people pennies to label images, transcribe audio, and rate results — the invisible human labor that trained the modern machine-learning era. It's a fittingly circular ending: the humans who taught the models are being retired by the models they taught.

Our take: the theme tying these together is verticalization. When the frontier gets expensive enough, everyone stops renting and starts owning — their chips, their data pipeline, their whole stack. The companies that sold picks and shovels are watching their best customers open their own mines.

Meta's $16.68 billion bill for hooking teens. Meta agreed to pay up to $16.68 billion to settle claims from 29 states that it deliberately engineered Instagram and Facebook to be addictive to children and misled the public about it. As part of the deal, Meta must add daily time caps and overnight limits for teens, tighten age checks, and submit to an independent auditor — while, naturally, admitting no wrongdoing. The largest settlement of its kind, paid out over a decade. Engagement optimization, meet the invoice.

OpenAI's own chip just beat Nvidia's on efficiency, and every big lab is building silicon. Is Nvidia's grip finally slipping?

You know ChatGPT and Claude. Here are five more tools worth adding to your kit:

This week's spotlight: Tiangong Ultra. At Beijing's World Humanoid Robot Games, the 1.8-meter, 55kg robot from the Beijing Humanoid Robot Innovation Center ran the 100 metres in the 8-second range — comfortably faster than Usain Bolt's 9.58s human record — improving its own time race after race across the Games. Yes, it needed a crash mat to stop and one run ended in a small torso fire, but the direction of travel is the story: China is iterating humanoid hardware in public, at competition speed, while most Western robots are still doing careful demos. Watch the full race below.

New arrivals in the AI Graveyard this week:

Want the autopsy? We dug into what actually kills AI tools — funding, acquisition, or just a dead domain — in our data report.

The ghost workers who built AI

Mechanical Turk shutting down is a footnote for most people, but it deserves a moment. For two decades, the "intelligence" in a lot of early AI was quietly human: people in Kenya, India, the Philippines, and rural America, paid cents per task to draw boxes around pedestrians, flag graphic content, and rate whether an answer was any good. Every model we now call miraculous was taught, in part, by workers whose names appear in no research paper and whose wages rarely cleared a few dollars an hour.

The uncomfortable truth is that this labor never actually went away — it moved and got renamed. Today it's "RLHF annotators" and "red-teamers" and "data raters," often through outsourcing firms, still doing the invisible work of making models seem effortless. The difference is that the frontier labs are worth trillions now, and the people teaching their models still mostly aren't.

You don't have to romanticize a gig platform to notice the asymmetry. As AI eats more knowledge work, the least-visible knowledge workers — the ones who labeled the training data — are worth remembering. If the machines are standing on anyone's shoulders, it's theirs.

With a third of new web pages now AI-written, telling real from filler is a survival skill. Next time a source looks off, paste it in and run this:

On the calendar:

Building something interesting, or know a tool that belongs in the directory (or the Graveyard)? Hit reply, or email [email protected] — we read everything. (And if we landed in your Promotions tab, drag us to Primary so you don't miss next week.)

— The ToolDirectory.AI team