Microsoft cuts 4,800 jobs, Nvidia's Kyber slips to 2028, and Amazon kills Mechanical Turk

Microsoft lays off 4,800, Nvidia's next rack server delayed a year, Amazon kills Mechanical Turk, and HuggingFace ships LeRobot v0.6.

The big picture

A quiet but revealing batch: the big infrastructure players are either stumbling (Nvidia’s hardware delay), restructuring around AI (Microsoft’s layoffs), or closing chapters that AI made obsolete (Mechanical Turk). One robotics drop from Hugging Face and some regulatory moves round things out.

Microsoft and Nvidia both stumble on the same week

Microsoft is cutting 4,800 employees today — roughly 2.1% of its total workforce — as it enters a new financial year. The cuts land heaviest in commercial sales and Xbox. The internal memo from chief people officer Amy Coleman pointed to AI reshaping how the company needs to operate. That framing is doing a lot of work: Microsoft has been aggressively selling AI Copilot products, and if the sales motion is shrinking, it suggests either the tooling is closing deals more efficiently or those deals aren’t closing at all. Either way, this is the second significant round of cuts in 13 months (9,100 went last year), and it tells you Microsoft is seriously reorganizing around AI, not just adding it on top. The Verge

Nvidia’s Kyber NVL144 — the next-generation AI server rack — has been pushed back more than a year to 2028, according to analyst firm SemiAnalysis. The culprit is circuit board manufacturing problems. Worse, the Rubin Ultra variant has been outright canceled. Asian suppliers dropped double digits in market value on the news. This is a legitimately significant delay: Kyber was supposed to be the infrastructure backbone for the next wave of large-scale AI training. A 12+ month slip opens a real window for AMD and Google to compete on the high-end server rack market, which has been almost entirely Nvidia’s territory. Don’t write off Nvidia — but this is the kind of execution problem that gives hyperscalers a reason to accelerate their custom silicon programs. The Decoder

The end of an era: Mechanical Turk shutters

Amazon is closing Mechanical Turk to new customers starting July 30, 2026. MTurk launched in 2005 as a marketplace for tiny human tasks — transcription, labeling, content moderation — with the tongue-in-cheek tagline “Artificial Artificial Intelligence.” The irony is thick: MTurk spent two decades helping humans simulate AI, then quietly got replaced by the real thing. For ML engineers, this closes off a once-standard pipeline for cheap data labeling. If you’re still using it, plan your migration now. The broader lesson is that synthetic data generation and automated annotation pipelines have matured enough that a crowdsourced human workforce is no longer the default answer for training data. The Decoder

Regulation and robotics: two signals from opposite directions

China’s government has ordered ByteDance and Alibaba to shut down features that let users create and chat with custom humanlike AI personas — the kind of AI companion and roleplay functionality that’s grown popular in consumer apps. This is a meaningful regulatory tightening: Beijing is drawing a clear line against chatbots that blur the distinction between AI and person. If you’re building companion or persona-based AI products and you have any exposure to the Chinese market, this is a hard wall you just hit. Expect the policy to influence how other regulators think about AI identity disclosure requirements. The Decoder

On the more optimistic end, Hugging Face shipped LeRobot v0.6.0. The release is titled “Imagine, Evaluate, Improve,” which hints at capabilities around simulation, evaluation pipelines, and policy iteration for robotic learning. LeRobot has been steadily building toward a practical open-source robotics stack, and if you’re experimenting with physical AI — training policies for robot arms, mobile manipulation, that kind of thing — this is the library to watch. The source blog post was sparse on details at time of writing, so check the release notes directly before drawing conclusions. Hugging Face Blog

Quick hits

Station F, the Paris startup hub backed by Xavier Niel, is launching a new edition of its F/ai accelerator to position itself as a home base for European AI startups — worth watching if you’re building in the EU ecosystem. TechCrunch