New York City just drew the hardest line any American school system has drawn on generative AI, UiPath proved the automation business still grows — and still can't convince the market — and Nvidia quietly turned a hardware constraint into a model design rule.

New York City has imposed a one-year ban on student-facing generative AI through eighth grade, the broadest such restriction in the country. Mayor Zohran Mamdani and Schools Chancellor Kamar Samuels announced the moratorium on September 2, covering roughly 600,000 students — about two-thirds of the district's enrollment — and halting the use of around 40 educational tools currently in classrooms. Companion chatbots are banned at every grade level. High schoolers get something softer: five capped pilots for at most 50,000 students in general education classes, plus two 45-minute critical-thinking modules a year covering bias, risk, and how AI shapes careers. Teachers keep using AI for planning and scheduling, and students with disabilities, multilingual learners, and career-readiness participants are exempt. The city also added screen-time caps — 30 minutes a day for grades 3–5, 45 minutes for 6–8.

The framing is the story, not the fine print. Mamdani didn't argue the technology is dangerous; he argued the premise is wrong. "The tech industry wants us to believe that A.I.-powered early education is not only inevitable, but necessary. We do not see it that way," he said, positioning the year as a study period rather than a permanent rejection. A Technology in Schools Coalition of students, parents, educators, and union partners will publish recommendations for 2027–28.

That distinction matters for everyone selling into education right now. This isn't a Luddites-versus-progress story — it's a procurement story. Every tool used in the system now faces a review that will bar anything not judged mission-critical to learning, with vendors held to safety, transparency, and evidence-of-impact standards. Google, Anthropic, and Microsoft didn't respond to requests for comment. Los Angeles is reportedly reviewing its own AI policy, and if the largest district in the country can pause for a year and demand proof, the vendors who can't produce it just lost their easiest growth market.

UiPath beat on every headline number and watched its stock fall more than 7% after hours anyway. The automation company reported fiscal 2027 second-quarter revenue of $410 million, up 13% year over year, with annual recurring revenue of $1.94 billion, up 12%, and net retention rising to 109%. GAAP operating income came in at $32 million against a $20 million loss a year earlier — a real swing into profitability. It raised full-year revenue guidance and promoted chief accounting officer Hitesh Ramani to CFO. The shares initially jumped more than 10% before reversing.

The reversal is the tell. UiPath is up more than 70% over the past year but only 12% year to date, as investors weigh whether AI agents eat the company's customer base. Founder and CEO Daniel Dines argues the opposite: that AI expands what enterprises want automated while increasing demand for "orchestration, governance, and exactness" that deterministic automation provides — agents, robots, systems, and people running end-to-end processes together. The quarter offers some support: net new ARR of $37 million and a retention rate above 100% mean existing customers are spending more, which is harder to fake than a new-logo surge.

Our take: this is the central unresolved question in enterprise software, and UiPath is the cleanest public test case. A 13% growth rate at a $9.3 billion market cap is not a disruption story — but it isn't quite a durability story either. The company's argument is that agents need someone to govern them, and that's a real job. Whether customers pay for it separately, or expect it bundled into the agent platform they're already buying, is what the stock is arguing about at 7% down.

Nvidia published an official guide to speculative decoding that quietly turns a GPU constant into a model architecture decision. The September 2 technical blog post lays out six leading acceleration approaches with training costs and use cases attached, then makes an unusual leap: it derives the exact draft length a model should target from the hardware's attention tile size. Because the verification pass processes one real token plus D drafted tokens, and the GPU groups query and KV heads into blocks bounded by a tile size of 128, the constraint works out to G × (1 + D) ≤ 128 — meaning a model with 8 attention groups should draft 15 tokens, while one with 32 should draft three. Guess wrong and you pay for a full tile you only half-filled.

The practical payoff is large: DFlash, a diffusion-style approach from Jian Chen, Yesheng Liang, and Zhijian Liu, generates all seven draft tokens in a single forward pass rather than one at a time, and reaches roughly 6× lossless speedup — the output is bit-identical because the big model only accepts tokens it would have produced itself. EAGLE-3, built by researchers including Peking University's Yuhui Li and the University of Waterloo's Hongyang Zhang, is flagged as the best current choice for large models on GPUs, with the caveat that it has to be trained alongside the main model during pretraining. Nvidia's guide also puts a price on the oldest method: training a separate external draft model from scratch costs one to ten trillion tokens.

The significance isn't any single technique — it's who's writing the rules. A chipmaker has published a document telling model designers how many attention groups to pick based on the physical limits of the silicon their models will run on. Inference optimization used to be something engineers bolted on after training finished. Now it's an architecture decision made before the first token, and the labs that internalize that constraint will ship cheaper tokens than the ones that don't.

What to watch: the Technology in Schools Coalition's 2027 report — it will decide whether this moratorium was a pause or a permanent new baseline for how American schools buy AI.

Is a one-year classroom pause a responsible brake, or a year of lost ground for the students who need these tools most? Tell us in the comments.