curated · 2026-05-30

Simon’s demo matters because Pyodide moves from “Python runs in the browser” to “Python handles web traffic.” The mechanism is concrete: a Service Worker intercepts same-origin `/app/` requests, then routes them into a Python ASGI app running under Pyodide. The old Datasette Lite path used Web Workers and navigation interception, which broke `<script>` execution and many plugins. This version runs both a FastAPI demo and Datasette 1.0a31, so it is not a one-page stunt. Claude Opus 4.8’s role is the more 2026-shaped part. It did not “build an app”; it helped an expert thread Service Workers, Pyodide, and ASGI into a working browser runtime. That smells more durable than another AI-generated CRUD demo.

SoftBank’s €75B number is huge, but I’d discount it as a data-center intent signal for now. The article only cites La Tribune and the Financial Times on AI data centers in France. It gives no power capacity, GPU order, timeline, anchor tenant, or financing mix. This smells close to the SoftBank/OpenAI Stargate playbook in the US: lead with a giant CAPEX figure, secure policy attention and power access, then fill in the supply chain later. France is a good stage because nuclear power and sovereign-AI politics make the pitch easier. The hard constraint for AI data centers is not land; it is grid connection, H100/H200/B-series delivery, and signed compute contracts. Without those terms, €75B is an expensive placeholder.

DynoSim’s sharp move is shifting inference-serving tuning from live-cluster trial and error into a virtual timeline. NVIDIA’s concrete hooks are thousands of configurations, a Rust implementation, and tests running at 1,500x real time. For a stack like Dynamo, small changes in queues, KV cache policy, batching, and routing can swing GPU utilization and tail latency together, so simulation can kill bad candidates early. I don’t fully buy the “high-fidelity” claim yet. The snippet gives no error bounds, workload distribution, GPU type, or trace size behind the 1,500x number. vLLM, TensorRT-LLM, and Triton have all been fighting the online scheduling problem; NVIDIA is pulling that decision surface deeper into Dynamo. If the fidelity holds, this is real engineering leverage. If not, it is a good-looking prefilter.

NVIDIA dropped a Computex teaser for June 2: an ARM laptop chip called N1X that packs a Blackwell GPU and AI unit into a single SoC. Both sources covering this are working off the same official teaser, so the agreement doesn't add much confidence — it's one signal, not two independent confirmations. I'd hold off on getting excited. We've got a teaser image and some media paraphrasing, but zero hard specs: no core count, no GPU CUDA numbers, no TDP, no memory bus width. NVIDIA has tried ARM PC chips before with Tegra, and it never really stuck in consumer laptops. The wildcard this time is whether Windows on ARM and NVIDIA's driver stack are finally ready. If N1X actually delivers Blackwell-class inference in a thin laptop, it lowers the bar for running local models meaningfully. But right now it's a teaser. June 2 is when we'll see if the numbers back up the hype.

This reads like a clean kill shot, but SemiAnalysis gives a verdict without reproducible evidence. The hard hook is specific: xAI dropped JAX GPU, moved to a C training framework written with Grok Build, and allegedly saw under 10% MFU on its JAX stack. The missing pieces matter: model size, GPU type, parallelism plan, batch size, and network topology are not given. Without those, under-10% MFU can indict XLA/JAX, or it can indict xAI’s own cluster plumbing. I’m more skeptical of the “vibe-coded C trainer” angle. Training frameworks are not demos; one bad collective can waste millions on a frontier cluster. PyTorch/XLA, Megatron, and Triton already showed the fight sits in kernels, scheduling, and communication, not in the language flex.