The Commerce Department moved Monday to stop Chinese companies from reaching advanced U.S. AI chips through overseas data centers. Its interim final rule requires a license when a data center gives companies headquartered in China, Macau or other Country Group D:5 destinations access to more than 12 million TPP of aggregate compute, roughly 750 Nvidia H100s, and gives existing arrangements above that line 90 days to get one. It also adds case-by-case licensing for advanced chips shipped to Brazil, Indonesia, Malaysia, Singapore and Thailand, exempting U.S.-headquartered buyers if exporters file monthly reports, and formally rescinds the 2025 AI Diffusion Rule. The rule takes effect on Federal Register publication, expected Nov. 19, with comments due 60 days later (Federal Register)
Private Companies
Lambda Labslaunch
Lambda has finished deploying co-packaged optics (CPO) networking across more than 10,000 Nvidia Blackwell Ultra GPUs at its Quebec campus, which it calls the first large-scale CPO deployment. CPO moves the optics from plug-in modules onto the switch package. In Lambda's tests, the Nvidia Photonics switches matched pluggables on throughput and cut latency 3.8% at rack scale and 10.1% at pod scale. Optics ran at 36°C, versus 51°C for pluggables. Lambda first announced the rollout in November 2025 (Lambda press release)
NextSiliconcustomer
NextSilicon delivered four Maverick-2 accelerators to Los Alamos National Laboratory, where they sit in the Darwin testbed for emerging hardware. Lab researchers will run multi-physics simulation codes for national security science on them. Maverick-2 works alongside a host CPU and reconfigures its hardware while a program runs to fit the application. NextSilicon says the architecture already passed production acceptance at Sandia National Laboratories on unmodified code. No contract value was disclosed (NextSilicon press release)
Public Markets
Nvidia previewed Rubin HPC at the SC26 conference, a GPU with more than 50,000 CUDA cores, versus over 28,000 on the flagship Rubin, plus seventh-generation NVLink. It is expected in 2028. Nvidia reports fiscal third-quarter results after the close today (Nvidia)
Emerging
Model-specific silicon: A paper from Valavan Manohararajah and Vithursan Thangarasa, Exstill: The Most Efficient Single-Chip Processor for the Next Token, puts the energy floor for one Qwen3-1.7B token at 16K context at 0.89 mJ on 5nm logic, and measures an Nvidia B300 at 391x to 1,120x that floor. Their 5nm design hardwires one model architecture into the datapath and keeps weights and every user's KV cache on-die. After co-designing the model with the circuits, a token costs 0.52 mJ. Against a measured B300 serving the same 12 users, it decodes each user's tokens 110x faster at 690x lower energy. The design is pre-silicon and has not been fabricated (arXiv)
Offloading GPU-to-GPU traffic: A paper from Haoran Shang and 16 co-authors, SURMA: Efficient Fine-Grained Communication Offload for SuperPOD, adds a remote memory access engine to the GPU die so small transfers between GPUs, down to a 576-byte KV-cache entry, no longer tie up the GPU's streaming multiprocessors (SMs). On a two-FPGA prototype linked to two 8-GPU NVLink servers, it saturates bandwidth at 512-byte granularity with one SM, where NVLink needs over 14% of SMs. It speeds inference on four models by 1.05x to 1.26x and takes under 0.5% of GPU area (arXiv)