Newsletter

Nvidia lowers the memory bandwidth it lists for Vera Rubin AI systems

The NVL72 product page now rates each Rubin GPU at 19.2 TB/s of HBM4 bandwidth, down from 22 TB/s, and the rack total falls to 1,400 TB/s; Nvidia has not explained the change.

A black eight-GPU server board with finned heatsinks and NVIDIA-branded module covers, sitting on a workbench beside hand tools.
Lead image An Nvidia HGX B200 eight-GPU board, the Blackwell-generation system that Vera Rubin succeeds, photographed on a workbench. Nvidia has since lowered the memory and interconnect bandwidth it lists for Rubin. · Photo: Pokiiri / Wikimedia Commons · CC BY-SA 4.0
On this page

Nvidia has quietly lowered the memory and interconnect bandwidth it publishes for Vera Rubin, its next data-center platform. The NVL72 product page lists 19.2 TB/s of HBM4 bandwidth for each Rubin GPU, down from 22 TB/s. It also lists 3 TB/s of NVLink 6 bandwidth, down from 3.6 TB/s. At rack scale, aggregate memory bandwidth falls from 1,580 TB/s to 1,400 TB/s. NVLink falls from 260 TB/s to 216 TB/s. Nvidia’s page listed the lower figures on Sunday.

IT Home first reported the revision after comparing Nvidia’s October documentation with a copy saved in June. Wccftech then placed the edit between September 1 and September 11 using Wayback Machine snapshots. Nvidia made no announcement. The reductions are uneven. The eight-GPU HGX Rubin NVL8 board falls from 2.3 TB of HBM4 at 176 TB/s to 2 TB at 130 TB/s, a 26 percent loss. Scale-out networking rises about 12.5 percent. Memory per GPU stays at 288 GB, and the headline 50 PFLOPS of sparse NVFP4 inference is unchanged. Nvidia has not said whether the new numbers describe shipping hardware or a corrected target. That matters because Nvidia said in August that Rubin racks were already running at CoreWeave, Google Cloud and Microsoft Azure.

Featured source: TechRadar , NVIDIA Vera Rubin NVL72 specifications .

A federal judge lets the NAACP’s Clean Air Act suit against xAI proceed

Judge Debra M. Brown denied xAI’s motion to dismiss a suit brought by the NAACP through the Southern Environmental Law Center. The center says xAI installed and ran at least 35 combustion turbines at its Colossus site near Memphis without required air permits. xAI argued the turbines are mobile sources outside the permitting rules, and that the plaintiffs lacked standing. The judge found statutory standing and held that the turbines are not mobile sources under Mississippi’s implementation plan. She also rejected the argument that citizen suits violate Article II. A similar motion from the Justice Department failed earlier this month. Data Center Watch counted almost $200 billion of projects delayed, stalled or blocked in the first half of 2026.

Filed from: TechCrunch .

Binance says about 90 percent of the stolen $6.3 billion has been covered

Binance said about 90 percent of the $6.3 billion stolen from customer accounts has been covered, and it reopened withdrawals on Sunday. The exchange is covering the loss from its own funds while it buys back remaining tokens, and it says about $3 billion has been frozen. Its investigation pointed to a compromised laptop belonging to a DevOps engineer. TRM Labs and Chainalysis attribute the theft to TraderTraitor, a North Korean group.

Filed from: CNBC , TRM Labs .

OpenAI trims its latest fundraising target to $20 billion

OpenAI is seeking up to $20 billion in its current round, down from about $30 billion it had targeted earlier. CNBC and the Financial Times reported the reduction. The FT put the post-money valuation at $1.25 trillion, with investors from the United Arab Emirates leading alongside Coatue. The round is still open and terms could change, a person familiar with the deal told CNBC. The retreat followed a week in which AI-linked stocks fell and investors questioned whether heavy data center spending will pay off.

Filed from: CNBC .

An Anthropic team drops coding interviews from its hiring loop

Felix Rieseberg, a hiring manager at Anthropic who has led work on Claude Cowork and Claude Desktop, said his team removed data-structure and algorithm interviews. They gave no useful signal about candidates, he wrote in a blog post, and The Pulse reported the change on Friday. The newsletter also said incident.io has cut AI spending per developer from $4,500 a month in August to $1,600. Anthropic has not announced a company-wide change.

Filed from: The Pragmatic Engineer .

A year into Europe’s repair rules, parts are easier to find than repairs

Europe’s smartphone ecodesign rules took effect in June 2025. They require parts such as screens and batteries for seven years after a model leaves the market. A Right to Repair Europe and iFixit review of 2,334 model records in September found 80 percent lacked required repair information or parts prices. That is a share of models, not of units sold. iFixit’s Thomas Opsomer said batteries still sit behind glued displays. BEUC’s Silvia Barlassina said parts are easier to find, but it is too early to tell whether repair rates rose. One high-end screen replacement she examined cost more than €400, about 30 percent of the device’s retail price.

Filed from: TechRadar , Right to Repair Europe .

Rescue crews describe hikers arriving with confident, wrong chatbot advice

Three hikers in their twenties called for help from a rocky face on Mount Shasta. They had planned part of the trip with Google’s Gemini, according to a Wired report from High Country News. They summited at 7:30 pm, descended in the dark, and one hurt a knee. Nick Meyers of the US Forest Service said his team also met a camper who cited AI as permission to pitch beside a “No Camping” sign. Grand Teton superintendent Chip Jenkins said the share of visitors reporting AI use for trip planning rose from zero to 26 percent in a year. Wyoming rescue calls still fell, to 312 so far in 2026 from 466 in 2025.

Filed from: Wired .

From the Community

A language model that runs from a 512-byte boot sector

A developer published bootgpt, which fits a working language model in the 512 bytes of a master boot record. The assembly code loads an 11.5 KB int8 model from the next 23 sectors and runs it in 16-bit real mode on a 386. The model has one hidden layer and eight characters of context, and it generates plausible Linux boot logs for hardware that does not exist. A Python script simulates the same integer arithmetic, and its output matches the boot sector character for character.

Filed from: GitHub .

A father and son fit a handheld computer into an Altoids tin

A parent and his 12-year-old spent about two months building a pocket computer inside an Altoids tin. A free robot at the MIT Swapfest had supplied the 128x64 display. They moved to an Arduino Nano with a lithium battery and a boost converter, and the son did the CAD and 3D printing. It took five iterations before the lid closed with a real pair of Altoids inside. The display’s pinout came from a stranger on the Arduino forum, after they tried asking an AI and got a 20-pin answer for a 10-pin part.

Filed from: Altoids project .

Continue reading

Complete index →