Newsletter

Nvidia stops shipping Hopper chips to China as revenue doubles

Third-quarter revenue reached $115.8 billion on data center sales of $107.6 billion, and guidance for the current quarter assumes no data center compute revenue from China.

Four NVIDIA H100 accelerator cards seen edge-on, their cream-coloured fin stacks and black NVLink connector bars filling the frame against a blurred desk.
Lead image NVIDIA H100 accelerators, built on the Hopper architecture that also includes the H200 chips Nvidia has stopped shipping to China. · Photo: 极客湾Geekerwan / Wikimedia Commons · CC BY 3.0
On this page

Nvidia has stopped sending its licensed Hopper data center products to China. Interim chief executive and finance chief Colette Kress disclosed the change on Tuesday’s earnings call. The guidance issued with the results assumes no data center compute revenue from the country. Nvidia had won White House approval in December last year to sell H200 chips there. Jensen Huang, who left the chief executive role last week, had argued for the business.

Revenue rose to $115.8 billion, up 103% from a year earlier, with data center sales of $107.6 billion. Nvidia guided the current quarter to $130 billion, give or take 3%. Kress did not explain the halt or say whether other China business would continue.

Featured source: The Verge , Nvidia results .

Oracle patches critical Java flaw tracked as CVE-2026-114938

Oracle’s November security update fixes a flaw in Java SE, GraalVM for JDK and GraalVM Enterprise Edition. The CVE record scores it 8.1 and describes an attack reachable over a network without credentials, in the compiler component. Oracle rates exploitation as difficult. Researcher Quang Luong of Calif.io, who is credited in the advisory, publicized the bug as “Java4Shell”. He reported remote code execution in enterprise products including Jenkins, Jira and Keycloak, a claim not independently confirmed.

Filed from: TechRadar , Oracle advisory .

Compromised Optimizely scripts served malware to 178 sites

JavaScript signed by Optimizely and delivered from its content network had been altered to attack visitors. It swapped cryptocurrency wallet addresses, skimmed payment cards, copied cookies and screenshots, and pushed iOS exploit kits. ReportURI, which found 178 affected sites including CNN, the BBC and Salesforce, said the code came from Optimizely’s own infrastructure. It carried a valid certificate, so this was not a poisoned cache. How attackers obtained write access to so many customer projects is unknown, and Optimizely did not answer questions.

Filed from: Ars Technica .

Microsoft moves Office inside Copilot and gives agents their own PCs

At Ignite, Microsoft made Word, Excel and PowerPoint generally available inside Copilot. Satya Nadella called Copilot “the new OS for work.” Microsoft also introduced Copilot Autopilot, an agent that can be given standing responsibilities rather than single tasks. Agents can now be assigned managed Windows Cloud PCs with their own Entra identities, controlled through Intune. Nadella said in the keynote that AI systems cannot be trusted to govern themselves. Limits on what they may do, he argued, matter more than model behaviour.

Filed from: Ars Technica , Microsoft Azure blog .

Amazon and Google will hold nearly 30% of Anthropic after its IPO

Anthropic’s filings put Amazon’s stake at 14.7% and Google’s at 13.7%, both in non-voting shares. Amazon is holding back notes that would lift its holding to 23.8%. A group including AMD, Broadcom, Micron, Microsoft, Nvidia and TSMC has agreed to buy $26 billion of the non-voting stock. Google and Nvidia take $10 billion each. AWS also expects to earn from Anthropic’s commitment to spend more than $100 billion on cloud services over a decade.

Filed from: TechCrunch .

Molten Industries starts making battery graphite from methane

Molten Industries has begun running a demonstration plant in Oakland. It splits natural gas into carbon and hydrogen, producing about 300 metric tons of battery-grade graphite a year, roughly enough for 300 megawatt-hours of cells. Benchmark Mineral Intelligence estimates Chinese firms control about three-quarters of the graphite anode supply chain and 97% of synthetic anode output. Molten says its graphite matches Chinese prices once tariffs are added, a company estimate that has not been audited.

Filed from: TechCrunch .

Osmo says its model matches people at recognising smells

Osmo, founded in 2023 by former Google Brain researcher Alex Wiltschko, claims its foundation model for smell performs like an average person at telling odours apart. It says the model composes scents about as well as a junior perfumer. The company has assembled more than 10 million human labels, and says fragrance and consumer-goods firms already use the system. Smell scientist Avery Gilbert has called parts of the marketing undocumented. Other researchers note that human olfaction leans on context the model does not capture.

Filed from: MIT Technology Review .

Humyn Labs raises $47 million to film everyday tasks for robots

Bengaluru-based Humyn Labs has raised $47 million, split between $26 million in equity led by Accel and $21 million in venture debt. The company records household and workplace chores with two-camera rigs, then labels the footage as training data for robotics and world models. It works across more than 20 countries. It reports 12 paying customers, $10 million in revenue between April and October, and about 90% of sales in the United States.

Filed from: TechCrunch .

From the Community

Cerebras splits inference between GPUs and its wafer-scale chips

Cerebras says disaggregated inference is now in production. AMD Instinct GPUs handle prompt processing while its CS-3 systems generate responses. Separating the two stages lets the wafer-scale chips spend their capacity on token generation, where their on-chip memory helps most. Cerebras reports five times more tokens per megawatt than a CS-3-only setup, with no loss of generation speed. The figures come from its own deployment.

Filed from: Cerebras .

Energy-based reasoners cut Sudoku inference compute by 345 times

A research project describes recurrent reasoning models that learn an energy function over latent states. The models refine candidate answers by descending on that function and stop when the gradient is small. With 1.8 million parameters, the authors report 99.9% accuracy on Sudoku-Extreme using 345 times less inference compute than the next-best model. They also report strong coverage on N-Queens and graph colouring. The results come from the project’s own benchmarks and have not been peer reviewed.

Filed from: Energy-Based Reasoners .

A Berlin guide site says AI Overviews erased three-quarters of its traffic

Nicolas Bouliane, who runs the relocation guide All About Berlin, reports that visits fell by about 75% after Google’s AI Overviews began answering questions his pages once served. He links the decline to search results that summarise his material without a click. He describes rewriting pages and cutting costs in response. The figures come from one site owner’s own analytics rather than an independent study.

Filed from: All About Berlin .

Continue reading

Complete index →