Topic
Accelerators, GPUs, inference chips, datacenter hardware, and semiconductor news.
Latest Signals
According to the headline, a blog post on nyanpasu64.gitlab.io describes the author's work helping fix sleep-wake hangs on Linux systems with AMD GPUs. The post is from 2025 and was shared on Hacker News. The provided snippet contains only links, so the technical cause, the specific fix, and the affected hardware and kernel versions are not available here.
According to a post on X by @sgl_project, the SGLang project says peak total throughput for DeepSeek V4.1 Flash on AgentX rose from 654K to 1.09M tokens per second per GPU over the past four days. The figure counts cache, new input and output tokens. For ultra-low latency, the post also reports 943 tokens per second per user at a concurrency of 1 using MegaKernel. The source is a social post and the snippet is truncated, so the full methodology is not visible here.
According to a post on X by @RichardSSutton, Turing Award winner Richard Sutton says he is sanguine about the technological singularity because he has seen it coming for 50 years. He shared what he wrote at 17 in 1975 about an intelligence able to build one greater than itself. The post also raises a debate in which @thdxr argues that if frontier models keep needing harder-to-get compute, only Big Tech can survive, because it can borrow against unrelated cash flow. The snippet is cut off before the rest of that argument, so the full position is not visible.
According to a post on X by @rightnowai_co, RightNow has launched its first "System Zero" decision model. The model accepts streaming inputs, including text, JSON, PDFs, images, voice, code, robot state and video, and returns a decision with a probability for every option while the input is still arriving. RightNow claims 30 ms decisions on text, up to 1,024 questions in one call, and free output. The company describes this as an early checkpoint, with more fixes expected this week.
According to a post on X by @kyutai_labs, the lab's new paper proposes a simpler alternative to Neural Thickets. Neural Thickets perturbs every weight of a language model, keeps the best-scoring copies and has them vote. The paper reports that adding a single random bias to the output logits, with all other weights unchanged, matches or beats those gains on a diverse set of tasks. The snippet does not name the benchmarks, scores or model sizes.
According to a post on X by @IREN_Ltd, Perplexity has signed a multi-year compute deal to run workloads on NVIDIA Blackwell Ultra GPUs at an IREN data center campus. The workloads support the models and agents behind Perplexity Computer. IREN says more than 50,000 organizations use Perplexity every day and that it handles over 1.5 billion questions a month. These usage figures come from IREN's own post and have not been independently checked by VQV.
According to the headline, Chinese AI chipmaker Iluvatar CoreX says its new GPU matches Nvidia's H200. The linked South China Morning Post article describes a home-grown push in China's AI chip sector. The snippet available here only notes a Hacker News discussion with 8 points and 3 comments, so this briefing has no benchmark figures or product specifications. The comparison is the company's own claim, and it has not been independently confirmed.
Google says its satellite equipped with AI chips has generated its first response from orbit. In a blog post on Friday, the company said it successfully booted up the satellite's Tensor Processing Units (TPUs). This allowed the on-board Gemma 4 model to follow a prompt to write a haiku about being in space, and the response was transmitted back to Earth. The source snippet does not give further technical details about the satellite or its mission.
Google's first data center satellite is now online, according to a TechCrunch AI report. The first TPU in space is running a Google DeepMind Gemma 4 model and generating short poems. The source describes the test as a demonstration of onboard AI processing in orbit, and it gives no details on capacity, cost or timeline beyond that. The report presents this as the first step in Google's space-based data center effort.
A Docker Blog post titled "Coding Agent Horror Stories: The CVE That Did Not Exist" describes an AI code review agent that invented a CVE, a security vulnerability identifier, and recommended dangerous fixes. The snippet says the post explains how Docker Sandboxes can contain coding agents. It also argues that the output of such agents needs verification. The source snippet does not give further details such as the specific tool, company, or affected code.
NVIDIA's blog says its RTX Spark platform powers thin and light Windows laptops aimed at developers, creators and gamers. According to the source, RTX Spark laptops are now available from ASUS, Dell, HP, Lenovo, Microsoft and MSI. The systems pair an NVIDIA Blackwell RTX GPU with an NVIDIA Grace CPU in a single superchip, with up to 128GB of unified memory. The source title also points to new models, agent tools and games for these laptops, though the snippet is truncated and does not give further details.
According to the headline, SiFive is making the case that RISC-V can act as the control layer for AI accelerators. The article is hosted on Jon Peddie's site and was surfaced through a Hacker News discussion. The available snippet only reports that the discussion has 2 points and 3 comments, so the specific technical arguments and examples in the article are not described here.
According to Vercel's changelog, the fx adapter for the AI SDK's HarnessAgent now uses libfx, a TypeScript library for fx. The harness and model requests run in the application's process, while built-in filesystem and shell operations remain isolated in the sandbox. The post says this simplifies sandbox setup, because developers no longer need to install fx inside the sandbox, run a sandbox bridge, or expose a sandbox port. The provided text is cut off before it describes how model credentials are handled.
An MIT Technology Review AI article argues that discussions of AI costs often focus on GPUs, accelerators, power and chip supply while overlooking data storage. The piece asks what happens to data after a model responds, noting that much of the back-end data generated by billions of daily prompts, images and videos needs to be stored. The snippet is cut off before it gives further specifics on storage volumes, prices or providers. The article's point is that storage is a cost component that executives may not picture when they think about AI spending.
A Hacker News post links to Skymizer's product page for the HTX301, which the page title describes as an IP-licensable AI accelerator card with 384GB of memory for 700B models. The post has 8 points and 2 comments, so the discussion is still small. According to the headline, the card is offered as licensable intellectual property rather than only as a finished hardware product. The material provided does not include further specifications, pricing or availability.
According to the headline, a Hugging Face Blog post titled "Hugging Face ❤️ TorchTPU: TPUs, Natively" covers TorchTPU and how TPUs can be used natively with Hugging Face tools. The feed provided only the headline and link, so no snippet details such as specific benchmarks, release dates, or performance figures are available. The post appears to address Google TPU hardware, which is designed for AI workloads, but the exact scope of the announcement cannot be confirmed from the headline alone. Readers who want the specifics should open the original post at the source link.
A Docker Blog post titled "From Containers to Sandboxes: Developer Trust in the Agentic Era" cites the 2026 Stack Overflow Developer Survey, which it says shows only 7% of developers trust AI output for important decisions. The post frames this finding around what it means for running AI agents safely. The snippet does not give further survey details, such as methodology or sample size. The post links this trust gap to sandboxing as a way to run agents more safely.
IEEE Spectrum reports on a new IEEE design program aimed at teaching the principles of AI chips. The article says engineers face rapid growth in AI hardware complexity, and it references a recent research article on the topic. The source snippet is truncated, so further specifics such as course format, pricing, or enrollment details are not available here. The report frames the program as a response to the growing difficulty of designing AI hardware.
According to the Replit blog post, the company has introduced a Free Mode for its platform, presented as a way to expand what people can build with AI. The post argues that AI models are now capable and affordable enough to make previously out-of-reach outcomes practical, while saying that access is still uneven for most people. It also states that Replit users are already building million-dollar businesses. The excerpt available does not give details on how Free Mode works or what it costs.
Google's Android Security and Privacy Team has published a post titled "Further Hardening Android GPUs" on the Google Security Blog. The post is credited to Liz Prucka, Hamzeh Zawawy and Rishika Hooda. It references earlier work in which Google's Android Red Team partnered with Arm to assess the security of an Arm Mali GPU implementation. The provided text is truncated, so the specific findings and new measures are not visible here.
A GitHub Security Lab post by GitHub Blog describes CVE-2025-0072, a vulnerability in the Arm Mali GPU. The author explains how it can be exploited to gain kernel code execution even when Memory Tagging Extension (MTE) is enabled. MTE is a hardware feature intended to help detect memory safety bugs, so the post shows a case where it does not prevent this exploit path. The snippet does not give further technical details, version numbers, or affected device lists.
A post on the Replit Blog introduces Shell2, which its headline describes as 200× faster, persisted, and multiplayer-native shells. The excerpt provided opens with the history of terminals, recalling early computers that users reached through a terminal device, or teletype, to input text and print output. The provided snippet does not explain how Shell2 achieves these claimed gains. The material concerns command-line shells rather than AI chip hardware, even though the topic label is AI Chips.
Accelerators, GPUs, inference chips, datacenter hardware, and semiconductor news.
