AI Roundup: OpenAI Tests Jalapeño Chip, New Physics AI Startup Debuts, and Perplexity Teams with Nvidia

AI Roundup: OpenAI Tests Jalapeño Chip, New Physics AI Startup Debuts, and Perplexity Teams with Nvidia The latest edition covers benchmark results from OpenAI's custom Jalapeño inference chip, the emergence of a physical-AI startup using neural operators, Perplexity's on-device agent for Nvidia hardware, and several short industry briefs. OpenAI published benchmarks for its 700-watt Jalapeño inference chip, reporting higher speed and power efficiency than Nvidia's 1,200-watt systems. Caltech researchers launched Accelerated Understanding to simulate physical processes via neural operators instead of traditional transformers. Perplexity and Nvidia released Portable Computer, an on-device agent operating locally on Nvidia's DGX Spark desktop supercomputer. Briefs note executive departures at OpenAI, an Apple Mac Mini refresh, and enterprise developments from Anthropic and Google Cloud. OpenAI Jalapeño Benchmark Results OpenAI published initial internal benchmark results for Jalapeño, an in-house inference chip developed alongside Broadcom. According to OpenAI, the 700-watt chip delivered up to 3.6 times faster response times and 1.9 times greater efficiency per watt compared to 1,200-watt Nvidia hardware. OpenAI credited its Astra model and Codex with designing the hardware within nine months. Hardware VP Richard Ho stated the chip is dedicated entirely to internal inference needs, with deployment planned for data centers later this year and production expanding through 2027, while Nvidia hardware remains in use for training. Accelerated Understanding Launches Physical AI Models Caltech professor Anima Anandkumar and engineer Benedikt Jenik launched Accelerated Understanding, an enterprise-oriented startup designed to predict physical world dynamics rather than text sequences. Utilizing physics-based neural operators to track events across 3D space and time, the startup's model processed 5 trillion data points in a single benchmark run. The founders turned down roles and a 35% stake in Jeff Bezos' Prometheus venture to pursue their approach targeting robotics, extreme weather forecasting, and chip materials. Perplexity and Nvidia Unveil Portable Computer Perplexity teamed with Nvidia to release Portable Computer, an agent that operates directly on Nvidia's $4,699 DGX Spark hardware. The tool runs locally without using cloud credits, allowing users to choose between Qwen 3.8 27B and Perplexity's PPLX 27B models while retaining the option to connect to external cloud models with explicit approval. The launch follows reports that Nvidia is in talks to invest billions in Perplexity at a valuation surpassing $30 billion. Industry Quick Hits Anthropic introduced a unified Claude memory spanning standard chat and Cowork, while reports indicate the company plans to pitch a $30 trillion addressable market to IPO investors. Apple unveiled an $899 Mac Mini powered by M6 silicon for agentic workloads. Google Cloud previewed legal and financial editions of Gemini Enterprise, and OpenAI head of data centers Chris Malone reportedly left the company amid multiple executive changes.

OpenAI 公布首款自研芯片 Jalapeño 基准数据，行业迎来端侧与物理 AI 新突破 本期简讯涵盖多项重磅 AI 动态：OpenAI 披露其与博通联合研发的首款内部推理芯片 Jalapeño 测试成绩；新创公司 Accelerated Understanding 探索基于物理特性的“神经算子”模型；Perplexity 与英伟达联手推出本地端智能体系统。 OpenAI 联合博通研发的 700 瓦 Jalapeño 推理芯片基准测试显示，速度比英伟达旗舰系统快达 3.6 倍，能效提升高达 1.9 倍。 新初创公司 Accelerated Understanding 放弃传统 Transformer 架构，采用“神经算子”专注于真实物理世界预测，单次处理达 5 万亿数据点。 Perplexity 与英伟达联合发布 Portable Computer，可在英伟达 DGX Spark 硬件上免费、私密运行本地端 Agent 模型。 行业速报涵盖 Anthropic 统一 Claude 记忆系统并展望 IPO 市场前景、苹果推出搭载 M6 芯片的 Mac Mini、谷歌上线金融与法律行业定制版 Gemini 等。 OpenAI 自研芯片 Jalapeño 展现能效与速度优势 OpenAI 公布了首款用于运行（而非训练）AI 模型的自研芯片 Jalapeño 的基准测试数据。该芯片由 OpenAI 与博通合作开发，借助即将发布的 Astra 模型及 Codex 在 9 个月内完成设计到量产准备。在测试中，这款 700 瓦芯片相比英伟达 1200 瓦方案，回答响应速度最高达 3.6 倍，单位功耗工作量达 1.9 倍。该芯片仅供 OpenAI 内部使用而不对外销售，后续两代产品已在开发中，预计将于今年晚些时候进驻数据中心并在 2027 年前逐步扩大产能；模型训练目前仍依赖英伟达。 Accelerated Understanding 聚焦物理 AI 新架构 加州理工学院教授 Anima Anandkumar 与工程师 Benedikt Jenik 创立了新公司 Accelerated Understanding。与预测下一个词的 Transformer 模型不同，该公司采用“神经算子”架构模拟 3D 时空变化，在测试中单次运行处理了 5 万亿数据点。两位创始人曾拒绝杰夫·贝索斯旗下 Prometheus 提供的核心职位与 35% 股权，未来商业方向将聚焦芯片材料、极端天气预测和机器人等企业级物理领域。 Perplexity 携手英伟达推出本地端 Agent 设备 Perplexity 与英伟达联合发布了 Portable Computer，将 Computer 智能体搬到本地硬件上运行。该系统可在英伟达售价 4,699 美元的 DGX Spark 上一键安装，支持 Qwen 3.8 27B 或 Perplexity 自研的 PPLX 27B 模型，且本地运行不消耗积分、确保数据隐私，必要时也可经用户确认调用 15 款以上的云端模型。 行业综合动态 其他主要进展包括：Anthropic 推出跨聊天与 Cowork 的统一 Claude 记忆功能，并在筹备向 IPO 投资者宣介其超过 30 万亿美元的潜在市场空间；苹果发布搭载 M6 芯片、售价 899 美元的新款 Mac Mini，主打全天候智能体计算；谷歌云上线面向金融与法律领域的定制 Gemini Enterprise 预览版；OpenAI 数据中心负责人 Chris Malone 确认离职。

[Read Online]({{live_url}}) ^| [Sign Up](https://www.therundown.ai/subscribe?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) ^| [Advertise](https://therundownai.typeform.com/to/kraZ1TSO?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) **Good morning, {{ first\_name | AI enthusiasts }}, and welcome to our 3,221 new readers.** OpenAI named its first in-house AI chip Jalapeño, and the first benchmarks just came out appropriately spicy. The company's internal tests have the chip beating Nvidia's flagship GPUs on both speed and power efficiency, with a design OpenAI credits its Astra model and Codex for helping create in nine months. **In today’s AI rundown:** OpenAI's Jalapeño chip brings the heat New startup takes AI from words to physics Build a website hands-free with Claude Voice Perplexity, Nvidia go portable with Computer **LATEST DEVELOPMENTS** Image source: OpenAI **The Rundown:** OAI just [published](https://openai.com/index/jalapeno-first-results/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) the first benchmark results for Jalapeño, the custom chip it built with Broadcom to run AI models (not train them), with its tests showing wins over Nvidia's flagship systems on both speed and power efficiency. **The details:** In tests, OAI's 700-watt chip topped Nvidia’s own options rated at 1,200 watts, answering up to 3.6x faster with up to 1.9x more work per watt. The company said it leveraged its own upcoming Astra model and Codex to create Jalapeño, taking nine months from first design to manufacturing-ready. OAI won't sell the chip, with hardware VP Richard Ho saying it has "so much need for it" internally, and still leans on Nvidia to train new models. Two more generations are already in the works, with plans to reach OAI's data centers later this year and production set to ramp up through 2027. **Why it matters:** Google's in-house TPUs have been a good example of what a lab gains from owning its silicon, and the club is filling up fast with Anthropic, Amazon, and Microsoft also planning their own designs. If Jalapeño's numbers hold, OAI gets cheaper, faster tokens and a tighter, personalized design loop for each model after. **The Rundown:** Physical AI is the next frontier, but complex real-world variables and multimodal experimentation make building embodied systems a challenge. Our new e-book, Advancing Physical AI: From Learning to Embodied Intelligence, shares practical strategies to help you optimize robotics workflows and accelerate development. **Get the guide to learn:** Techniques to close the sim-to-real gap through real-time iteration Steps to ensure the safety of physical AI systems How to train VLAs at scale Image source: Accelerated Understanding **The Rundown:** Caltech professor Anima Anandkumar and engineer Benedikt Jenik just [launched](https://acceleratedunderstanding.com/longform?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) Accelerated Understanding, a new startup training AI to forecast how the physical world evolves rather than predict what word comes next. Instead of the traditional transformer, the model runs on “neural operators”, a physics-based design that follows events through 3D space and time*.* Anandkumar and Jenik were initially offered top positions and a 35% stake in Jeff Bezos’ Prometheus, but chose to build their own venture instead. In tests, AU’s model processed 5T data points in a single run, about 5M (!) times what Google and Anthropic's top models hit. Early business targets include chip materials, extreme-weather prediction, and robotics, with the company focused on enterprise over consumer models. **Why it matters:** The physical AI race is on, with several labs betting on world models — but this duo sees physics as the answer over text or video. Bezos’ Prometheus has since raised $12B chasing a similar goal, making the decision to pass on a 35% stake from one of the world’s richest men a serious conviction play. **The Rundown:** In this guide, you'll learn how to use Claude Voice to build better websites faster. We'll walk you through a technique that you can use every time you want to build a prototype. **Step-by-step:** Optional: If you have photos or design references, you can stash them in a folder. Now open Claude Desktop, start a regular chat, and click Voice Explain the website, then ask Claude to interview you until it has a design plan and brand direction At the end of the voice chat, have Claude write out the website building plan in full in the chat Now with Claude in text mode, drop in the photos and tell it to build the prototype based on the plan. If it skips images, say: “Use photos in this artifact” **Pro tip:** “Static site” is the keyword for portfolios, landing pages, and mockups. It keeps the complexity way down for you and the agent. **The Rundown:** Scribe Optimize automatically maps how work actually happens across your org — so before you build an AI strategy, you're working from reality, not assumptions. **With Scribe Optimize, you can:** Map real workflows — no surveys or consultants Surface inefficiencies leadership can't currently see Build your AI strategy on ground truth, not gut feel Image source: Perplexity **The Rundown:** Perplexity and Nvidia just [launched](https://www.perplexity.ai/hub/blog/introducing-portable-computer-for-local-first-ai?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) Portable Computer, a new on-device take on Perplexity’s Computer agent that keeps files private and is free to run locally on Nvidia’s DGX Spark consumer hardware. Users can select between Qwen 3.8 27B or Perplexity’s own PPLX 27B as the local model, with the option to call in 15+ cloud models when needed. On-device work doesn’t use any credits, with connectors able to integrate with popular apps and all cloud use requiring approval by the user. Portable installs in a single click on Nvidia’s $4,699 DGX Spark desktop supercomputer, with certain PCs also able to run the agent in the near future. The release comes just days after reports that Nvidia is in talks to invest billions of dollars in Perplexity in a new equity round at a valuation of over $30B. **Why it matters:** Perplexity already brought its agent to Macs with Personal Computer, but Portable moves the actual AI work from the cloud onto the hardware itself. It’s a privacy and cost pitch that should only strengthen as smaller open models improve, turning frontier AI into the occasional backup instead of the default. **QUICK HITS** ▸ [A reader is using AI to manage a $13.5M church renovation](https://app.therundown.ai/community/posts/cd4df2ad-d27c-451c-a62c-d3feb2787248?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) Today’s workflow comes from reader **Anonymous**: I’m 71 years old, retired from owning a large cattle-feeding operation, and fairly new to AI and software development. I’ve discovered that I really enjoy using AI to build practical tools in Replit. My latest project is an app for our church to help us oversee a $13.5M renovation of a 65,000-square-foot former movie theater into our new church facility. I’m not a programmer or a construction expert, so I’m using AI to help bridge both gaps. I built the app almost entirely by describing in plain English what I wanted it to do, testing what it built, and working back and forth with AI to improve it. The app uses AI to read meeting notes, emails, texts, and general project updates and identify decisions, action items, important dates, budget changes, and project events. Nothing is accepted automatically. I review, edit, approve, or reject what AI finds before it becomes part of the project record. The app also keeps our budgets, documents, meetings, and project history together. *See the full workflow here. How do you use AI? Tell us for a chance to be featured.* ⚙️ [Adapt](http://adapt.com/?utm_source=www.therundown.ai&utm_medium=referral&utm_campaign=) - Your company, your agent. One agent connected to every tool your team uses. Start building free\* 🔌 [Portable Computer](https://www.therundown.ai/tools/portable-computer?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) - Perplexity's agent that runs fully on Nvidia hardware 🔍 [Keenable](https://www.therundown.ai/tools/keenable?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) - Search API that lets agents query and learn from the live web 🏷️ [Stickers in ChatGPT](https://x.com/ChatGPT/status/2091996384954069032?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) - Build sticker packs from photos to send on iMessage *\*Sponsored Listing* **Anthropic** [rolled out](https://claude.com/blog/claudes-memory-works-everywhere-and-you-decide-whats-in-it?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) a single Claude memory shared between chat and Cowork, now also able to save topics in real time mid-conversation. **Apple** [introduced](https://tech.therundown.ai/p/apple-mac-mini-makes-ai-comeback?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) its new $899 Mac Mini, describing it as its “leading desktop for always-on agentic computing”, with new M6 chips to handle workloads up to 4x faster. **OpenAI head of data centers Chris Malone** reportedly [left](https://www.bloomberg.com/news/articles/2026-08-25/openai-data-center-executive-chris-malone-departs-the-ai-startup?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) the company, the latest in a flurry of executives to depart in recent months. **Google Cloud** [launched](https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-for-legal?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) industry-tuned Gemini Enterprise editions for financial services and legal teams in preview, with healthcare and life sciences coming next. **Anthropic** is reportedly [preparing](https://www.wsj.com/tech/ai/anthropic-expected-to-tell-investors-it-sees-over-30-trillion-in-potential-revenue-a611efea?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) to tell IPO investors it sees a $30T+ addressable market, topping the $28.5T figure SpaceX floated in May. Read our last AI newsletter: [SpaceX-Nvidia partnership heads to orbit](https://www.therundown.ai/articles/the-spacex-nvidia-partnership-heads-for-orbit?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) Read our last Tech newsletter: [Apple’s Mac Mini makes an AI comeback](https://tech.therundown.ai/p/apple-mac-mini-makes-ai-comeback?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) Read our last Robotics newsletter: [Humanoids beat Bolt’s 100M record](https://robotnews.therundown.ai/p/humanoids-beat-usain-bolt-100m-record?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) Today’s AI tool guide: [Build a website hands-free with Claude Voice](https://app.therundown.ai/guides/build-a-website-hands-free-with-claude-voice?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) RSVP to next workshop on Aug. 27: [Build your first AI Eval](https://app.therundown.ai/live/how-to-build-your-first-ai-eval?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-0923) See you soon, *Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown*