Happy Thursday! Last week OpenAI's Jalapeño chip outpaced rivals in speed, while Perplexity partnered with Nvidia to bring AI agents fully local. Starcloud raised $250M for orbital AI data centers and Thomson Reuters launched a proprietary legal AI model.

NEW AI CHIP LAUNCH

OpenAI's Jalapeño Chip Beats Rivals in Speed

OpenAI has unveiled its new AI chip called Jalapeño, claiming it outperforms competing AI systems in both efficiency and response speed. The announcement was made in a blog post published Tuesday. OpenAI hardware vice president Richard Ho said Jalapeño delivers the best of both worlds, achieving lower latency and higher throughput simultaneously. This is notable because AI systems typically must trade off one for the other.

NEW GENERATIVE AI LAUNCH

Perplexity Brings AI Agents Fully Local With Nvidia

Perplexity is launching Portable Computer, an agentic AI platform that runs entirely on local hardware, starting with Nvidia's DGX Spark desktop supercomputer and Linux machines with Nvidia RTX GPUs. The move marks one of the boldest efforts to shift serious AI workloads off the cloud. All processing, files, and tasks stay on the user's device by default, with no token costs for local work. The system only contacts cloud-based frontier models with explicit user permission, giving users full control over privacy and spending.

AI STARTUP FUNDING

Starcloud Raises $250M for Orbital AI Data Centers

AI hardware startup Starcloud has raised $250 million at a $2.3 billion valuation, led by Manhattan West with backing from Nvidia and Cisco Investments. The funding extends a Series A round announced in March, following the company's landmark first-ever orbital AI training run aboard a small satellite. Starcloud plans to launch 88,000 satellites with 20 gigawatts of combined computing capacity, citing advantages over ground-based data centers including simpler cooling, continuous solar power, and weather immunity. Its next satellite, Starcloud-2, launches in 2027, with more advanced generations planned, including the cylindrical Starcloud-4 attached to a 2.5-square-mile solar array.

NEW AI LEGAL LAUNCH

Thomson Reuters Debuts Proprietary Legal AI Model

Thomson Reuters has launched Thomson, its first proprietary large language model tailored for legal work. Built on an open-weight model enhanced with the company's vast legal content and professional expertise, it will debut in the CoCounsel Legal AI assistant's Tabular Analysis feature. The project cost roughly $40 million over two years, though the final training run came in at just $450,000. Thomson outperformed general models in internal tests when connected to Thomson Reuters content, including its Westlaw platform spanning 40,000 databases. Only 10% of the company's content has been used so far. Plans include an open-weight academic release, an API portal for developers, and potential customization options for large law firms and corporations.

NEW GENERATIVE AI LAUNCH

Glean launches Tau desktop AI, beats Claude on cost

Glean Technologies unveiled Glean Tau, a desktop workspace connecting enterprise AI to local files, apps and code. The company claims a 5.2-times token-cost advantage over Anthropic's Claude Cowork and user preference of 3.6 times, with Tau capable of planning and executing multistep tasks autonomously. The launch headlined Glean's GO conference in San Francisco, alongside six other products covering usage analytics, proactive task management, interactive dashboards, team chat, AI governance and security threat detection. Several features are in beta or not yet released. Glean is valued at $7.2 billion following a $150 million Series F in June 2025.

NEW GENERATIVE AI LAUNCH

DeepMind Spinout's Tiny AI Beats OpenAI, Anthropic

London startup Inherent, founded by Google DeepMind alumni, claims its AI agent Faraday has outperformed frontier models from Anthropic and OpenAI at independently replicating published scientific research. Notably, Faraday runs on a 27-billion-parameter model, far smaller than the systems it beat. The company, which emerged from stealth with a $50 million seed round, uses reinforcement learning to teach AI research instincts rather than rigid rules. Its longer-term goal is building AI capable of discovering genuinely new scientific knowledge, not just verifying existing results.

AI INFERENCE OPTIMIZATION

Harness beats model in Nvidia's AI research

Nvidia research shows the software harness around an AI model matters more than the model itself for long-horizon tasks. Using a custom harness with a supervisor component, researchers pushed Claude Opus 5 from a 30% to a perfect 100% score on the ARC-AGI-3 benchmark. The findings echo similar research from Databricks, which found harness choice can double AI costs regardless of model selection. Nvidia argues open harnesses give users far more control over accuracy and performance than most realize, and its researchers built a custom system called Agentic Variation Operators to prove the point.

AI MODEL EFFICIENCY

Nvidia cuts AI handoff costs with simple linear math

Nvidia researchers have developed a cross-model KV cache transfer technique that eliminates costly recomputation when agentic AI systems hand tasks between models. Instead of forcing the receiving model to reprocess entire conversations from scratch, the method maps cached data directly from one model to another using simple linear math. The technique runs 2.7 to 25 times faster than traditional recomputation while retaining up to 98% accuracy. Unlike previous approaches, it requires no expensive deep learning model, making it a practical solution for enterprises running long-horizon, multi-LLM workflows.

NEW GENERATIVE AI LAUNCH

Claude Now Reads Slack Chats, Joins Unprompted

Anthropic has updated Claude Tag, its Slack-based AI agent, to read entire conversation threads rather than individual messages. The change makes Claude roughly 30% better at deciding when to join discussions without being directly asked. The update reflects Anthropic's broader push toward what it calls multiplayer AI, moving beyond single-user chatbots toward agents that operate across teams and proactively insert themselves into workplace conversations. The company believes the biggest enterprise AI bottleneck is not model intelligence but the fact that most people still use AI alone.

NEW GENERATIVE AI LAUNCH

Claude Gets Smarter Memory With Privacy Controls

Anthropic has updated Claude's memory system to give users full visibility and control over what the AI remembers, organized topic by topic. Users can now edit or delete any stored information, while sensitive subjects like health, race, religion, and political beliefs are excluded by default unless manually enabled. The updated memory works across Claude's chat interface and Cowork desktop agent, persisting context over months without users needing to repeat themselves. Certain data, including Social Security numbers, government IDs, and immigration status, is never stored regardless of user settings.

Upcoming Data & AI Conferences

NVIDIA GTC Berlin 2026

Tue, Oct 20 - Thu, Oct 22, 2026 | Berlin, Germany

Curated with AI by EMC2 AI

Brought to you by Estevan McCalley

Share this issue

Follow Estevan

Like this? Subscribe for the weekly AI roundup.