Samwise Tech/AI/Robotics Newsletter
Thursday, May 7, 2026
Anthropic Secures Full Colossus 1 Capacity in xAI Neocloud Deal
Anthropic has secured all computing capacity at xAI's Colossus 1 data center — more than 300 megawatts backed by roughly 220,000 NVIDIA H100, H200, and GB200 GPUs. xAI had already migrated model training to a newer Colossus 2 facility, making the capacity available. The arrangement positions xAI as a “neocloud” operator, selling GPU access to third-party developers. For Anthropic, the deal immediately doubles Claude Code's five-hour rate limits for Pro, Max, Team, and Enterprise tiers and removes peak-hours restrictions. The two companies also announced interest in developing gigawatt-scale orbital AI compute infrastructure.
Sources: TechCrunch
Samsung Crosses $1 Trillion Market Cap on AI Memory Chip Surge
Samsung Electronics crossed the $1 trillion market capitalization threshold on Wednesday, becoming the second Asian company after TSMC to join the trillion-dollar club. Shares surged more than 10 percent after the company reported profits eight times higher year-over-year, driven by surging demand for high-bandwidth memory chips used in AI training and inference workloads. Samsung announced in February that it had become the first manufacturer to begin mass production of HBM4 chips and start deliveries to undisclosed AI clients. The stock has gained more than 120 percent in 2026, outpacing global chip peers as AI-driven memory demand tightens supply.
Sources: TechCrunch
Genesis AI Unveils GENE-26.5 Foundation Model and Full-Stack Robotic Hand
French robotics startup Genesis AI, backed by Eclipse and Khosla Ventures, has unveiled its first foundation model, GENE-26.5, alongside a proprietary dexterous robotic hand. The model is hardware-agnostic, capable of controlling robots from multiple manufacturers via a physics-based simulation platform that generates synthetic training data up to 430,000 times faster than real-world collection. Demo footage showed GENE-26.5-powered hands cracking eggs, slicing tomatoes, playing piano, and solving a Rubik's Cube. CEO Guillaume Collignan said the company shifted to a full-stack approach after concluding software-only control was insufficient for complex manipulation tasks. The startup raised $105 million in a July 2025 seed round.
Sources: TechCrunch
MCP Design Flaw Leaves 200,000 AI Agent Servers Vulnerable to Command Injection
Security researchers at OX Security have identified a design flaw in Anthropic's Model Context Protocol that leaves an estimated 200,000 MCP server instances vulnerable to command injection attacks. The vulnerability stems from MCP's default STDIO transport, which executes any incoming OS command without sanitization or execution boundaries, enabling four distinct exploit families including zero-click attacks on AI coding IDEs. Anthropic updated its SECURITY.md file nine days after initial contact in January 2026 to add a caution about STDIO adapters but made no architectural changes. The flaw propagated into every official MCP SDK — Python, TypeScript, Java, and Rust.
Sources: VentureBeat
Google Updates AI Search to Surface Reddit Threads and Personalized News Sources
Google has begun rolling out two updates to its AI-powered search interface: excerpts drawn from Reddit threads and expert-sourced web forums are now surfaced alongside standard link results, and links from a user's subscribed news outlets are highlighted to improve source transparency. The Reddit integration is notable given Google's 2024 licensing agreement with the platform, which gave the company access to its real-time data fire hose for AI training and search. Google said the updates are rolling out to United States users first across web and mobile search, with broader availability to follow.
Sources: TechCrunch
OpenAI Launches GPT-5.5 Instant as New ChatGPT Default with Improved Benchmarks
OpenAI released GPT-5.5 Instant on Monday as the new default model powering ChatGPT, replacing GPT-5.3 Instant. The model is a faster, lower-latency option compared to the full GPT-5.5 released earlier this year. OpenAI said GPT-5.5 Instant scored 81.2 on the AIME 2025 mathematics benchmark and 76.0 on the MMMU-Pro multimodal test — improvements of roughly 25 percent and 10 percent respectively over its predecessor. The model also introduces a new memory transparency feature that shows users which prior context shaped a given response. The update applies to all free, Plus, and Pro ChatGPT users.
Sources: TechCrunch
AI Evaluation Platform Braintrust Confirms AWS Breach, Urges All Customers to Rotate Keys
AI evaluation platform Braintrust has confirmed a data breach after unauthorized actors gained access to one of its Amazon Web Services cloud environments. Braintrust, which bills itself as an operating system for engineers building AI applications, notified all customers and instructed them to immediately rotate API keys and other sensitive credentials. The company said it detected the intrusion and notified affected users within 24 hours of confirmation. Braintrust's platform is used by engineering teams to run automated evaluations, benchmark model performance, and track regression across AI product releases. The company has not publicly disclosed how many customers were affected.
Sources: TechCrunch
Tech Pulse
Top Frontier Models: GPT-5.5 (72.4%) | Claude Opus 4 (70.1%) | Gemini 2.5 Ultra (65.8%)
Top Open Source Models: Llama 4 Maverick (60.3%) | Qwen3-72B (57.6%) | DeepSeek-V4 (55.2%)
Top Small Models (15–50B): Qwen3-32B (52.8%) | Phi-4 (48.1%) | Mistral Medium 3 (44.6%)
Top Edge Models (0–15B): Phi-4-mini (40.3%) | Gemma 3-12B (37.2%) | Qwen3-14B (35.8%)
AI Leaders: NVIDIA $3.2T | Microsoft $3.0T | Alphabet $2.1T
Robotics Leaders: Intuitive Surgical $186B | ABB $73B | Fanuc $31B
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.