Today’s lead · x.com
NVIDIA introduces revenue-sharing AI factories for cloud partners
NVIDIA is partnering with AI clouds on large-scale, multi-tenant AI factories using a revenue-sharing and credit-support model. The first deployments include Sharon AI with up to 40,000 Grace Blackwell GB300 GPUs and Firmus with a Batam, Indonesia campus expected to scale to 360 megawatts and up to 170,000 NVIDIA GPUs.
Top signals
3 more
Sysdig says an AI agent automated a Langflow-to-database ransomware attack
Sysdig says the JADEPUFFER operator used an LLM-driven agent to exploit the already-patched Langflow CVE-2025-3248, harvest cloud and model API credentials, pivot to MySQL and Nacos, and encrypt or delete production data. The attacker also hit weak defaults such as MinIO's minioadmin:minioadmin login and a longstanding Nacos signing key.
Anthropic is reportedly discussing a custom AI chip with Samsung
TechCrunch reports that Anthropic is in contact with Samsung about a potential custom AI chip, following earlier reports that the company was exploring its own hardware. Anthropic told TechCrunch that a diversified hardware stack across Google, Amazon, and NVIDIA remains central to its compute strategy.
Meta reportedly tells staff AI agents are progressing slower than expected
Reuters, via TechCrunch, reports that Mark Zuckerberg told staff Meta's AI agent work had not accelerated as executives expected and that the upside from its AI-focused reorganization had not yet come to fruition. Meta has reportedly laid off about 8,000 employees, reassigned 7,000 to AI groups, and is expected to spend as much as $145 billion on AI infrastructure this year.
Tools & repos
6 selected
Safari MCP server
Safari Technology Preview 247 adds a Model Context Protocol server that lets MCP-compatible agents inspect a live Safari window. Agents can use DOM data, network requests, screenshots, console output, performance signals, accessibility checks, and user-state verification without bouncing through separate browser tooling.
DeepSeek DSpark in vLLM
vLLM nightly now runs DeepSeek's DSpark speculative decoding natively, using a semi-autoregressive drafter that proposes multiple tokens and verifies them in one pass. The implementation reuses SparseMLA backends, captures the draft and sampling loop in one CUDA graph, and works with prefix caching and FP8 KV cache.
Claude Platform API rate limits
Anthropic raised Claude Platform API rate limits for all users and simplified tiers so they are no longer based on API spend. The latest Sonnet and Haiku models now provide 5x higher rate limits at the highest tier.
Claude Code Artifacts on Pro and Max
Claude Code Artifacts are now available on Pro and Max plans. Claude can write code, publish a private self-contained page to claude.ai, and keep updating it live while it continues working in the session.
Context.dev
Context.dev offers one API to scrape, enrich, and extract data from the internet. For AI teams, the useful angle is turning messy web pages into structured context for agents, RAG systems, and internal automation without maintaining a custom scraping stack.
Claude Code /dataviz skill
Claude Code v2.1.198 adds a built-in /dataviz skill that loads chart and dashboard design guidance into context. It also includes a runnable color-palette validator so Claude can check contrast and accessibility programmatically instead of guessing.
Blogs worth your time
6 reads
Hardware-Rooted AI Security That Won't Slow You Down
NVIDIA details how Blackwell confidential computing protects model weights, data, and code during inference with hardware roots of trust, NVLink encryption, and remote attestation. The useful part is the benchmark: Qwen 3.5 397B on HGX B300 sees mostly single-digit overhead while SGLang and FlashInfer optimizations keep secure inference close to native performance.
Your coding agent bill doubled. Here's how to fix it.
LangChain explains why multi-agent coding setups make spend hard to attribute across Claude Code, Cursor, Copilot, Codex, OpenCode, and Pi. It lays out a practical control loop: normalize traces, compare cost per session, identify wasteful tool use, and add gateway-level caps and routing.
Multi-Agent Teams Hold Experts Back
Apple researchers test self-organizing LLM teams and find they often underperform their best individual expert agent, with losses up to 41.1% on ML benchmarks. The practical warning is clear: adding agents can create consensus averaging unless the system explicitly routes authority to the right expert.
VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization
Apple presents a video tokenizer that emits variable-length coarse-to-fine token sequences instead of fixed 3D grids. For video model builders, the signal is efficiency: comparable generation quality with a 1.1B model versus a 5.2B baseline and 10-second, 81-frame video represented with 672 tokens.
Conformal Thinking: Risk Control for Reasoning on a Compute Budget
Apple and Johns Hopkins reframe adaptive reasoning as a risk-control problem: stop early when the model is confident, stop unsolvable cases before wasting tokens, and tune thresholds against a validation set. It is a useful pattern for teams trying to control reasoning cost without blindly capping token budgets.
New analytics and cost controls are available for Claude Enterprise
Anthropic adds richer Claude Enterprise admin analytics, model-level entitlements, spend alerts, and an Analytics API. The important operational detail is that admins can now break down cost by group, user, product, model, skills, artifacts, Claude Code sessions, and connectors before usage surprises become budget problems.
Funding & acquisitions
2 moves
Microsoft Frontier Company
Microsoft launched Microsoft Frontier Company, an operating business focused on enterprise AI deployments using Microsoft's existing AI tools. The effort is backed by a $2.5 billion Microsoft investment and 6,000 industry and engineering experts, with early partnerships cited across London Stock Exchange Group, Unilever, Land O'Lakes, and Accenture.
Neo
Bhavin Turakhia is self-funding Neo, a Bengaluru enterprise work platform combining project management, documents, file storage, and AI. The company says it is model-agnostic, has been in internal use since April, and plans to roll out to mid-sized businesses in technology, consulting, and professional services.
Bengaluru radar
5 eventsAI Coding Summit, London Edition
A near-term summit focused on AI-powered software development, with talks and hands-on workshops around coding workflows and developer tooling.
Agentic AI Workshop at BITS Goa
A three-day hands-on workshop covering agent foundations, memory, tools, workflows, API and no-code builds, personal AI assistants, startup use cases, and a mini hackathon/demo showcase.
GenAI Summit 2026
A San Francisco summit with speakers listed from OpenAI Codex, Anthropic, Google DeepMind, AWS, Stripe, Harvey, Webflow, and major AI investors, with tracks on vertical agents, orchestration, autonomous workflows, and AI-native enterprises.
Think Like an AI Hacker: Intro to Offensive AI Security
A one-day Singapore workshop on AI-powered ethical hacking, offensive security tactics, and defensive implications, with instructor-led demos and interactive activities around AI-assisted reconnaissance, phishing, exploitation, and evasion.
DataHack Summit 2026
Analytics Vidhya's Bengaluru AI gathering returns with a theme around Human × AI and the agentic operating layer, listing 75+ sessions and 10+ hands-on workshops for real-world AI applications.