Safe Superintelligence inks $5 billion Nvidia deal to scale AI research
Plus: Anthropic clarifies its position on open-weight models and opposes bans. Microsoft releases a specialized cybersecurity AI. Delhi court rules AI training is fair dealing.
TL;DR
- Safe Superintelligence lands $5B from Nvidia to access Vera Rubin GPUs, the largest independent AI funding deal this year.
- Anthropic rejects calls for open-weight AI bans, backing chip controls and safety testing instead of restrictions.
- Delhi High Court rules AI model training qualifies as fair dealing, giving OpenAI a landmark copyright victory.
- Microsoft launches specialized cybersecurity AI that scores 96% on vulnerability detection while cutting costs in half.
Models and research
Anthropic stakes out ground against AI bans while courts rewrite training rules
As pressure mounts for restrictions on open-source models, Anthropic and courts in India clarify what AI companies can legally do.
Dario Amodei broke Anthropic's silence on open-weight AI by publishing a formal position stating the company has never called for blanket bans. Instead, Amodei backs chip export controls, cracking down on distillation of frontier models, and mandatory safety testing for all sufficiently capable systems. His core concern is authoritarian governments building more powerful AI than the US while Amodei sees open-weight models without dangerous capabilities as a public good (via Anthropic).
Delhi High Court classifies AI training as fair dealing. A landmark ruling rejected a major Indian news agency's copyright claim against OpenAI, holding that training ChatGPT on ANI's published reports qualifies as protected research under India's fair-dealing doctrine. The decision marks the country's first substantive judicial finding on whether AI companies can use copyrighted material to train models without a license—a precedent that contrasts sharply with ongoing suits in the US and Canada (via The New Publishing Standard).
METR introduces expenditure horizon, a metric for when AI agents beat humans. Researchers at METR published a framework measuring the point at which paying for AI agent optimizations costs less than hiring humans to do the same work. On tested problems, frontier models show crossover points around $2–3K, suggesting AI agents are already competitive with human researchers on specific high-skilled tasks (via The Decoder).
OpenAI research shows employees using ChatGPT to do other people's work. Analysis of 800,000+ work messages revealed 43.5 percent of job-specific queries involved tasks outside the asker's field, suggesting ChatGPT is becoming an internal tool for cross-functional collaboration within OpenAI itself (via The Decoder).
Industry and business
Nvidia backs Sutskever's AI startup with $5B, Cursor heads to India
Two major bets signal confidence in AI infrastructure and the economics of coding assistants.
Ilya Sutskever's Safe Superintelligence announced a $5 billion investment from Nvidia to scale its research after operating in stealth for two years. The deal grants SSI access to Nvidia's Vera Rubin GPU platform, increasing the startup's compute by an order of magnitude and positioning Sutskever as an independent force in frontier AI research after his departure from OpenAI (via TechCrunch).
Cursor launches India tier at $7 ahead of $60B SpaceX deal. The coding assistant expanded to India with a localized $7-per-month tier—called Cursor Start, far below its standard $20 Pro plan—and hired staff across four Indian cities. India is now Cursor's third-largest market. SpaceX's acquisition of parent Anysphere, valued at $60 billion, is expected to close in Q3 2026 (via TechCrunch).
Cognizant expands Claude partnership to embed AI in enterprise platforms. The services giant reached Global Premier Partner status in Anthropic's Claude Partner Network, scaling from internal testing to deploying Claude across its own business, engineering platforms, and client-facing systems in manufacturing, life sciences, and insurance. Cognizant has the highest number of Claude certifications globally (via Anthropic).
Products and tools
Microsoft deploys specialized cybersecurity AI while Threads and Claude expand reach
New tools embed AI deeper into enterprise workflows and consumer platforms, raising stakes around security and privacy.
Microsoft launched its first specialized cybersecurity model, MAI-Cyber-1-Flash, designed to find vulnerabilities in complex codebases. The compact model scores 96% on CyberGym—a benchmark measuring how well AI systems reason over large repos to find real bugs—while cutting costs roughly in half compared to Microsoft's production configuration. Microsoft's Project Perception platform enters public preview August 3, deploying teams of agents to automate security workflows (via TechCrunch).
Claude shared chats reappeared in Google search, briefly exposing conversations. Hundreds of Claude AI shared chats became searchable via Google over the weekend before Anthropic added noindex tags. Reddit users discovered site:claude.ai/share queries surfaced conversations containing legal advice, engineering work, health records, and personal documents. The issue stemmed from share links lacking proper meta tags after initial indexing (via TechCrunch).
Threads rolls out Meta AI in direct messages globally. Meta expanded its AI assistant into Threads' private messaging, letting users DM Meta AI with conversations, images, and links. The move follows public-feed testing and positions Meta's chatbot as an in-app alternative to third-party assistants like ChatGPT and Gemini (via TechCrunch).