OpenAI Close to Landing $500 Billion Ohio Data Center New York Times / Bloomberg OpenAI is finalizing a lease on a 10-gigawatt facility in Piketon, Ohio — a former uranium enrichment plant being redeveloped by SoftBank — that could cost more than $500 billion including chips. Nvidia is in talks to provide a $250 billion financial backstop, partly because OpenAI lacks an investment-grade credit rating. The deal still needs Commerce Secretary Howard Lutnick’s sign-off, and the first phase is slated for completion in 2028.

🏗️ Infrastructure & Compute

OpenAI Close to Landing $500 Billion Data Center With Nvidia’s Backing — The New York Times

The proposed Ohio campus would be the largest data center project ever announced. Nvidia would stand behind financing instruments to reassure lenders, with OpenAI also reportedly discussing a chip purchase valued up to $350 billion. The scale of the bet underscores how central compute access has become to frontier AI development.

Safe Superintelligence Partnered with Nvidia — TechCrunch

Ilya Sutskever’s secretive SSI announced a long-term partnership with Nvidia that includes an undisclosed investment and access to Vera Rubin GPU clusters. Nvidia reportedly received a rare glimpse into SSI’s research before committing — SSI had previously relied primarily on Google chips.

Recursive Superintelligence Signs $410M Compute Deal with Amazon — TechCrunch

The self-improving AI startup, which routes most of its budget into compute rather than headcount, signed a $410M AWS deal to automate its own product development pipeline. A recurring theme: AI labs betting on compute-forward rather than people-forward scaling.

Data centers may face temporary power cuts to prevent blackouts on largest US grid — TechCrunch · Grid operators are considering interruptible load agreements with data centers as the breakneck pace of AI build-out strains US electricity infrastructure.

ASML Slides on Report of China Starting DUV Tool Production — Bloomberg · A Shanghai-based firm has begun mass-producing immersion deep-ultraviolet lithography tools, targeting ~5 machines this year and ~20 next — early-stage but a notable milestone in China’s effort to circumvent chip export bans.

🤖 Frontier Models

Releasing the Model Weights and Technical Report of Kimi K3 — Moonshot AI

Moonshot has open-sourced Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window and native visual understanding. Built on a new hybrid attention architecture (Kimi Delta Attention), it claims 2.5× the intelligence per unit of compute over its predecessor. Alongside the weights, Moonshot is releasing high-performance attention kernels, a MoE communication library, and infrastructure for agent environments — making this the most comprehensive open-source release in the 3T-parameter class to date.

Microsoft Introduced MAI-Cyber-1-Flash Cybersecurity Model — Microsoft

Microsoft launched a specialized model for finding hard-to-spot vulnerabilities in large codebases. It powers MDASH, a new platform for identifying and remediating software security flaws. Coming alongside the launch of the Open Secure AI Alliance (see below), it signals Microsoft’s push to make AI a core pillar of enterprise security posture.

How we built and benchmarked VR-1, our frontier cyber reasoning model — Cogent · VR-1 autonomously investigates environments, crosses system boundaries, and executes attack chains; it achieved 2× pass@3 over the strongest frontier baseline on the new IntrusionBench enterprise intrusion benchmark.

Bringing MCP 2026-07-28 to Claude — Anthropic · Anthropic rolled out the latest MCP spec revision to Claude, keeping the platform current with the evolving Model Context Protocol standard.

🔒 Security & Safety

PSA: Your Claude Shared Chats and Artifacts May Have Ended Up on Google — TechCrunch

Reddit users discovered that Google had indexed hundreds of publicly shared Claude conversations containing sensitive data — resumes, financial spreadsheets, patient names, and API keys — because shared-chat pages lacked a noindex tag. Anthropic quickly updated robots.txt to wipe the results, but the incident highlights a gap between user expectations (“shared link = private”) and reality (“shared link = public URL”). Users should audit and unpublish sensitive shared chats.

Hugging Face is Being Used to Easily Undress Women and Children — The Verge

A report by European nonprofit AI Forensics found that seven of the top nine image-editing models hosted on Hugging Face readily comply with nonconsensual deepfake requests, with Hugging Face doing little to prevent it. The finding renews pressure on open-model repositories to enforce meaningful content policies for hosted weights.

Industry Leaders Unite in Open Secure AI Alliance — NVIDIA

NVIDIA, Microsoft, and other industry leaders launched the Open Secure AI Alliance to build open-source defensive AI tools and promote transparency in cybersecurity. The initiative also calls on policymakers to treat open models as strategic assets rather than liabilities — a direct counterpoint to calls for open-weight restrictions.

Anthropic Rejected Blanket Bans on Open-Weight Models — Anthropic · Dario Amodei clarified Anthropic’s position: the company does not advocate banning open-weight models, but fears Chinese AI capabilities and supports chip controls, anti-distillation measures, and mandatory safety testing for sufficiently capable models.

🛠️ Developer Tools & Platforms

Perplexity’s Personal Computer Turns Windows PCs into AI Agents — The Verge

Perplexity expanded its agentic “Personal Computer” tool to Windows, giving the world’s most popular desktop OS a locally-running general-purpose digital worker that can access local files and apps. The Mac version launched in April; the Windows release significantly broadens the potential install base.

Gemini Distillation Service — Google Cloud · Google launched a production distillation pipeline that trains a smaller student model from a larger Gemini teacher, targeting high-volume, latency-sensitive workloads where frontier-grade reasoning is needed at inference efficiency.

PorTAL — Ramp Labs · Open-sourced framework for cross-model LoRA adaptation: train a task once, adapt it to any supported frozen base model, export as a standard Hugging Face PEFT adapter.

Molt Agentic Reinforcement Learning Framework — NVIDIA NeMo · PyTorch-native RL framework that treats the agent as the program, supporting custom Python rewards, tool use, multimodal environments, and scaling to trillion-parameter MoE models via Ray + vLLM + FSDP2.

CLI: Spec Drafting and Plan Auto-Execution — Kiro · Kiro’s CLI adds guided /spec new and automatic plan execution after approval — no more manual mode-switching.

🏢 Enterprise & Market

OpenAI’s Report on How AI is Expanding What People Do at Work — OpenAI · Analysis of 800,000 US ChatGPT sessions found significant “task crossover” — workers using AI to perform tasks traditionally associated with other occupations, with implications for how skills and job functions evolve.

Satya Nadella Says Companies That Trust One AI for Everything May Not Survive — TechCrunch · Nadella argued enterprises need AI gateways to insulate their data from any single model vendor — pointing toward Microsoft’s own Azure AI infrastructure play.

Fish Audio Raises $52M Seed to Build AI Voice Models for Creators and Enterprises — TechCrunch · The startup has 8M users and $21M ARR from its open-source and hosted voice models.

Cursor Makes Its Biggest India Push Ahead of SpaceX Acquisition — TechCrunch · Cursor introduced localized pricing for India, now its third-largest market, as the SpaceX acquisition approaches.

Generated by claude-sonnet-4-6 · 2026-07-28T10:00:00Z