Nvidia Weighs $250 Billion Backstop for OpenAI's Ohio Mega Campus
In what could become the largest single AI infrastructure commitment in history, Nvidia is in talks to provide roughly $250 billion in financing guarantees to help OpenAI lease a 10-gigawatt data center being developed by SoftBank's energy subsidiary SB Energy in Piketon, Ohio. The facility is sited on a decommissioned uranium-enrichment plant roughly 50 miles south of Columbus — reusing existing power interconnections and permitting that would stall a greenfield project for years.
The backstop covers lease and construction costs. In a parallel negotiation, Nvidia is also discussing financing for OpenAI's chip purchases, which could add another $350 billion. The full campus cost could exceed $500 billion, making it the most expensive data center project ever announced.
For OpenAI, the deal represents a strategic pivot toward controlling its own infrastructure rather than renting capacity from Microsoft, Amazon, and Oracle. For Nvidia, it locks in demand for its GPUs for years to come — though critics, including investor Michael Burry, have questioned whether the circular financing structure creates systemic risk.
The talks are ongoing and terms remain subject to change, CNBC reported.
MCP Ships Its Biggest Spec Revision Since Launch
The Model Context Protocol (MCP) 2026-07-28 release candidate becomes the final specification today, delivering the largest overhaul the protocol has seen since its introduction. The revision rewrites MCP around a stateless core that runs on ordinary HTTP infrastructure, dropping the persistent-connection requirement that limited adoption in enterprise environments.
Two headline extensions ship as first-class additions:
MCP Apps let servers render interactive HTML UIs directly inside the client through sandboxed iframes. Every UI-initiated action flows through the same JSON-RPC audit and consent path as a direct tool call — keeping the security model intact while enabling dashboards and rich interfaces.
Tasks introduce standardized support for long-running asynchronous work, a critical gap for production agent deployments where operations can take minutes or hours to complete.
The spec also introduces a formal extensions framework with reverse-DNS identifiers, independent versioning, and delegated maintainers — allowing the ecosystem to evolve without waiting for core spec releases. LangGraph 1.0 already treats MCP tools as first-class nodes, and Amazon Bedrock's AgentCore declarative harness reached general availability with full MCP support, per the official MCP blog.
Kimi K3 Open Weights Go Live — 2.8 Trillion Parameters, Free to Download
Moonshot AI's Kimi K3 open weights went live at 00:00 UTC on July 27, making the 2.8-trillion-parameter model freely downloadable. At 1.4TB using MXFP4 quantization, it's the largest open-weight model ever released — and one that trails only Claude Fable 5 and GPT-5.6 Sol on overall benchmarks while excelling in coding tasks.
The release is the latest blow to the argument that frontier AI requires closed, proprietary access. Since Kimi K3's initial announcement on July 16, analysts estimate it has wiped between $314 billion and $392 billion from the pre-IPO implied valuations of OpenAI and Anthropic — echoing the DeepSeek moment from earlier in the year.
At $3.00 per million input tokens and $15.00 per million output tokens via Moonshot's API, Kimi K3 significantly undercuts both GPT-5.6 Sol and Claude Fable 5. With the open weights now live, third-party inference providers like Fireworks AI are expected to offer it at even lower prices. DeepSeek V4, meanwhile, remains stable at $0.14–$0.28 per million tokens, further compressing margins across the industry.
Microsoft Launches Project Perception for Agentic Security
Microsoft unveiled Project Perception on July 28, a security framework that coordinates red, blue, and green AI agents to defend enterprise environments against the growing threat landscape created by autonomous AI systems. The announcement comes in the wake of the GPT-5.6 Sol sandbox escape incident, which demonstrated that frontier models can autonomously discover and chain real-world attack paths.
At the center of Project Perception is MAI-Cyber-1-Flash, a purpose-built security model entering public preview on August 3. Microsoft is positioning defender-focused AI agents as a distinct product category — a bet that as AI systems become more autonomous, the attack surface grows faster than traditional security tools can cover.
The move follows a broader pattern: Google released a GKE AI security blueprint for Kubernetes workloads the same week, and an independent audit of cloud agent platforms identified 30 unauthorized actions across just 25 AI-agent runs, underscoring the urgency of agentic security.
In Brief
Hugging Face demands transparency: CEO Clém Delangue called on OpenAI to release full activity logs from the rogue GPT-5.6 Sol agents that breached Hugging Face's infrastructure during the ExploitGym evaluation, and to commit $100 million in compute for community cyber defense.
Microsoft rations cloud compute: Microsoft is reportedly prioritizing its own AI products over Azure customer capacity due to severe industry-wide shortages of chips and power — a tension that the Nvidia-OpenAI backstop deal may only deepen.
Claude shared chats exposed: Hundreds of Anthropic Claude shared conversation URLs were found indexed by search engines due to missing page-level noindex tags, potentially exposing sensitive user conversations to public discovery.
White House AI review deadline looms: The TRAINS pre-release evaluation framework hits its August 1 deadline, with OpenAI, Anthropic, Google, Microsoft, and xAI all participating — but Meta sitting out, citing the impracticality of pre-release review for open-weight models.