AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Anthropic Lines Up $36B Off-Balance-Sheet Debt for Google TPUs, Backed by Broadcom Guarantee
Apollo and Blackstone arranged a roughly $36 billion private debt package to fund Anthropic's purchase and lease of Google TPUs, with Broadcom providing a residual-value guarantee that lowers borrowing costs.
Key topics and reactions
Anthropic Lines Up $36B Off-Balance-Sheet Debt for Google TPUs, Backed by Broadcom Guarantee
Apollo Global Management and Blackstone have assembled a private debt package of around $35–36 billion to finance Anthropic's purchase and lease of Google Tensor Processing Units, with Broadcom backstopping the bulk of the senior tranches. A special-purpose vehicle acquires the TPUs and leases them to Anthropic, keeping the debt off Anthropic's own balance sheet.
The deal hinges on Broadcom's residual-value guarantee, which delivers near-investment-grade credit quality and drives down funding costs. Tranches are reported at roughly $6 billion (A1), $25 billion (A2) and $4.5 billion (B), with senior yields near 5.75% and junior yields around 8.5%. Proceeds fund TPUs across data centers in New York, Texas, Louisiana and Indiana.
The structure echoes Meta's 'Beignet bonds,' treating committed AI chip demand as collateral. Anthropic's revenue run rate has surged from $9 billion at the end of 2025 to over $30 billion, with more than 1,000 enterprise customers spending over $1 million per year. The company committed in October 2025 to up to one million TPUs and over 1 GW of capacity in 2026.
Anthropic Says Claude Now Writes Over 80% of Its Internal Code
Anthropic has disclosed that Claude now produces over 80% of its own internal codebase, citing the development as evidence of accelerating AI 'self-improvement' while warning about the associated risk management challenges.
The disclosure comes alongside reports that the company is preparing for an IPO and is rapidly scaling its compute footprint, including the large off-balance-sheet financing for Google TPUs.
OpenAI Overhauls ChatGPT Memory With New 'Dreaming System'
OpenAI has revamped ChatGPT's memory capability using a new architecture it calls the 'Dreaming system,' introducing a summary page that lets users review what the assistant retains about them.
The update is part of a broader push to make ChatGPT's personalization more transparent and controllable for users.
Category highlights
Chrome Previews 'Modern Web Guidance' for AI Coding Agents
Google's Chrome team released an early preview of 'Modern Web Guidance,' which injects modern web development expertise into AI coding agents like Claude, Cursor and Gemini CLI via SKILL.md-format skills. Tied to the Web Platform Baseline, the evergreen skills—validated by Chrome, Microsoft Edge and community contributors—aim to stop agents from defaulting to outdated patterns and deprecated APIs, covering Core Web Vitals, Passkeys, WebMCP integration and cross-browser fallbacks. It is open for contributions on GitHub.
NVIDIA Unveils 550B Open Model Nemotron 3 Ultra
NVIDIA announced Nemotron 3 Ultra, a 550-billion-parameter open frontier model optimized for AI agents. It joins a busy week of open-weight releases including Google's Gemma 4 12B and MiniMax's M3, reinforcing a trend toward selecting and mixing multiple frontier models.
Thinking Machines Previews 'Interaction Models'
Mira Murati's Thinking Machines Lab previewed 'interaction models' that process voice, text and video in 200ms increments, signaling a push toward real-time multimodal interaction.
Grok Build and Agent Mode See Rapid Updates
Grok Build shipped versions v0.2.28 through v0.2.31 with marketplace display, session restore, prompt ordering and faster large-repo search. Users with little coding experience report building usable agent systems with Grok alone, while Agent Mode enables full agent capabilities via a single config.toml line and supports low-cost setups with DeepSeek or Ollama. Complaints persist about slow resumption of large sessions and marketplace discoverability.
Grok Imagine Agent Mode Speeds Up Video Creation
Creators are using Grok Imagine's Agent Mode to turn ideas into short videos quickly, directing character design, lighting, camera angles, action and pacing before finishing edits in CapCut. While praised for enabling polished output from quick concepts, users note that truly commercial results still require manual work and multiple revision passes.
Key trends
NVIDIA Demos Sim-to-Real Robot Fleets at GTC Taipei
At GTC Taipei, NVIDIA showcased Isaac-powered Sim-to-Real robot fleets, emphasizing scaling 'physical AI.' xAI also announced a new model focused on understanding and manipulating the physical world, underscoring a broader push beyond purely digital tasks.