The $10 Billion
Handshake

Anthropic is in early talks to lease compute from Meta in a deal that could reach $10 billion over two years. CNBC reported the story July 17, citing a person familiar with the matter. The New York Times confirmed the figure. Neither company commented.

The mechanism: Anthropic proposed the arrangement in June. Meta's data centers, built for LLaMA training at a scale that Mark Zuckerberg has since acknowledged may have been overbuilt, have headroom. Anthropic needs capacity. The reported structure runs in monthly installments with an early-exit option for either party. The practical effect: Anthropic would add Meta's infrastructure alongside the SpaceX Colossus deal it signed weeks prior, reducing dependence on any single cloud partner before what looks like a coming IPO.

The blast radius here is not the compute. It is the optics.

Anthropic has built its identity around Constitutional AI and the argument that capability research and safety constraints are compatible. Meta has publicly dismissed AI safety concerns as overblown, open-sourced its frontier models, and lobbied against proposed regulation. The two companies represent opposite ends of the philosophical spectrum on how the frontier governs itself.

What this deal says is that infrastructure transcends philosophy. Anthropic does not have the luxury of ideological consistency when compute ceilings limit what paid subscribers can do. Meta does not have the luxury of refusing revenue that validates its data center spending to investors. Both parties understand this. The deal is not a contradiction; it is how capital-intensive industries work when product cycles are short enough that ideology cannot wait.

The pattern is real: every major lab is diversifying compute dependency right now. OpenAI extended its Microsoft partnership while co-building Stargate. xAI built Colossus and is now renting capacity to other labs. Google runs its own TPU stack while still serving external inference. The cloud oligopoly that AI labs inherited from general software is fracturing into a multi-vendor infrastructure layer where the labs themselves are becoming each other's cloud providers.

The contrast: Meta's compute posture is two-sided. The company is simultaneously negotiating to lease infrastructure to a direct competitor (Anthropic) and publishing foundational open-source research (DINOv3, below). Google, the only lab with comparable infrastructure ownership, is doing neither right now. Its flagship model missed its third consecutive launch deadline yesterday, and its cloud is not where frontier labs are routing inference capacity.

Code Claude Code ships twice in 24 hours

Two Claude Code releases in one window. That cadence is not typical; most weeks produce one. Two in a day, with the second a security hardening release arriving the morning after the feature push, describes a specific development pattern worth understanding.

v2.1.212 (July 17) is the feature release. The biggest structural change: /fork now copies a conversation into a background session visible in claude agents, while the old in-session subagent behavior moves to /subtask. A /resume command opens a picker of past sessions, including deleted ones, for resumption as background sessions. Both commands suggest Anthropic is building toward a model where persistent background agents are a normal part of how Claude Code operates, not an edge case.

The resource caps matter for production users. Claude Code now enforces session-wide limits: 200 WebSearch calls (configurable via CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION) and 200 subagent creations (CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION). Long-running MCP tool calls over two minutes automatically move to background. A /clear resets the subagent budget. Anyone running automated Claude Code sessions in CI will hit these caps if their sessions branch heavily. The ceiling is generous for most workflows; large parallel migrations or research runs can approach it.

v2.1.214 (July 18) is the hardening release, arriving twelve hours after the feature push. Eleven permission and security fixes: Bash permission checks were bypassed on Windows PowerShell 5.1 when using dir/** allow rules, which were auto-approving writes to nested directories anywhere in the tree rather than only within the specified directory prefix. A worktree symlink vulnerability allowed file creation outside the repository boundary. File-descriptor redirects and commands over 10,000 characters were escaping permission checks. docker commands with daemon-redirect flags now require explicit permission. help and man commands with dangerous options were being auto-approved.

The worktree symlink fix is the one to audit if you run Claude Code in shared CI environments with isolation mode. An attacker with code execution in a worktree could previously create files outside the repo boundary. That path is now closed.

The EndConversation tool is new: Claude Code can now terminate sessions with abusive users or jailbreak attempts, matching behavior already live on claude.ai. A progress heartbeat fires for long-running tool calls so the terminal is no longer silent during extended operations.

The pattern: Feature push on Friday, security hardening on Saturday. The last two weeks also included v2.1.208 and v2.1.210. The velocity is not slowing, and the security-follows-feature cadence is deliberate. Anthropic is shipping fast and patching same-cycle.

News Gemini 3.5 Pro misses its third deadline

Google targeted July 17 for the general availability of Gemini 3.5 Pro. The model had already slipped twice: first from a June launch, then again in early July, after Google scrapped the original architecture for a ground-up rebuild. On July 17, nothing appeared. No model card. No pricing page. No API listing in the Gemini developer documentation.

The reported specs remain unconfirmed third-party accounts from enterprise preview participants: a 2 million token context window, a Deep Think reasoning mode, autonomous workflow chaining, pricing at $15 per million input tokens and $60 per million output. None of that is official.

The contrast is clean. OpenAI launched GPT-5.6 Sol, Terra, and Luna on July 9, ten days before Google's missed target. Anthropic has Fable 5 in production. xAI shipped Grok 4.5 on July 16, one day before Google's failed date. The only major frontier lab without a live, purchasable flagship as of July 18 is Google. Three missed deadlines in sequence is an execution story, not a model quality story. Reports from TechTimes (July 16) describe Google evaluating a stopgap enterprise preview release while the full launch slips further.

Research News Meta publishes DINOv3, xAI opens its government business

Two labs, two different definitions of shipping.

Meta's DINOv3 (published July 17) is a 7B-parameter self-supervised vision backbone trained on 1.7 billion images: a 6x model size increase and 12x data increase over DINOv2. The key technical contribution is Gram anchoring, which solves a known failure mode where dense feature maps degrade over long training schedules as the self-supervised signal becomes too smooth. Gram anchoring preserves local structure during extended training. The result: for the first time, a single frozen backbone outperforms task-specific architectures simultaneously on object detection, semantic segmentation, and satellite canopy height mapping. Meta released the full suite open-source under a commercial license, including a dedicated backbone trained on MAXAR satellite imagery.

For builders: computer vision tasks that previously required fine-tuning a task-specific backbone can now use DINOv3 frozen. The Gram anchoring technique is the transferable insight; the paper is worth reading if you are running any vision pipeline at scale.

xAI's Grok for Government launched July 17. Federal, state, and national security customers now get access to Grok 4 with Deep Search, Tool Use, and custom application development capabilities. The announcement coincided with DoD confirmation of $200 million contracts awarded to each of xAI, Anthropic, Google, and OpenAI for agentic AI workflows in defense and national security, announced July 14. xAI is the last of the four to formalize a government offering; Anthropic and OpenAI have had government programs running for longer.

The contrast: Meta is shipping foundational open-source research and positioning itself as critical infrastructure for its competitors. xAI is building a commercial government business off Grok 4 and Colossus revenue. OpenAI's July 17 entry was a CFO blog post about enterprise ROI metrics, which is positioning, not a release. Google missed its deadline. Four labs, four different postures on the same day.