Google to invest up to $40 billion in Anthropic
Google is investing up to $40 billion in Anthropic — $10 billion in immediate cash plus $30 billion contingent on performance milestones — as the competition for frontier AI compute intensifies. Meta signed a multi-year deal for millions of Amazon Graviton 5 CPU cores specifically for agentic inference workloads, signaling that the agent-era hardware stack is diverging from the training-era GPU focus. DeepSeek published detailed architecture specs behind V4’s agent-optimized 1M-token context, while GPT-5.5’s API access finally opened after its launch-day delay.
Funding & Business #
Google Plans to Invest Up to $40 Billion in Anthropic #
Bloomberg / TechCrunch / Ars Technica / Hacker News (602 points)
Google will invest $10 billion immediately at a $350 billion valuation, with an additional $30 billion contingent on Anthropic hitting performance milestones. Google Cloud will also provide 5 gigawatts of computing capacity over five years using its TPU infrastructure. The deal follows Amazon’s recent investment in Anthropic and intensifies the pattern where hyperscaler capital and compute commitments — not model capability alone — determine which frontier labs survive.
ComfyUI Hits $500M Valuation #
TechCrunch
ComfyUI raised $30 million in Series B funding at a $500 million valuation, led by Craft Ventures. The node-based workflow platform for AI image, video, and audio generation claims over 4 million users, with “ComfyUI artist” now appearing as a job title on studio job boards. The valuation reflects growing demand for fine-grained control over generative media — where prompt-based tools “get only 60-80% there,” ComfyUI’s modular interface enables precise element-level adjustments.
Meta’s Loss Is Thinking Machines’ Gain #
TechCrunch
Thinking Machines Lab is hiring from Meta faster than Meta is poaching from TML, pulling in prominent researchers including Soumith Chintala (PyTorch co-creator) and Piotr Dollar (Segment Anything co-author). TML’s $12 billion valuation gives it equity leverage against Meta’s cash-heavy compensation packages. The bidirectional talent flow, occurring alongside Meta’s 10% layoffs announced this week, suggests that the AI talent market is increasingly favoring well-funded startups over incumbents cutting headcount.
Model Releases #
DeepSeek-V4: A Million-Token Context That Agents Can Actually Use #
Hugging Face Blog / DeepSeek / MIT Technology Review / The Register / Hacker News (1,930 points)
The Hugging Face deep-dive on V4’s architecture reveals how DeepSeek achieved a practical 1M-token agent context: hybrid attention alternates Compressed Sparse Attention (4x KV compression plus top-k block selection) with Heavily Compressed Attention (128x compression with dense attention over the compressed sequence), reducing KV cache to roughly 2% of standard implementations. Agent-specific features include interleaved thinking that preserves reasoning across tool-call boundaries, XML-based tool calls with a dedicated delimiter token to eliminate JSON escaping failures, and a Rust-based sandbox for RL-based agent training. V4 also extends support for Huawei Ascend NPUs, making it the first frontier-competitive model runnable outside the NVIDIA ecosystem — a detail with geopolitical implications given ongoing US-China chip restrictions.
GPT-5.5 API Access Opens with Prompting Guide #
OpenAI / Simon Willison
GPT-5.5’s API access, delayed at yesterday’s launch, is now live. OpenAI published a detailed prompting guide for the model, including a recommendation for streaming thinking indicators in applications with extended reasoning time. Simon Willison’s llm tool immediately shipped support (v0.31), adding a verbosity parameter for GPT-5+ models. The API delay at launch — which prompted Willison to build a Codex CLI credential workaround within hours — underscored developer impatience with model access gatekeeping, a friction that open-weight alternatives like DeepSeek V4 avoid entirely.
Infrastructure #
Meta Signs Deal for Millions of Amazon Graviton 5 CPU Cores #
TechCrunch / The Register
Meta will deploy tens of millions of AWS Graviton 5 ARM-based CPU cores for AI agentic workloads including real-time reasoning, code generation, search, and multi-step task coordination — tasks that differ fundamentally from GPU-intensive model training. The multi-year deal represents one of the largest cloud compute commitments in AWS history and signals that agentic inference is creating a distinct hardware demand category. For infrastructure planners, the deal validates the thesis that agent workloads will run on CPUs at scale, not GPUs, shifting the cost structure of AI deployment.
Intel Bets on AI Inference to Restore CPU to the Top Table #
The Register
Intel CEO Lip-Bu Tan argued that the CPU is “reinserting itself as the indispensable foundation of the AI era” through inference and agentic workloads on edge devices, robots, and agents. Intel’s Xeon 6 processors have been selected as host CPUs for NVIDIA’s DGX systems, and the company is co-developing Infrastructure Processing Units with Google. Intel expects its 14A process node to generate design commitments in H2 2026 — the same unfinished node Tesla is betting on for its AI chips.
Marked-Up Mac Minis Flood eBay as AI Demand Drives Shortage #
TechCrunch
Apple’s M4 Mac mini is experiencing unprecedented shortages as the compact desktop has become the preferred device for running local AI models. The machines are valued for power efficiency and quiet 24/7 operation, making them ideal for on-device AI testing and deployment. The shortage — and resulting eBay markups — demonstrates that local AI inference demand is real enough to disrupt consumer hardware supply chains.
Samsung Worried About First-Ever Smartphone Loss from AI Memory Shortage #
Ars Technica
Samsung executives are concerned the company could lose money on smartphones for the first time, with the AI-driven memory shortage cited as a key factor. The cascading effect — AI demand for HBM and DRAM driving up memory prices, which increases component costs for consumer electronics — illustrates how AI infrastructure investment is creating collateral damage in adjacent hardware markets.
Security #
Open Source Models Can Match Mythos for Bug-Finding, Says OpenAI’s First Security Hire #
The Register
Ari Herbert-Voss, CEO of RunSybil and OpenAI’s first security hire, argued at Black Hat Asia that organizations can achieve Mythos-comparable bug detection using multiple open source models working together in scaffolding arrangements. While Mythos excels at finding both simple and complex vulnerabilities individually, Herbert-Voss contends the cost and accessibility of open source alternatives makes them necessary for broad adoption. He dismissed job displacement concerns, predicting that security workers will remain busy managing the high volume of AI-generated bug reports rather than being replaced by them.
System Prompt Prohibitions Decay in Long-Context LLM Agents #
arXiv:cs.AI
In a 4,416-trial study across three LLMs, prohibition-type constraints in system prompts (e.g., “never reveal credentials”) degraded by 47-78% at 128K tokens, while requirement-type constraints persisted — an asymmetry the authors term “Security-Recall Divergence.” The finding means prohibitions cannot be trusted to hold across extended agent conversations, and supplementary enforcement mechanisms are required. For anyone deploying long-context agents with safety-critical system prompt instructions, this maps a structural vulnerability that context pressure alone — without adversarial prompts — can exploit.
Comic-Template Jailbreaks Achieve High Attack Success Against Multimodal LLMs #
arXiv:cs.AI
Embedding harmful goals inside three-panel comic narratives and prompting the model to “complete the comic” bypasses safety alignment in multimodal LLMs at rates significantly above text-only jailbreaks. The attack exploits the visual grounding pathway, which receives less safety training than the text channel. For teams deploying multimodal models with image input, this maps a vulnerability class where existing text-focused safety measures provide no protection.
Developer Tools #
llm 0.31: GPT-5.5 Support and Verbosity Controls #
Simon Willison
Simon Willison’s llm CLI tool shipped GPT-5.5 support and a new verbosity parameter for GPT-5+ OpenAI models (-o verbosity low). The rapid turnaround — shipping model support within hours of API access — demonstrates the speed of the open-source tooling ecosystem in tracking frontier model releases. The verbosity control is notable as a first-class API parameter, giving developers explicit control over output length that was previously only achievable through prompting.
Research & Papers #
Tool Attention Is All You Need: Eliminating the MCP/Tools Tax #
arXiv:cs.AI
Quantifies the “MCP Tax” — the per-turn overhead of stateless, eager schema injection in multi-server tool deployments, which practitioner reports place between 10K and 60K tokens. This payload inflates the KV cache, wastes compute, and is associated with reasoning degradation as context utilization rises. The proposed solution — dynamic tool gating that injects only schemas the model is likely to need, plus lazy schema loading that defers full definitions until selection — cuts overhead to near-zero. For anyone building agentic systems with MCP-based tool access, this paper names and quantifies a pain point many teams have encountered empirically but lacked a systematic framework to address.
Speculative Actions: Lossless Acceleration for Agentic Systems #
arXiv:cs.AI
Applies the speculative decoding principle to agent actions: predict likely future actions, execute them in parallel, and verify correctness afterward, discarding incorrect speculations. The framework is lossless — producing identical results to sequential execution while significantly reducing wall-clock time in environments where each action incurs substantial API latency. For teams deploying agents in latency-sensitive settings, this converts the serial bottleneck of action-observation loops into a parallelizable pipeline without sacrificing correctness.
Thinking with Reasoning Skills: Fewer Tokens, More Accuracy #
arXiv:cs.AI
Instead of reasoning from scratch each time, this approach distills successful reasoning patterns from extensive deliberation into reusable “skills” that can be retrieved and applied to similar problems at inference time. The method claims accuracy gains while reducing token usage — directly attacking the cost-quality tradeoff in reasoning models. For production systems paying per-token for chain-of-thought, the approach offers a path to amortizing reasoning costs across similar queries.
Regulatory & Policy #
Anthropic Updates Election Safeguards Ahead of 2026 US Midterms #
Anthropic
Anthropic published updated election safeguards for Claude, reporting 95-96% scores on impartial engagement evaluations for Opus 4.7 and Sonnet 4.6, and 99.8-100% compliance on election-related safety requests. Claude now displays election banners directing users to nonpartisan voting resources and triggers web search 92-95% of the time for candidate-related queries to ensure current information. The publication establishes audit benchmarks ahead of the midterms that regulators and the public can reference.
Threads to Watch #
Agentic workloads are reshaping the hardware stack. Meta’s Graviton deal for agentic CPU inference, Intel’s bet on CPUs for agent workloads, DeepSeek V4’s agent-optimized architecture, and Mac mini shortages from local AI inference all point in the same direction: agentic inference creates hardware demand patterns fundamentally different from model training. The CPU, not the GPU, may be the primary scaling constraint for agent deployment.
The MCP overhead problem is getting formal attention. The “Tool Attention” paper quantifying 10-60K token waste per turn from MCP schema injection, combined with DeepSeek V4’s XML-based tool call system designed to reduce escaping failures, suggests that the agentic tool-use interface layer — not just the underlying model — is becoming a first-class optimization target.
Compute access is consolidating as the primary competitive moat. Google’s $40B Anthropic deal (including 5GW of compute), Meta’s AWS Graviton commitment, Amazon’s recent Anthropic investment, and Thinking Machines Lab’s Google Cloud partnership form a pattern where infrastructure partnerships — not model capabilities alone — determine which labs can compete at the frontier.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Stanford HAI — scrape: content_truncated