Self-replicating AI worms now operationally viable, frontier labs urge government intervention

Storage becomes hot inference path: NAND repricing, token caching economics, and NVIDIA's architected stack signal fundamental AI infrastructure shift

The AI infrastructure stack is undergoing a structural realignment around persistent state storage. NAND vendors booked record $18.5B in a single quarter (up 86% sequentially), with TrendForce explicitly attributing the surge to agentic AI services requiring persistent KV cache storage rather than stateless inference. Simultaneously, OpenAI, Anthropic, Google, and DeepSeek have all priced cached context at 90-99% discounts to fresh compute—with DeepSeek achieving ~1% through disk-resident cache—creating unambiguous market signals that stored state replaces recomputation for long-context and agentic workloads. NVIDIA has formalized this shift by introducing BlueField-4-based flash pods with claimed 5x tokens-per-second throughput and new KV-cache I/O profiles, paired with partnerships across the storage industry (DDN, Dell, HPE, Pure, VAST, WEKA). This signals that storage has transitioned from commodity plumbing to an architected infrastructure component with direct implications for inference economics, hardware design, and the competitive positioning of storage vendors in the AI stack.

DeepMind's compute allocation drops to 15% of Google's AI resources as GCP revenue growth takes priority

Google's internal resource allocation has shifted decisively toward commercial cloud AI over moonshot research. DeepMind's compute share fell from 45% to 15% of Google's total AI resources, coinciding with Demis Hashimoto stepping back to chief scientist and Jeff Dean departing after 27 years. The move signals a strategic pivot toward margins and accelerating GCP revenue growth over long-term fundamental research.

Chinese researchers scale agent-based simulation to one billion agents using knowledge-distilled surrogate models

Researchers have demonstrated a billion-agent simulation framework (Light Society) that uses mixture-of-models optimization and efficient event queuing to reduce computational costs by orders of magnitude while maintaining behavioral fidelity. The work grounds agents in real World Values Survey demographic data from 96,000 respondents and reveals scaling laws on Barabási–Albert scale-free networks: demographic factors (education, income) sharply correlate with influence and persuasion resistance, opinions spread via cascades rather than pairwise influence, and surrogate models closely preserve full LLM dynamics with low variance. This represents a significant engineering advance in scaling agent-based modeling from millions to billions of agents for social science applications and demonstrates emergent collective behaviors at previously unattainable scales.

Figma Q2 revenue grows 48% YoY to $370M with 136% NDR, but market reprices software valuations downward

Figma reported Q2 revenue of $370M (up 48% YoY) with 136% net dollar retention and added 1,635 customers with $100k+ ARR (up 46% YoY), achieving 84% gross margins despite AI feature costs. Despite strong fundamentals, the stock fell 17% post-earnings as the market reprices software valuations; FIG now trades at 7.3x EV/ARR, down 80% from its July 2025 IPO close.

Import AI 467: Self-sustaining AI viruses; pacing AI progress; confusion about AI and creativity

Jack Clark in Import AI (Jack Clark) published a report titled "Import AI 467: Self-sustaining AI viruses; pacing AI progress; confusion about AI and creativity". Main takeaways:

• Researchers from University of Toronto, Vector Institute, Cambridge, and ServiceNow demonstrated a self-replicating AI worm using open-weight LLMs on compromised GPU resources, achieving ~37% end-to-end attack success rate with 80% vulnerability detection, 53% exploitation, and 88% self-replication rates—proving autonomous generative cyber-threats are operationally viable.
• 1,337 employees from OpenAI, Anthropic, Google DeepMind, Meta, and other frontier labs issued a joint statement requesting US government support for international governance tools to deliberately pace automated AI development, citing competitive pressure preventing unilateral slowdown.
• Shadow evaluation of Claude Opus 4.8 on unpublished NeurIPS 2026 research found AI agents excel at engineering but fail at creative research ideation, scoring 1-2 (Strong Reject/Reject) on novel research contributions despite solving engineering prerequisites.

The Inference Unbundling: Why Prefill and Decode Are Splitting the GPU

Data Gravity published an essay on the architectural disaggregation of inference workloads across heterogeneous silicon. Main takeaways:

• Prefill (prompt processing) and decode (token generation) have fundamentally different compute/memory characteristics (100-400 FLOPs/byte vs 1-2 FLOPs/byte), requiring different silicon architectures to optimize economics
• Nvidia's $20B Groq acquisition and Cerebras' AWS Trainium partnership both implement the same strategy: pairing compute-optimized silicon (Blackwell, Trainium) with memory-optimized silicon (LPU, WSE-3) via routed inference stacks
• Output tokens command 4x price premium over input tokens (GPT-4o: $10M vs $2.5M per million tokens), making decode the margin lane and creating economic incentive for disaggregated stacks
• HBM supply constraints on decode workloads create structural advantage for SRAM-based alternatives; migration off HBM-dependent silicon reshapes HBM4 timing and frees capacity for training
• Inference orchestration layer (routing, scheduling, KV-cache management) becomes the new control point and margin layer, shifting competitive advantage from single-architecture incumbents to software-defined routing platforms