Political constraints now the binding limit on AI infrastructure buildout

Open model inference platforms accelerate: Fireworks serving 40T+ tokens daily with 2x growth in customized model vertical

Nathan Lambert's analysis of public disclosures from Baseten, Fireworks, and Together AI reveals sustained acceleration in the open model inference market, with Fireworks reporting 2x growth in their customized model vertical since summer. This matters for understanding whether open models represent a meaningful economic category or remain marginal relative to closed model providers.

Democratic sweep threatens AI data center buildout; Oracle scraps $7B Tencent chip lease; SpaceX-Qualcomm ink 5G deal

Democratic gubernatorial wins in Texas, Georgia, Ohio, and Pennsylvania mean approximately half of pending AI data center projects now face tougher approval processes, with voters in 38+ jurisdictions passing restrictions and 56% citing data centers as a major election factor. This creates material risk to hyperscaler capex plans and could accelerate interest in alternative infrastructure like space-based data centers. Separately, Oracle has scrapped a ~$7B, five-year deal to lease as many as 100K advanced AI chips to Tencent after bowing to pressure from the Trump administration, while Chinese companies ordered ~6M Nvidia RTX Pro 5500 chips worth ~$70B but expect only ~1% delivery within four months, significantly constraining China's AI buildout capacity. In parallel, SpaceX selected Qualcomm's Dragonwing platform (QRU100, QDU100, 64T64R Massive MIMO radios) to build Starlink Mobile's terrestrial 5G network, marking the first new national US wireless carrier since T-Mobile/Sprint merged in 2020 and opening a potential path to AI-native 6G infrastructure.

Claude accelerates custom AI chip design with 70% faster iteration; Anthropic and OpenAI demonstrate model-driven hardware feedback loops

Anthropic's 15-person chip-design team is using Claude to accelerate custom AI chip development, with engineers reporting significant speedups compared to traditional Nvidia/AMD processes. OpenAI similarly used its models with Broadcom to complete the Jalapeño chip from design to tape-out in nine months, exemplifying the "Escher Hands" feedback loop where better models enable better hardware that runs better models.

OpenAI's autonomous cyber pipeline patches 948 Linux kernel vulnerabilities for $1.2M, discovers 10+ novel exploitation techniques

OpenAI's security researchers built an autonomous pipeline for vulnerability discovery, exploitation and patching that validated and generated patches for 948 local privilege escalation (LPE) vulnerabilities in the Linux kernel at a total API cost of $1.2M. The agents discovered and used at least 10 Linux exploitation techniques that OpenAI's team and consulted experts consider novel, demonstrating models going beyond applying known methods. Researchers typically find about 20 Linux LPEs per year; the model found approximately 1,000 in one week, with OpenAI now allocating "a significant amount" of frontier compute toward pro bono open-source patching.

Open-weight model wave: Kimi K3.1 Pro ranks #1 on Agent Arena; Ripple-65B, SWE-2 Flash, Hy4 intensify agentic coding competition

Moonshot released weights for Kimi K3.1 Pro, a post-training update with expanded RL for coding, financial research and long-horizon agentic work. On Agent Arena it ranks #1 among open models and #13 overall, with +5.1% net improvement across more than 17.5K agentic sessions and +4.2pp improvement over K3, with Day-0 support across AMD, NVIDIA, Rubin and TPU via vLLM and SGLang. Ricursive's Ripple-65B, a 65B dense model pretrained from scratch, claims best US open-weight agentic coding performance at 80.5% on Terminal-Bench 2.1 with 12T training tokens and a $32M budget. Cognition's SWE-2 Flash, post-trained on Inkling-Small, costs 20% of SWE-2 and runs 3x faster, built in 6 weeks. Tencent's Hy4 reports 75.2 on APEX-Agents with API pricing at $0.834/$2.501 per million input/output tokens.

CrowdStrike acquires $495M in AI security startups, launches agent kill switch with Nvidia Sentry integration

CrowdStrike is acquiring Dreadnode ($95M) for pre-deployment vulnerability testing and Fiddler AI ($400M) for post-deployment monitoring, plus announcing an AI kill switch to revoke rogue agent access. The company is integrating with Nvidia Sentry to enforce security rules on separate hardware outside the agent's control, addressing risks from legitimate-but-bad AI decisions.

AI on the Ballot

Stratechery published a strategic essay analyzing how the 2026 midterm election results—particularly Democratic victories on anti-data-center platforms in Texas and elsewhere—represent a fundamental shift in AI infrastructure constraints from technical (compute, memory, power) to political. Main takeaways:

• Political opposition to data center construction has become the binding constraint on AI buildout, with nearly every local ballot measure opposing data centers winning or winning decisively across the US.
• Hyperscalers (AWS, Google, Microsoft, Meta) are positioned to gain market share relative to neoclouds (CoreWeave, Lambda Labs, etc.) because they already have permitted sites and greater political sway, while neoclouds face dual constraints of capital scarcity and permitting difficulty.
• Nvidia faces a mixed outcome: fewer total data centers built reduces chip demand, but constrained power/capacity makes Nvidia's efficiency advantage (tokens/megawatt) worth significantly more money than in an unconstrained scenario.
• China gains relative advantage because it can build power generation and operate iso-compute (power-agnostic) rather than iso-power, offsetting its lack of leading-edge chips through brute-force electricity deployment.
• SpaceX data centers in space become economically viable if terrestrial permitting becomes prohibitively expensive, shifting from theoretical upside to realistic scenario.

Interconnects Dashboard: Open Model Inference Share, Downloads, and Adoption Metrics

Interconnects published a comprehensive dashboard tracking open model adoption, inference volume share, and competitive dynamics across regions. Main takeaways:

• GLM-5.3-Flash leads adoption at 12.2× relative adoption metric (RAM) with 11M downloads; Chinese models dominate top 5 with GLM-5.3, Qwen variants, and DeepSeek-V4.1-Flash
• OpenRouter weekly open-weight token share shows Chinese models at 46% of frontier open-weight inference volume as of Oct 30, 2026
• Hugging Face derivative model creation peaked in mid-2026; tracked open-weight model families now span 51 families with daily download attribution by region
• arXiv paper mentions of LLMs reached 61.05% of AI papers in Nov 2026 (up from 53.14% in Jul 2026), with regional breakdown showing adoption patterns
• Inference volume data from Together AI, Baseten, Fireworks, and OpenRouter shows fragmentation: Fireworks at 43T+ tokens/day, Together AI at 5T+ tokens/day, indicating no single dominant inference platform

Riot Signs $28 Billion AI Lease for Entire Corsicana Campus

The Defiant reported on Riot Platforms' 15-year, $28 billion lease of its entire Corsicana campus (910 MW) to Anthropic, representing a major infrastructure commitment and shift in Riot's business model. Main takeaways:

• Riot's total contracted data center capacity now reaches 1,151 MW (~$38B long-term revenue), with Anthropic accounting for 1,101 MW across Corsicana and Rockdale leases
• Powered-land lease structure shifts capital and operating costs to Anthropic; Riot's $5.5B capex covers site/electrical work and module reimbursements, financed via project debt
• Riot generated $116M from bitcoin mining in Q3 2026 (64% of revenue) and sold 20,264 BTC for $1.5B YTD to fund data center buildout, indicating mining revenue now funds infrastructure expansion
• Phased delivery: 400 MW by Sept 2027, 200 MW by Jan 2028, 310 MW by July 2028, with rent due 60 days after each phase delivery
• Anthropic's dual commitments (Corsicana + Rockdale) represent ~$37B of its disclosed $518B infrastructure commitments, signaling major capacity concentration with single operator

FlashProve: Provable Inference via zkSNARKs for Cryptographic Verification of AI Model Outputs

Anthropic published a technical report on provable inference using zero-knowledge succinct arguments (zkSNARKs) to cryptographically verify that AI model outputs come from specific model weights. Main takeaways:

• FlashProve generates cryptographic proofs at ~110 tokens/second on a single H100 GPU for a 400B-parameter MoE model (Llama 4 Maverick), with proof size ~19 MB per 1,024-token request, representing 70-95x GPU overhead versus vLLM inference.
• Proofs enable verification that responses came from committed model weights without revealing weights or architecture, addressing tamper detection, regulatory pacing verification, and text provenance use cases.
• System includes executable C++ prover/verifiers, Lean formal verification proofs of security guarantees, and protocol specification released under Apache 2.0 license, designed for auditor/regulator tooling and infrastructure security.