Anthropic Productizes OpenClaw Agents Amid Compute Crunch
Anthropic shipped enterprise-grade agents in 10 weeks using OpenClaw primitives, with safeguards like per-app permissions; agents explode per-user compute needs, fueling $1T Nvidia revenue forecasts and supply chain battles.
Enterprise Agents Emerge from Open Source Signals
OpenClaw's 333,000 GitHub stars validated demand for persistent, multi-device agents, prompting Anthropic to ship production-ready features in 10 weeks: Cowork (Jan 12) evolved to phone-to-desktop threads via Dispatch (Mar 17), direct computer use (Mar 23), plugins, admin controls, and scheduled tasks. Paid users now delegate browser control, app integrations, and mouse operations when APIs fail. Anthropic prioritized enterprise safety—connectors first, explicit per-app permissions, prompt-injection scanning—using Claude Code for rapid iteration. This wraps raw open source primitives in deployable permission models, turning a single user into a persistent compute load via token-intensive tasks like spreadsheet automation.
Cursor's Composer 2 applies the same formula: fine-tune Moonshot's open Kimi K2.5 with domain RL for coding agents scoring 61.7 on Terminal-Bench 2.0 and 73.7 on SWE-bench Multilingual, at $0.50/M input tokens. Initial non-disclosure of the base drew criticism, but authorized Fireworks partnership confirms the template: strong open base + RL + superior UX beats closed models in apps.
Supply Chain Lags Fuel Massive Buildouts
Agent workflows multiply token use beyond chat, amplifying infrastructure strain. Nvidia's Jensen Huang forecasts $1T+ cumulative Blackwell/Rubin revenue by 2027 (conservative, excluding extras), calling Anthropic's $1T non-infra AI revenue by 2030 "very conservative" as enterprises resell tokens. Nvidia invests $4B in optical suppliers like Coherent/Lumentum to break photonics bottlenecks.
Downstream hesitancy persists: Broadcom flags TSMC as a wafer bottleneck until 2027, with lasers/PCBs scarce; memory prices tripled; Samsung demands 3-5 year contracts. Elon Musk's Terafab counters with a $20-25B initial Austin fab (scaling to 1TW annual compute output, vs. U.S.'s 0.5TW grid), 80% for space data centers. Tesla's AI5/AI6 chips start Samsung Texas production H2 2027; SpaceX builds North America's largest PCB facility. Yet fab ops gap looms—TSMC's 50,000 engineers and ASML's €39B-backlogged $200-400M EUV machines (165 tons each) block fast ramps. Terafab echoes Tesla's 4680 battery delays but forced supplier curves up.
Google counters via energy: $4.75B Intersect buy, 500MW Kairos SMRs by 2035, 615MW Duane Arnold restart, 200MW fusion, 600MW+ nuclear sites, 1GW curtailments. Ruth Porat warns U.S. energy lags AI. Meta rents/buys Google TPUs; Anthropic taps 1GW+.
OpenAI-Nvidia's $100B deal deploys 10GW Rubin systems from H2 2026, with progressive investment.
Model Releases Prioritize Efficiency and Agents
OpenAI's GPT-5.4 mini (2x faster than GPT-5 mini, 54.4% SWE-Bench Pro, $0.75/$4.50 per 1M tokens, 400K context, text+image+tools) and nano ($0.20/$1.25, for classification/coding subs) target high-throughput agents. Mistral Small 4 (119B MoE, 6-6.5B active, 256K context, text+image) unifies reasoning/coding/agents with 40% latency cuts, 3x throughput, configurable reasoning. Xiaomi MiMo-V2-Pro (1T MoE/42B active, 1M context) excels on agent evals (PinchBench 81.0, ClawEval 61.5). Nvidia Nemotron-Cascade 2-30B-A3B (~3B active) hits gold on 2025 IMO/IOI/ICPC via multi-domain RL distillation. Mamba-3 SSM beats transformers on prefill+decode latency with exponential-trapezoidal recurrence and complex states.
Agents demand early adoption: top-stack conviction outpaces mid-chain investment, settling in 2 years. Recursive self-improvement enables AI-first fabs to leapfrog incumbents.