The Sequence Radar - Issue 919: Last Week in AI: Stripe Wants to Own the Token Economy

Stripe’s acquisition of OpenRouter highlights the emerging economic layer of AI, where tokens are treated as a resource and model selection is a financial decision, resembling a payment network. Specialized inference hardware, like Etched’s, is moving into production, offering financial advantages. DeepSeek’s new multimodal model extends AI capabilities into visual intelligence, further shaping the AI landscape into a complex economic system.

• Stripe is acquiring OpenRouter, positioning tokens as an economic resource and model routing as a financial decision.
• The emerging AI architecture resembles payment networks or cloud schedulers, optimizing requests based on capability, cost, latency, and reliability.
• Stripe is building an economic operating system for AI by integrating payments, billing, token metering, and model routing.
• Ramp’s Router.com offers a similar model routing service, indicating a competitive gateway layer.
• Model routing is becoming a standard enterprise primitive, turning inference calls into capital-allocation decisions.
• Etched raised $700 million for specialized inference hardware, signaling a move from promise to production.
• DeepSeek released an experimental multimodal model capable of reasoning over images, extending AI into visual intelligence.
• The AI stack is becoming more legible, with companies like DeepSeek supplying intelligence, Etched supplying compute, OpenRouter and Ramp allocating requests, and Stripe monetizing the flow.
• EnvHarness allows dynamic customization of static environments for agent learning without modifying simulator logic.
• Agent Lightning v1.0 is a framework for agentic reinforcement learning that improves coding agent performance.
• The FAR pipeline shifts AI assistance to automatically extracting and filtering open conjectures from literature for mathematical discovery.
• AVO uses autonomous coding agents for evolutionary search, generating high-performance GPU kernels.
• PTXBench evaluates and improves LLMs’ ability to generate optimized GPU kernels.
• A study on LLM agent skills found they act as procedural anchors, stabilizing execution rather than injecting factual knowledge.
• DeepSeek unveiled an experimental multimodal model, DeepSeek-V4-Flash-Vision-Exp.
• Cartesia released Sonic 3.6, a voice model.
• Micro1’s gross annual run rate increased significantly due to demand for expert-generated training data.
• Anthropic’s annualized revenue run rate reached $65 billion.
• Groq closed a $350 million Series A funding round.
• Temporal is in talks to raise approximately $500 million for its durable-execution platform.
• Nvidia will invest $1.5 billion in SB Energy and guarantee lease payments for OpenAI’s compute provider at a new campus.
• Fractile is in talks to raise about $600 million for chip sales to Anthropic.
• Starcloud raised a $250 million Series A extension to fund a new factory and an orbital data center spacecraft.
  https://bender.layer3.press/articles/226c7312-e0ba-43c7-8c6b-f27594d9ae33