Today's News
AINews
Stripe is negotiating to acquire OpenRouter, a model-routing aggregation layer, at a valuation around $7B—roughly 50x its $140M annualized run rate and reflecting 70% gross margins. OpenRouter handles approximately 250 trillion tokens per month, positioning the deal as a concrete signal of where infrastructure value is accruing relative to GPU supply or frontier model labs themselves.
Bloomberg via Techmeme
Anthropic has achieved an annualized revenue run rate exceeding $65 billion as of late July 2026, up more than sevenfold from approximately $9 billion at the end of 2025, according to Bloomberg reporting on figures shared with investors. The extraordinary growth demonstrates rapid monetization of frontier AI capabilities and signals competitive pressure on OpenAI's market position, with concrete metrics underscoring adoption velocity and the economics driving AI infrastructure buildout.
Artificial Analysis via AINews
Qwen 3.8-27B has achieved frontier-tier performance on the Artificial Analysis Intelligence Index, marking the first local open-weight model to compete directly with DeepSeek V4-Pro and GPT-5.6 Luna on aggregate benchmarks. The result crystallizes accelerating convergence between proprietary and open-weight models, with implications for inference economics and local deployment viability.
Cornell Researchers Expose Deep-Research Agent Vulnerability: 13-Word Reddit Poisoning Achieves 38–62% Citation Rates
Brian Roemmele from The Brian Roemmele Report.
Cornell researchers have demonstrated that deep-research agents powering systems like ChatGPT Deep Research and Gemini Deep Research can be poisoned with as little as 13 words of Reddit text, achieving citation rates of 38–62%. The WARP attack exploits how these systems repeatedly retrieve the same user-generated content, with Reddit supplying 54–71% of all UGC pages across topic clusters. This reveals a structural vulnerability in production AI systems that companies have licensed or scraped while knowing data quality was poor—compounded by the fact that OpenAI and Google formalized dependency on Reddit through 2024 licensing deals while Sam Altman held approximately 8.7% of Reddit's equity and served on its board, creating potential conflicts of interest in how these systems treat the platform as authoritative.
Nvidia's $26B Open-Model Investment and Meta's Llama Strategy Reshape AI Economics Through Indirect Monetization
Interconnects by Nathan Lambert
Nvidia is investing $26 billion to democratize model training through near-open-source releases like Nemotron (which includes training code and data), betting that open-source demand drives chip sales growth by creating a world where many companies build their own models rather than buying from Anthropic or OpenAI. Simultaneously, Meta is flooding the market with strong open-weight models like Llama to commoditize competitors' token revenue. Both strategies represent indirect monetization paths: Nvidia generates inference demand across the ecosystem, while Meta hampers competitors' API revenue growth. These approaches only succeed if economic returns exceed the investment within years or if open-model companies develop sustainable financial models competitive with API providers.