Nvidia halves its OpenAI data-center backstop: Per a WSJ exclusive, Nvidia is cutting its planned $250 billion guarantee for OpenAI’s Ohio data center to roughly the low $100 billions, under investor pressure — an early sign that AI-infrastructure financing is being re-priced. WSJ · Reuters
Anthropic explains how Claude’s new text watermarks work: The company says the EU-driven watermarking works by adjusting the choice of “inconsequential words” in generated text, turning content-provenance tracing into a shipping product feature. TechCrunch
Google now lets users remove visible watermarks from AI generations: The opposite of Anthropic’s approach — Google lets users strip the visible watermark from AI images and video, showing how provenance policies are diverging across platforms. TechCrunch
Apple is training its own China AI model with Alibaba’s help: Apple is reportedly building a China-specific model with Alibaba and won unprecedented Beijing clearance — the first foreign firm to do so, marking a shift in how Western companies enter the China market. 9to5Mac
Anthropic’s Q2 revenue tops $11.5B, but its IPO valuation hinges on a $190-200B 2028 forecast: Revenue reportedly jumped at least 14x year over year, fueling IPO expectations, yet Reuters reports the valuation is tied to 2028 revenue projections. Frontier-lab revenue is now being stress-tested in public. Fortune · CNBC
Watermarking is becoming the new provenance battleground: Anthropic adds watermarks for transparency regulation while Google removes them for user convenience — provenance and authenticity policy is turning into a product differentiator. TechCrunch · TechCrunch
Thrive’s Joshua Kushner warns Silicon Valley VCs over AI euphoria: The head of one of AI’s biggest backers is publicly cautioning against valuation euphoria, adding to signals that capital markets are re-examining the AI growth narrative. TechCrunch
The gap for AI provenance and authenticity tooling: As watermarking and content labeling become mandatory, tools that verify authenticity and log provenance for generated content can absorb real compliance demand. TechCrunch
Hedge model and infra costs ahead of re-pricing: Nvidia’s scaled-back backstop points to repricing in large-scale cloud and infra capacity. Builders should structure costs around multi-provider and open-weight models rather than a single vendor. Reuters · The Decoder
A local-partner playbook for entering China: Apple’s move to train a local model with Alibaba shows product teams a workable “local partner + custom model” path into the China AI market. 9to5Mac
Anthropic’s $200B-by-2028 bet: A valuation pegged to a three-year-out revenue forecast means a growth stumble could sharply reset expectations. The bubble debate is unlikely to quiet soon. Fortune · The Decoder
The final size of Nvidia’s backstop: The WSJ exclusive says the guarantee was cut sharply, but the confirmed figure and terms are still unclear — and so is the ripple effect across AI-infrastructure financing. WSJ
User backlash against watermarking: Some Claude users are protesting that watermarks could be used to catch cheating or track their work. How the tension between compliance and user trust resolves is an open question. TechCrunch