AI products are accumulating dependencies on data capture, bundled cloud services, persistent memory, and scarce silicon. Meta's workflow traces and ChatGPT's background memory enlarge the data surface, while Lovable's Google agreement and TSMC's capacity warning show how commercial and physical concentration can constrain model choice.
1. Meta's workflow-capture tool reaches across privacy boundaries
Meta's Model Capability Initiative collects AI training data from mouse movements, clicks, navigation, and other activity across more than 200 apps and websites on participating US employees' computers. Internal material reviewed by Reuters said messages from non-US colleagues could also be captured when they interact with those employees.
Meta says the system focuses on interaction patterns and includes privacy mitigations. Complete task traces can nevertheless join clicks, messages, code, and clipboard content into a training record whose sensitivity exceeds any individual event.
Sources: Reuters investigation of Meta's workflow-capture tool
2. Lovable commits to a much larger Google Cloud footprint
Lovable and Google Cloud announced an expanded multiyear agreement covering infrastructure, Gemini models, Claude access, a Wiz security integration, and distribution through Google's enterprise agent marketplace. TechCrunch reported, citing a person familiar with the deal, that Lovable's cloud footprint would increase fivefold.
The agreement bundles compute, model access, security tooling, procurement, and marketplace distribution under one provider. That shortens Lovable's route into enterprise accounts while moving switching costs from application code into contracts and integrated cloud services.
Sources: Google Cloud's Lovable partnership announcement · TechCrunch on the reported fivefold expansion
3. ChatGPT's Dreaming system synthesizes memory in the background
OpenAI began rolling out a new memory architecture to US Plus and Pro users, with other countries and lower-priced plans to follow. The Dreaming method continuously synthesizes past chats into reviewable memory, aiming to preserve relevant context, follow preferences, and retire facts that become stale.
OpenAI says recent changes cut the system's compute cost about fivefold. Continuous synthesis makes correction and deletion harder than removing one chat because a mistaken fact can influence a derived memory before the original exchange disappears.
Sources: OpenAI's Dreaming memory release
4. TSMC warns AI chip demand still exceeds available supply
TSMC CEO C.C. Wei said customer demand was so high that the company could support only part of it, despite ongoing expansion. He also said fulfilling customers' needs through US production could take a very long time; TSMC has one Arizona factory operating and further plants planned.
The shortage sits upstream of model roadmaps and cloud quotas across fabrication, advanced packaging, memory, and power. Long commitments gain leverage under scarcity, leaving smaller deployments more exposed to allocation changes and component-price inflation.
Sources: Reuters on TSMC's AI-demand outlook · The Verge on TSMC's capacity warning