AI Sector Daily Digest: August 1, 2026 — MediaTek’s $5B chip bet and the new AI guardrails

In brief

1. MediaTek opens a $5 billion financing framework for AI chips

• MediaTek’s board approved a $5 billion discretionary financing budget for long-term growth, including its expansion into custom data-center AI chips. This is an authorization framework, not a disclosed debt or equity raise. 1
• The company says its first custom AI chip is ready for production in the fourth quarter, while a second chip is scheduled for volume production in 2028. MediaTek expects its data-center AI-chip business to generate more than $2 billion in 2026 revenue. 1
• MediaTek raised its estimate of the 2027 custom-ASIC market to $80 billion and its target share to 15%–20%. The important caveat is that these are management targets arriving alongside a 20% year-over-year drop in mobile-chip revenue, not yet delivered data-center results. 1

2. Google rolls back Google Earth’s Nano Banana feature after one day

• Google launched a Google Earth feature that let users use Nano Banana 2 to generate custom images on top of satellite, aerial, and 3D imagery. The stated uses included visualizing historical scenes and planning changes to real locations. 2
• One day later, Google said it was rolling the feature back after people shared generated screenshots that appeared to violate its policies. TechCrunch reported the criticism centered on the risk that fabricated images could be mistaken for geospatial evidence. 3
• This is a rollback while guardrails are rebuilt, not a permanent cancellation. Google said generated images did not appear in the main Earth experience for other users and carried AI watermarks, but those controls did not prevent the screenshots from circulating outside the product. 2

3. Snapchat removes wholly AI-generated videos from Spotlight recommendations

• Snapchat changed Spotlight’s recommendation system so wholly AI-generated videos are no longer eligible for recommendation, effective this month. The change affects distribution, not merely labeling. 4
• Videos enhanced or edited with Snapchat’s own AI tools can still be recommended and will carry transparency indicators. Snapchat is therefore drawing a line between AI-assisted creation and fully synthetic submissions. 4
• The move gives creators a direct incentive to keep a human-made component in content that depends on algorithmic reach. It also shows platforms treating synthetic-content quality as a ranking and rewards problem, not only a moderation problem. 5

4. Smallest.ai raises $13 million for real-time voice agents

• Smallest.ai raised a $13 million Series A led by Seligman Ventures, with participation from Sierra Ventures and 3one4 Capital. The round brings the startup’s total funding to more than $21 million. 6
• Its specialized voice model is designed to listen, reason, and speak in parallel, reducing the pauses that make phone-based AI feel artificial. When a request falls outside its narrow knowledge base, the system can hand it to a larger foundation model instead of pretending to know. 6
• RingCentral and Truecaller are among its named customers. The company is betting that customer-support providers will buy a voice-specific layer rather than build one themselves, a specialization trade-off that will be tested against ElevenLabs, Cartesia, and regional competitors. 6

5. Ellis AI emerges with $10 million for private-credit operations

• Ellis AI emerged from stealth with a $10 million seed round from investors including First Round Capital, 645 Ventures, Harlem Capital, Khosla Ventures, and Thrive Capital. The company was founded by Ryan Williams, who previously co-created real-estate investment platform Cadre. 7
• Its agents connect the documents, spreadsheets, accounting systems, and correspondence that private-credit firms use to monitor portfolios, investigate discrepancies, prepare reports, and close a fund’s books. The product is aimed at a fragmented workflow rather than a general-purpose chatbot. 7
• Material decisions and actions remain with human experts, according to Williams. That constraint is the practical test for this class of enterprise agent: whether it can remove repetitive reconciliation work without taking ownership of judgment-heavy financial decisions. 7

The read-through