1｜OpenAI Partners with Visa, Agent Starts Accessing Payment Network
Bloomberg Law and WSJ report that OpenAI has expanded its partnership with Visa, allowing the AI agent within ChatGPT to initiate online shopping payments upon user authorization. Visa is responsible for payment authorization, risk management, and user-defined spending limits. On the same day, Mastercard launched Agent Pay for Machines, targeting machine-driven high-frequency, low-value, low-latency transactions, and stated support for multi-rail settlement with cards, accounts, and stablecoins.
This marks a significant milestone as AI agents move into the economic system. While agents' past capabilities were mainly in search, comparison, and recommendation generation, the actual transactional actions still relied on human intervention and web interaction. Card networks are now attempting to encapsulate identity, limits, authorization, clearing, and dispute resolution into underlying capabilities that agents can invoke. Whoever controls this interface holds the key to the business entry point for AI agents transitioning from suggestions to execution.
(Source: Bloomberg Law / WSJ / Mastercard / Visa)
2｜Enterprises Reassess Model Permissions After Fable 5 Release
Stratechery placed Fable 5 within Anthropic's model tiering strategy discussion, seeing it as a new precedent for opening cutting-edge capabilities to the general users. The Verge later reported that Microsoft restricted internal employee use of Claude Fable 5 due to Anthropic's data retention requirements. VentureBeat also reported that Anthropic CEO Dario Amodei called for strong model establishment similar to FAA regulation.
Yesterday, Fable 5 was about model release; today, it has turned into a governance issue. The ability of more powerful models to enter enterprise workflows depends not only on their ranking but also on factors like data retention, routing of sensitive requests, and who has the authority to prevent model deployment. While Anthropic is unleashing capabilities through tiered models, it is assigning higher-risk aspects to regulatory narratives. What enterprise clients see is not just a new model but an entire set of permissions, compliance, and responsibility boundaries.
(Source: Stratechery / The Verge / VentureBeat)
3｜Anthropic and Super Micro Secure Funding, AI Infrastructure Pressure Shifts to Capital Structure
Axios and FT report that Apollo, Blackstone, and Broadcom lead a $35 billion financing round to support Anthropic's computing power and chip leasing. Reuters and Bloomberg report that Super Micro plans to raise $7 billion to meet AI server demand. Tom's Hardware also notes that Google has reportedly booked over 3 million TPU package units of Intel's 2028 capacity.
These signals point in the same direction: AI infrastructure is not just about buying chips but about rewriting the balance sheet. Anthropic uses private credit and SPVs to underwrite computing power, Super Micro supports orders with equity and convertible securities, and Google secures advanced packaging capacity in advance. Model companies, server manufacturers, and chip packaging factories are being tied together by the same capital chain. The cost of the AI race is shifting from R&D expenses to long-term financing and capacity allocation.
(Source: Axios / FT / Reuters / Tom's Hardware)
4 | Google Faces Data Expansion and AI Search Responsibility Simultaneously
The Verge reports that Google will launch a new Search Services History setting to save interactive data such as Google Lens images, Search Live recordings, voice searches, and Translate audio, which can be used to improve services and train AI models. Ars Technica and The Decoder also report that a German court, in a preliminary ruling, held Google responsible for erroneous statements in AI Overviews.
These two threads put together show that platforms are getting more input while also taking on more output responsibility. Multimodal search allows users to submit photos, sound, and video to Google, making training data more closer to daily life. However, once an AI summary is deemed by a court as the platform's own statement, search companies can no longer claim to be mere link intermediaries. The commercial logic of AI search is tightening: to get more data, you must take on more judgment consequences.
(Source: The Verge / Ars Technica / The Decoder)
5 | CFTC Advances Prediction Market Rules, Kalshi Begins Guarding Against Information Advantage
Axios reports that the CFTC is advancing the prediction market rule, setting boundaries for event contracts tradable on platforms like Kalshi and Polymarket. WSJ and The Verge report that Kalshi has started requiring some high-risk market users to disclose their employer and has introduced risk scores and whistleblower tools to identify traders who may have access to non-public information. The Block continues to track the disagreements in the crypto tax law hearing.
The prediction market is evolving from a financial fringe product to a quasi-exchange. The issue is not just about whether you can bet on sports, politics, or corporate events, but about who knows the outcome, who can trade ahead of time, and how platforms can prove that prices are not based on insider information. Kalshi's employer disclosure rule illustrates that information asymmetry has become an institutional cost that must be addressed post-product growth. The closer event contracts get to the real market, the more they need traditional financial market integrity constraints.
(Source: Axios / WSJ / The Verge / The Block)
Also worth knowing ↓
Warner Music acquires AI attribution company Sureel AI. Music rights companies are starting to embed model training and content tracking capabilities, shifting copyright protection from post-litigation to pre-technology. (Source: TechCrunch)
ServiceNow discloses a vulnerability that exposed some customer data on the internet. Security issues in the enterprise automation platform will extend to internal processes and customer data layers. After an AI agent enters the enterprise system, such permission boundaries become more sensitive. (Source: TechCrunch)
Google DeepMind releases DiffusionGemma. Ars Technica notes that this is a new member of the Gemma 4 series, using a non-autoregressive generation path and focusing on inference speed on local hardware. (Source: Ars Technica)
Japan's three largest banks plan to start stablecoin spot trading by March 2027. The banking system continues to integrate on-chain settlement into real financial processes, as stablecoins move from the entry point of crypto assets to interbank infrastructure testing. (Source: The Block)
Hyatt CEO says AI sales tools save salespeople about a day's worth of work per week. The adoption of enterprise AI is transitioning from pilot narratives to job efficiency ledgers, with the hotel sales sector, a non-tech scenario, beginning to provide quantifiable examples. (Source: Fortune)
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
