NVIDIA Releases "Transaction Foundation Model" for Financial Institutions

On June 2, NVIDIA released "Build Your Own Transaction Foundation Model," a developer sample that enables financial institutions to build their own Transaction Foundation Model. The model provides a unified understanding of consumer behavior from billions of financial transactions, and can consolidate task-specific models that have traditionally been built separately — such as fraud detection, credit decisioning, and recommendations — into a single foundation model.

According to NVIDIA's 2026 State of AI in Financial Services report, 65% of financial institutions are already using AI, and nearly 90% have adopted it or are evaluating adoption. Revolut partnered with NVIDIA to build PRAGMA, a transformer-based foundation model trained on 24 billion events, which outperforms traditional task-specific models without any feature engineering.

Mastercard is also developing its own large-scale tabular foundation model trained on billions of anonymized transactions, working with AWS and Databricks to apply it to fraud detection, cybersecurity, and loyalty programs. Similarly, Stripe leverages a foundation model that understands transaction behavior in its full context, and reportedly blocked approximately $112 billion in fraud last year.

ARM May Hit Its $15 Billion AI Chip Revenue Target "Ahead of Schedule"

In an interview with Bloomberg Television, ARM Holdings CEO Rene Haas said that, with demand from the AI boom exceeding expectations, the company may achieve its $15 billion chip sales target — set for the late 2020s — "ahead of schedule." Overwhelming demand for data centers and AI services is acting as a tailwind.

NVIDIA Releases "Cosmos 3," an Omnimodal World Model

NVIDIA has released a new world model, "Cosmos 3," on Hugging Face. It is an omnimodal model that can generate from any combination of text, images, video, audio, and action commands, and is designed as a foundation for Physical AI and robotics. It comes in two sizes: Nano (16B parameters) and Super (64B parameters).

Tencent to Begin Public Testing of WeChat AI Agent

Tencent is testing a prototype of an AI agent built into WeChat and expects to begin a public launch as early as this month after completing compliance processes, the Financial Times reports. The company plans a phased rollout following testing with a small group of external users, and on the news, Tencent's stock saw its sharpest rally in more than three years.

Chinese Lab-Grown Diamonds Emerge as Coolant for AI Semiconductors

Lab-grown diamonds made in Chinese laboratories are drawing attention as an unexpected beneficiary of the AI boom. Synthetic diamonds, once used mainly for jewelry, are now being adopted as coolants (heat spreaders) for AI semiconductors. Because they enable denser, higher-performance chips, several Chinese producers have reportedly completed customer validation and begun commercial shipments.

SBI Group Partners with Anthropic to Deploy Claude Company-Wide

SBI Group has announced a partnership with Anthropic to roll out Claude to all of its officers and employees. In addition, the company will conduct the first joint validation at a Japanese financial institution for the adoption of "Claude Security," a security tool that automatically scans for vulnerabilities and generates patches.

Moss TTS 1.5 — The Best English Voice Cloning So Far

Moss TTS 1.5 8B, an open-source voice cloning model, is attracting attention in the LocalLLaMA community. Even with default settings, it delivers quality surpassing Fish Audio S2 Pro and Qwen 3 TTS, and is said to improve further with parameter tuning.