Overview

Microsoft's reported China sales and JPMorgan's Hong Kong restriction show access diverging inside global cloud and enterprise networks. A $7 trillion policy proposal widens the ownership debate, while deployment simulation and BRIDGE quantify model behavior under production-like and clinical conditions.

1. Microsoft expands sales of OpenAI models in China

Microsoft has made inroads in China's enterprise AI market by selling access to OpenAI models through its cloud business, Bloomberg reported. The arrangement shows how cloud distribution can extend a model company's commercial reach even when geopolitical constraints and local competitors complicate direct market entry.

Cloud distribution gives OpenAI commercial reach without a conventional direct launch, but it also places contracting entity, inference location, and cross-border data movement inside the product path. A policy change by either government could alter that path independently of model performance.

Sources: Bloomberg

2. JPMorgan restricts Anthropic access for Hong Kong employees

JPMorgan Chase cut off Anthropic access for employees in Hong Kong, the Financial Times reported. The restriction demonstrates how a provider's geographic access policy can propagate through a global company's internal tools even when colleagues in other offices continue to use the same model.

The restriction makes office location an entitlement input for a service used elsewhere in the same company. Remote access, prompt routing, and approved alternatives can consequently vary by employee geography even when the underlying application and corporate data classification stay constant.

Sources: Financial Times

3. Sanders proposes a $7 trillion public program for AI and workers

US Senator Bernie Sanders unveiled a proposed $7 trillion plan intended to give the public a larger role in AI investment and governance, Ars Technica reported. The proposal links infrastructure and industrial policy with worker protections, framing model development as more than a private capital market.

The $7 trillion figure describes proposed investment, not enacted spending, and the plan's financing and legislative path remain unresolved. Its scale reframes AI policy around public ownership, worker transition costs, and distribution of productivity gains in addition to model safety rules.

Sources: Ars Technica

4. OpenAI releases a simulation for testing deployment behavior

OpenAI replayed de-identified production conversation prefixes with candidate models and compared simulated behavior with post-release traffic. Across about 1.3 million GPT-5-series conversations, predicted undesirable-behavior rates had a median multiplicative error of 1.5 times.

The method surfaced calculator hacking before release and reduced evaluation-recognition signals; a tool-rollout discriminator scored rich simulations at 49.5%, near chance. Sampling cannot estimate behaviors rarer than roughly one in 200,000 messages, leaving targeted tail-risk tests as a separate evaluation layer.

Sources: OpenAI

5. BRIDGE exposes a 47-point gap between exam and clinical-text performance

BRIDGE contains 87 tasks from 59 clinical data sources across nine languages, eight task types, and 14 specialties. Researchers evaluated 95 language models on work including triage, information extraction, diagnosis, prognosis, and billing codes.

The leading model scored as high as 92 on standardized medical exams but 44.8% on BRIDGE, according to Mass General Brigham. Open models sometimes matched proprietary systems, while medical fine-tunes built on older backbones often trailed newer general models.

Sources: Nature Biomedical Engineering and DOI 10.1038/s41551-026-01719-2 · Mass General Brigham summary